Koutsos E 2017 PHD Thesis

real time semg based muscle fatigue
monitoring using low power integrated

circuits
BY
Ermis Koutsos
January 27, 2017
A thesis submitted for the degree of Doctor of Philosophy of Imperial

College London
Department of Electrical and Electronic Engineering

Imperial College of Science, Technology and Medicine
Declaration of originality
`I hereby declare that this thesis and the work reported herein was composed by and originated
entirely from me. Information derived from the published and unpublished work of others has been
appropriately referenced. Any contribution made to this research by others is explicitly acknowl-
edged in the text.'
Copyright declaration
`The copyright of this thesis rests with the author and is made available under a Creative Commons
Attribution Non-Commercial No Derivatives licence. Researchers are free to copy, distribute or
transmit the thesis on the condition that they attribute it, that they do not use it for commercial
purposes and that they do not alter, transform or build upon it. For any reuse or redistribution,
researchers must make clear to others the licence terms of this work.'
To my beloved parents and dearest brother Alexandros!
Acknowledgements iv
Acknowledgements
First and foremost, I am grateful to my supervisor Dr. Pantelis Georgiou for being extremely
supportive and providing me with help, guidance and motivation throughout this project. His
passion and expertise have shaped my personality as a researcher. But above all, he believed in me
from the rst day of this project and encouraged me to always move onwards and upwards.
Many thanks to Prof. Tor Sverre Lande (Bassen), Dr. Timothy Constandinou, Dr. Yan Liu, Dr.
Yuanqi Hu, Dr. Mohammadreza Sohbati, Dr. Song Luan, Dr. Deren Barsakcioglu, Dr. Lieuwe
Leene, Dr. Sivylla E. Paraskevopoulou, Dr. Amir Eftekhar, Dr. Paul Strutton and Dr. Chloe
Chiou for the countless hours of technical advice, guidance and discussions. Furthermore, I wish to
thank the group administrators, Wiesia Hsissen and Izabela Wojcicka-Grzesiak, and all my friends
in B422 for the warmth they bring to the group. Special mention must go to Maria Khaleeq from the
Department of Physics for her willingness to help with her bonding skills and Raymond Thompson
for his CAD contributions.
Many thanks to the students Jack Heaey, Vlad Cretu, Daiwen Sun, Usama Arif, Silvia Gallo and
fellow Ph.D. candidates and friends Nicolas Moser and Nikolaos Miscourides, that contributed to
my vision of this project.
I am thankful to my good friends Dr. Radu Berdan, Michail Daktylidis, Nikolaos Daktylidis,
Alexandros Sourlas, Katerina Sourla, Antonios Andreou and Dimitrios Andreou for sharing my
anxiety and stress, and supporting me throughout my Ph.D. with advice and fun, relaxing moments.
Special thanks to a true friend and a brilliant engineer, Ilias Pagkalos, for all the problems we solved
and all the good times we enjoyed.
I am deeply grateful to my beloved Dr. Tatiana Trantidou for her support through thick and
thin, her patience, immense help and advice during this long endeavour. Brightening my days and
encouraging me to be myself was the biggest help of all.
Last but most importantly, I would like to express my gratitude to my family, my parents Paraskevi
Koutsou, Konstantinos Koutsos and my brother Alexandros Koutsos, for all their sacrices and for
always showing me the utmost faith, support and encouragement to follow my dreams.
Abstract v
Abstract
Electromyogram (EMG), the recording of the electrical impulses of the muscles, is a rich
source of information, which can facilitate such an insight into our muscles and especially their
activation and fatigue level. Muscle fatigue has been shown to be one of the most important
biofeedback parameters of EMG in rehabilitation, ergonomics and training, by using measured
results from the body to change the way we behave, improve our performance and achieve better
compliance to rehabilitation.
This thesis addresses the challenge of reliably and eciently estimating a muscle's fatigue
state though monitoring surface EMG signals, with the use of low power integrated circuits. CMOS
technology facilitates localised real-time processing to achieve complete miniaturisation, resulting
in an information driven system rather that conventionally data driven system. Thus, reducing
requirements on data transmission, saving power and increasing the degree of freedom for the user.
Several EMG properties progressively change during muscle fatigue and can be quantied
in the time and frequency domains using dierent processing techniques. CMOS technology allows
to signicantly reduce the power and size of the developed EMG processing technology. Firstly, a
CMOS system is presented, capable of measuring the instantaneous Median Frequency (iMDF) of
the EMG signal, which is considered the golden standard for muscle fatigue assessment. Continuing,
a novel bit-stream cross-correlator design that greatly simplies the sEMG signal without any loss
of information is presented for the estimation of the EMG conduction velocity, which is associated
with the physiological changes of the muscle during fatigue. Furthermore, a new metric similar
to iMDF is introduced, combining the advantages of the bit-stream approach that can accurately
track the spectral compression of the sEMG during fatigue with one bit representation. Lastly, a
complete muscle fatigue monitoring System-on-Chip (SoC) is presented, oering complete insight
into the underlying mechanisms and physiological changes during muscle fatigue through sEMG
analysis while operating under both static and dynamic contractions. The proposed approach is
scalable, as several muscle fatigue monitoring SoCs can operate in parallel and periodically relay key
information about the muscle, thus reducing data transmission costs and bandwidth requirements.
Finally, a succession of wearable EMG devices are presented, introducing the use of custom
Application Specic Integrated Circuits in wearable electronics for unsupervised muscle fatigue
monitoring. The wearable nodes are wireless while user ergonomics, power and weight were the
primary design considerations.

vi
Contents
Acknowledgements iv
Abstract v
Contents vi
List of Figures x
List of Tables xvi
List of Abbreviations xviii
Chapter 1. Introduction 1
1.1 Motivation for Muscle Fatigue Monitors . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Thesis Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
References 6
Chapter 2. Fundamentals of Electromyography and Muscle Fatigue Analysis 8
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Electromyography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2.1 Muscle contraction and EMG formation . . . . . . . . . . . . . . . . . . . . . 8
2.2.2 EMG detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Standardised sEMG Detection Protocol . . . . . . . . . . . . . . . . . . . . . 14
2.3 Myoelectric Manifestations of Muscle Fatigue . . . . . . . . . . . . . . . . . . . . . . 17
2.3.1 Processing in Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.1.1 Root Mean Square, Spike Counting, Zero Crossing Rate . . . . . . . 19
2.3.1.2 Muscle Fibre Conduction Velocity . . . . . . . . . . . . . . . . . . . 20
2.3.2 Processing in Time-Frequency Domain . . . . . . . . . . . . . . . . . . . . . . 21
2.3.2.1 Short Time Fourier Transform . . . . . . . . . . . . . . . . . . . . . 22
2.3.2.2 Continuous Wavelet Transform . . . . . . . . . . . . . . . . . . . . . 22
2.3.2.3 Other Time-Frequency Methods . . . . . . . . . . . . . . . . . . . . 23
2.3.3 Fatigue Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

Contents vii
2.3.4 Repeatability of sEMG Muscle Fatigue Estimates . . . . . . . . . . . . . . . . 26
2.4 Current Approaches for Muscle Fatigue Monitoring . . . . . . . . . . . . . . . . . . . 26
2.4.1 Approach in this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
References 33
Chapter 3. An ASIC for sEMG Spectrum Monitoring using a Median Frequency
Estimator 45
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.2 Methods for real-time iMDF tracking systems . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Proposed IC iMDF Tracking System . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.0.1 iMDF Model Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3.0.2 iMDF Model Simulations . . . . . . . . . . . . . . . . . . . . . . . . 53
3.4 iMDF Tracker Circuit Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5 Measurement Results and Reections on iMDF Tracker . . . . . . . . . . . . . . . . 65
3.5.1 Design Considerations for Future Improvements . . . . . . . . . . . . . . . . . 65
3.5.2 A novel Neuromorphic based iMDF Tracking System . . . . . . . . . . . . . . 67
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
References 74
Chapter 4. A wearable MFM System based on the Muscle Fibre Conduction Ve-
locity 76
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.2 Methods for Conduction Velocity Estimation . . . . . . . . . . . . . . . . . . . . . . 77
4.3 MFCV Tracking Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.3.1 Design Considerations for the CMOS Cross-Correlator . . . . . . . . . . . . . 83
4.3.2 Proposed Bit-Stream Cross-Correlator . . . . . . . . . . . . . . . . . . . . . . 88
4.4 MFCV Tracking ASIC Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.4.1 sEMG Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.5 MFCV Tracking ASIC Experimental Results . . . . . . . . . . . . . . . . . . . . . . . 97
4.5.1 System Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.5.2 Validation on Human Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.5.2.1 Experimental Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.5.2.2 Results from Human Trials . . . . . . . . . . . . . . . . . . . . . . . 103
4.5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
4.6 Wearable Muscle Fatigue Monitor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
References 111
Contents viii
Chapter 5. A MFM ASIC exploring the iMDF and Zero Crossings relationship 117
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.2 Median Frequency and sEMG Spike Duration . . . . . . . . . . . . . . . . . . . . . . 118
5.2.1 Algorithm for Spike Duration Based iMDF Approximation . . . . . . . . . . . 123
5.3 SDMF Tracking ASIC Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.3.1 sEMG Front End Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . 129
5.3.1.1 Choice of Electrodes . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.3.1.2 Analogue Front End . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.3.1.3 sEMG Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.4 SDMF Tracking ASIC Experimental Results . . . . . . . . . . . . . . . . . . . . . . . 134
5.4.1 System Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.4.2 Physiological Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.4.4 The second version of the Wearable Muscle Fatigue Monitor . . . . . . . . . . 144
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
References 148
Chapter 6. A sEMG Analysis System-on-Chip for Muscle Fatigue Monitoring 152
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.2 sEMG SoC Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.3 Adaptation for Dynamic Contractions . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.4 Circuit Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.4.1 Digital Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.5 Mucle Fatigue Monitoring ASIC Overview . . . . . . . . . . . . . . . . . . . . . . . . 162
6.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
References 165
Chapter 7. Conclusion 167
7.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.2 Recommendations for Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.2.1 System Optimisation and Enhancement . . . . . . . . . . . . . . . . . . . . . 168
7.2.2 Next Generation Wearable Electronics . . . . . . . . . . . . . . . . . . . . . . 169
7.2.3 Multimodal Biosensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
List of publications 170

Appendix A. 171
A.1 VHDL code for neuromorphic iMDF tracker . . . . . . . . . . . . . . . . . . . . . . . 171
Appendix B. 175
B.1 VHDL code for MFCV ASIC (OPTIMUS) . . . . . . . . . . . . . . . . . . . . . . . . 175
Contents ix
B.2 VHDL code for SDMF ASIC (BATMAN) . . . . . . . . . . . . . . . . . . . . . . . . 179
B.3 VHDL code for MFM ASIC (ATLAS) . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Appendix C. 195
C.1 Bit-Stream Cross-Correlator Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . 195
C.2 Noise Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199

x
List of Figures
1.1 CMOS technology can unlock the potential of future EMG biofeedback systems by overcoming
the limitations current technologies are facing. Integration of detection and processing stages will
signicantly reduce data transmission and allow several EMG channels to be processed concurrently,
thus increasing the insigh and eectiveness of biofeedback in the future. . . . . . . . . . . . . . 3
2.1 (a) Muscle anatomy showing motor unit, motor neuron and muscle bre. Figure from [2], (b) Muscle
cell anatomy. Figure from [3]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 (a) The actin-myosin binding cycle. Figure from [3], (b) Muscle cell membrane potential due to N a+
and (K + ) concentrations. Once the membrane threshold is exceeded, the contraction/relaxations
cycle of the muscle begins, based on the sliding lament model. . . . . . . . . . . . . . . . . . 11
2.3 Factors aecting the shape, amplitude and frequency of the detected MUAP on the skin surface
during a sustained contraction. Figure from [6]. . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 The eect of monopolar, bipolar (single dierential) and double dierential electrode congurations
on the recorded sEMG signal. Figure from [15]. . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 Spatial ltering eect of (a) monopolar , (b) bipolar and (c) double dierential electrode congura-
tions. (left) conguration topology, (middle) spatial lter impulse responce and (right) lter output
signal. Figure from [16]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6 Electrode positioning between myotendonous junction and innervation zone. . . . . . . . . . . . 15
2.7 Eect of interelectrode distance and electrode location on the sEMG amplitude. Figure from [21]. . 16
2.8 The power spectral density of the sEMG signal during a static, sub-maximal, fatiguing contraction. 19
2.9 Progression of the spectral compresion of the sEMG signal during fatigue. The iMDF is marked in
red. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.10 The progression of dierent sEMG variables in the time and frequency domian during a fatiguing
contraction. Figure from [63]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.11 Joint Analysis of myoelectric Spectrum and Amplitude decision case algorithm. . . . . . . . . . . 26
3.1 (a) Power Spectral Density function of EMG signal during fatigue, (b) top: sEMG signal from static
fatiguing cobntraction, bottom: iMDF of the recorded sEMG signal. . . . . . . . . . . . . . . . 46

List of Figures xi
3.2 Partitioning the sEMG spectrum with the use of two tunable lters (high-pass and low-pass) with
the same corner frequency. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3 System architecture for the estimation of iMDF, utilising tunable lters and a feedback controller.
Recreated from [3]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4 iMDF estimation algorithm: The power of the up-modulated sEMG signal is partitioned using a
high order bandpass lter. The dierence in power of the original singal and the up-modulated one
is integrated and used to control the VCO frequency. . . . . . . . . . . . . . . . . . . . . . . 48
3.5 System-level block diagram of the chosen iMDF tracking implementation. . . . . . . . . . . . . . 49
3.6 Simulink model for tracking iMDF. The signal passes through an antialliasing lter and then proc-
cesed by the algorithm described above. The output of the system is the dierence between fc and
fV CO . The output is low-pass ltered for a smoother response. . . . . . . . . . . . . . . . . . . 50
3.7 A 250 ms sEMG waveform is sown in a) and the corresponding signal transformations as it goes
through nodes B-F in the system model Fig. 3.6 (b-f respectively). . . . . . . . . . . . . . . . . 54
3.8 IMDF tracking during fatiguing contraction from muscles a) LES T12, b) RES T12, c) LES L4, d)
RES L4, e) L RA and f) R RA. Blue line is iMDF using system described. Red line serves as a guide. 54
3.9 Circuit schematic of the gm-C VCO. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.10 Frequency responce of the gm-C VCO with a bias current in the range of 100-300 nA. . . . . . . . 56
3.11 Current mode, four quadrant, translinear loop multiplier. . . . . . . . . . . . . . . . . . . . . 56
3.12 Harrison-amplier architecture [7] for a variable gain, subtracting amplier. . . . . . . . . . . . . 57
3.13 Dierential in/out switched-capacitor integrator with oset compensation. . . . . . . . . . . . . 58
3.14 Schematic of the trasmition gate used with MOSFET capacitors for charge injection absorbtion. . . 58
3.15 FLF eliptic lter architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.16 Biquad architecture circuit diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.17 OTA used in biquad lter with source degeneration. . . . . . . . . . . . . . . . . . . . . . . . 60
3.18 Current mode adder circuit diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.19 Simulation of the bandpass elliptic lter gain response: a) ideal lter, circuit implementation and
cascaded lter response, b) closeup of a) around the 3dB point, c) Monte-Carlo extremities of the
cascaded lter and d) eect of Q and centre frequency variation on the FLF structure frequency
response. Figure from [11]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.20 Blue line: cadence simulation iMDF tracking during fatiguing contraction from muscles a) LES T12,
b) RES T12 and c) LES L4. Red line: iMDF tracking using MATLAB® simulation. . . . . . . . 64
3.21 Die micro-photograph of the iMDF tracking ASIC. . . . . . . . . . . . . . . . . . . . . . . . 65
3.22 (a) The dierential output of the PGA durring and after a RESET period, (b) The dierential
output of the SC integrator following the PGA durring the same period. . . . . . . . . . . . . . 66
3.23 Block diagram representation of the second version of the iMDF Tracking System. . . . . . . . . 67
List of Figures xii
3.24 Second order Gm -C High Pass Filter Schematic. By operating current source transistors M13 and
M14 in the weak inversion region, the cuto frequency depends on the biasing current. An array of
binary scaled current mirrors is used to control the bias current and adjust the cuto frequency of
the lter. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.25 Analogue Current Mode Squarer Circuit. The circuit also acts as a transconductor which converts
the voltage mode signal into current mode signal which can be used to drive the integrate-and-re
neuron circuit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.26 Integrate-and-Fire Neuron Circuit Implementation. The circuit integrates the signal power over
an adjustable time window which gives the average signal power. The integrated signal power is
represented as the amount of spikes red from the neuron circuit. The neuron is tuned to be linear. 70
3.27 Blue Trace: Median Frequency during fatigue using proposed system from (a) Subject 1, (b) Subject
2, (c) Subject 3, Red Trace: Median Frequency using MATLAB® medf req() function. . . . . . . 71
4.1 Distance between reference points: a) Seven sEMG signals detected along the muscle bre, b)
Alignment of EMG segments from reference points. Figure from [20]. . . . . . . . . . . . . . . . 79
4.2 Cross-correlation result of two EMG signals detected along the muscle bre path. (a) Reference and
Comparison (delayed) EMG signals, (b) Cross-correlation coetient. Figure from [26]. . . . . . . 81
4.3 The propagation of MUAPs along the muscle bre in two directions, originating from the innervation
zone. Figure from [28]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.4 Cross correlation detection points on bicepts brachii muscle. (a) single electrode, monopolar detec-
tion, (b-c) double electrode, bipolar detection. Figure from [32]. . . . . . . . . . . . . . . . . . 82
4.5 Lande et al. bit-stream cross correlator. The two signals are buered and latter correlated using an
AND or XOR gate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.6 Proposed architecture of the bit-stream cross correlator. The buering blocks have been replaced
with delay blocks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.7 Performance of the proposed bit-stream cross-correlator with varying threshold, correlation window
and sampling frequency: (a) MARD between MATLAB® xcorr() function and modelled system
with a varying threshold, (b) MARD between MATLAB® xcorr() function and modelled system,
where W is the correlation window (s). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.8 Correlation stage with delay block, counter and correlator (XNOR). . . . . . . . . . . . . . . . 91
4.9 Correlation stage signals. (a) EMG signal, (b) xd (n) bit-stream, (c) x(n) bit-stream delayed by
10 previous correlation stages, (d) x(n) bit-stream delayed by all previous stages, (e) XNOR bit
correlation result between (b) and (c), (e) counter result after adding all XNOR values in real time. 92
4.10 Maximum Function Flow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.11 Timing control unit circuit diagram. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

List of Figures xiii
4.12 Timing Control Unit signal ow. (a) sampling clock, (b) correlation time clock, (c) delay clock, (d)
enable maximum function, (e) read data for output and (f) reset counters. . . . . . . . . . . . . 95
4.13 System block architecture of the Muscle Fibre Conduction Velocity tracking ASIC. . . . . . . . . 95
4.14 Die micro-photograph of the MFCV tracking ASIC. . . . . . . . . . . . . . . . . . . . . . . . 97
4.15 (a) Input-referred noise measurement (red) and simulation (black) results of the IA channel 1, (b)
Common-mode rejection ratio measurement of the IA (chan. 1) using a 500mVpp input signal. . . . 99
4.16 Top: sMEG segment from Channel 1, middle: sMEG segment from Channel 2. Delay between the
two channels is easily detectable by eye, bottom: Comparator output from channel 2 for reference. . 100
4.17 Cross-correlation delay estimation under static contraction from rst test subject. sEMG was
recorded from the inner left biceps brachii. Sampling clock is 10 kHz and window size clock is
1 Hz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4.18 The recorded sEMG signals from Subject 2 are played back to the system using a funciton generator.
Thus, the eect of thresshold can be studied. . . . . . . . . . . . . . . . . . . . . . . . . . . 101
4.19 Experimental set-up protocol. The subjects keep the force output constant and within limmits with
the aid of a visual, real-time representation of the applied force. . . . . . . . . . . . . . . . . . 102
4.20 Left: Architectural diagram of experimental setup. Middle: Electrode conguration for bipolar
single dierential sEMG amplication. Muscle zones are displayed. Right: Experimental setup
including microcontroller for data transmission, clock generation and threshold position. Faraday
cage used in noise measurements only. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.21 (a) MFCV trial results from subject 11. The rested state MFCV is 6 sec
m
and progresses to 4 sec
m
as the muscle fatigues, (b) Fatigue trial results comparison between MFCV tracking ASIC and
MATLAB® . The MARD is presented in percentile. . . . . . . . . . . . . . . . . . . . . . . . 105
4.22 (a) Minimum area for 4 electrode conguration. (b) PCB dimensions constrains based on 4 electrode
conguration. (c) Bottom side of the case with electrode snap sockets. (d) Completed PCB with
soldered components. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.23 Exploded view of the case with the PCB inside.Snap sockets are screwed in place in the hole
oppenings. The device is attached on the user by the electrode adhesive. . . . . . . . . . . . . . 108
4.24 (a) Calibrate parameters tab of the Android application. The used can specify the common mode
voltage, threshold voltage, sampling frequency and correlation window, (b) Live capture tab of the
Android application. The user can visualise sEMG and muscle fatigue data, as well as EEPROM
saved data and choose to export the recorded values. . . . . . . . . . . . . . . . . . . . . . . 109
4.25 (a) Wearable Muslce Fatigue Tracking System, showning the Android application and the wearable
device, (b) Complete Muslce Fatigue Tracking System, showning the Android application and the
wearable device attached on a user. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

List of Figures xiv
5.1 sEMG segment indicating the zero crossings and the time interval between these crossigns. . . . . 120
5.2 SDMF and iMDF plots of biceps brachii sEMG recordings of three dierent subjects (a), (b), (c)
respectively). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.3 (a) sEMG segment showing the addition of low frequency (10 Hz) motion artefacts, (b) iMDF
computed for the unaltered and the altered sEMG signal. . . . . . . . . . . . . . . . . . . . . 122
5.4 SDMF and iMDF plots of biceps brachii sEMG recordings of three dierent subjects (a), b), (c)
respectively). The SDMF was computed using the artefact infected sEMG signal and the iMDF
using the clean sEMG signal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.5 Architectural block diagram illustration of the proposed SDMF tracking algorithml. . . . . . . . . 124
5.6 (a) is the undelayed positive toggle bit-stream, (b) the undelayed negative toggle bit-stream, (c) the
already delayed positive toggle bit-stream and (d) the comparison between the two versions of the
same bit-stream. If the two signals overlap, then the toggle length is greater or equal to the added
delay at that stage and the comparison output increments a counter. . . . . . . . . . . . . . . . 125
5.7 SDMF tracking system circuit implementation. . . . . . . . . . . . . . . . . . . . . . . . . . 125
5.8 Circuit implementation of the delay & overlap stage, including an overlap correction circuit. . . . 126
5.9 (a) the counter value from all stages at three distinct points in time. All three plots show a decreasing
trend with time. (b) the counter values for three distinct stages for the complete duration of the
sEMG signal of a fatiguing contraction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.10 Estimated SDMF comparison between the median() function in MATLAB® and the proposed
algorithm. The error between them is given by MARD = 6.5%. . . . . . . . . . . . . . . . . . 128
5.11 System block architecture of the SDMF tracking ASIC. . . . . . . . . . . . . . . . . . . . . . 128
5.12 Skin-electrode impedance model for dry and wet electrodes. . . . . . . . . . . . . . . . . . . . 129
5.13 (a) transconductance, (b) transimpedance (single ended), (c) DC servo and (d) transimpedance
(double ended) circuit implementations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.14 Second order, low pass lter, gm-C architecture. . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.15 Die micro-photograph of the SDMF tracking ASIC. . . . . . . . . . . . . . . . . . . . . . . . 134
5.16 (a) Input-referred noise measurement results of the analogue front-end signal path (IA, PGA, LPF),
(b) Common-mode rejection ratio measurement of the IA using a 500mVpp input signal. . . . . . . 136
5.17 Transient responce of the IA start-up. The ouput reached a stadle operating point after 10 seconds
due to the large time constant involved with the DC servo loop. . . . . . . . . . . . . . . . . . 137
5.18 PCB measurement board for the SDMF ASIC. The dry electrodes are made of steel disks, 1cm in
diameter. The ASIC communicates withthe microcontroller via SPI and the microcontroller to the
PC via UART. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.19 Transient reconding from the upper part of the inner left biceps brachii muscle. ECG heartbeats
were detected and shown in detail in the left gure. . . . . . . . . . . . . . . . . . . . . . . . 138

List of Figures xv
5.20 Shown are an sEMG recording and the position of the selected hysteresis thresholds. Hysteresis is
introduced with the aid of positive feedback, ass seen in the bottom left conrer. . . . . . . . . . . 139
5.21 (a) sEMG recording of a contraction starting from a rested state until exhaustion, (b) smoothed
sEMG amplitude using a 300 ms moving average window. . . . . . . . . . . . . . . . . . . . . 140
5.22 (a) Comparison of the SDMF (red) and iMDF (blue) trends before and after the muscle contraction
(green regions), (b-f) SDMF (red), iMDF (blue) and their smoothed (black) trends for Subjects 1-5. 142
5.23 (a) Smoothed SDMF responce with sEMG playback from Subject 1 and a varying PGA gain from
1 to 7, (b) SDMF plot and smoothed responce with sEMG playback from Subject 2 and a varying
PGA gain from 3 to 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
5.24 (a) Revised Wireless MFM node, bottom view, (b) Revised Wireless MFM node, top view, (c) Size
comparison between old and new MFM versions, (d) Size comparison between old and new MFM
versions with their cases on, (e) Size comparison between wet electrode sockets and custom made
dry electrodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
5.25 (a) Application window where the user can select number and location of wireless MFM nodes, (b)
Application window where the user can save muscle names and locations used. . . . . . . . . . . 146
6.1 Onset detector circuit on prototyping board. . . . . . . . . . . . . . . . . . . . . . . . . . . 155
6.2 (a) sEMg recording of a dynamic contraction, (b) close-up section of dynamic contraction, (c) SDMF
output for the dynamic contraction recording, (d) synthesised dynamic contraction (blue) and the
onset pulse (red), (e) SDMF output for the synthesised dynamic contraction and (f) SDMF output
for the synthesised dynamic contraction with the onset detector on. From (a) it is clear that the
SDMF algorithm is better at detecting the start/end point of the contraction. . . . . . . . . . . 158
6.3 Onset detector circuit implementation for the MFM ASIC. . . . . . . . . . . . . . . . . . . . . 159
6.4 (a) Frequency responce of the fth order SC LPF with CLK = 4 kHz, (b) Transient responce of the
onset detector. The sEMG signal is approximated with a 100 Hz sine wave. . . . . . . . . . . . . 159
6.5 FSM state diagram. The SPI[15:14] bits control the state transitions. . . . . . . . . . . . . . . . 161
6.6 System block architecture of the Muscle Fatigue Monitoring ASIC. . . . . . . . . . . . . . . . . 162
6.7 Die micro-photograph of the MFM ASIC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

xvi
List of Tables
2.1 Repeatability of Initial Values of Mean Power Spectral Frequency (MNF), Average
Rectied Value (ARV), and Muscle Fibre Conduction Velocity (MFCV) during Iso-
metric Voluntary and Electrically Elicited Contractions in Dierent Muscles. Table
recreated from [8]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Major commercial EMG acquisition systems. . . . . . . . . . . . . . . . . . . . . . . 29
2.3 Comparison amongst the presented methods for muscle fatigue tracking. . . . . . . . 31
3.1 6th order FLF bandpass elliptic lter parameters. . . . . . . . . . . . . . . . . . . . . 59
3.2 Biquad circuit parameters. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.3 Comparison between circuit simulations and MATLAB

® . . . . . . . . . . . . . . . . 62
3.4 System-Level power breakdown (C35, 3.3V). . . . . . . . . . . . . . . . . . . . . . . . 63
3.5 Results comparison between proposed system and MATLAB® . . . . . . . . . . . . . . . . . . 71
3.6 Power Consumption Breakup (C35, 3.3V) . . . . . . . . . . . . . . . . . . . . . . . . 72
4.1 Factors aecting MFCV estimates. Each primary factor call inuence MFCV esti-
mates depending on sEMG generation and detection [5]. . . . . . . . . . . . . . . . . 83
4.2 Error results between xcorr() and the linear t of the added time delay between
the two sEMG recordings. Correlation windows and sEMG sampling frequency are
varied to obtain optimum working region. . . . . . . . . . . . . . . . . . . . . . . . . 85
4.3 Error results between xcorr() and the linear t of the added time delay between the
two resulting bit-streams. Correlation windows and sEMG sampling frequency are
varied to obtain optimum working region. . . . . . . . . . . . . . . . . . . . . . . . . 87
4.4 Error results between the cross-correlator proposed in [33] and the linear t of the
added time delay between the two resulting bit-streams. Correlation windows and
sEMG sampling frequency are varied to obtain optimum working region. . . . . . . . 87
4.5 MFCV ASIC parameters, size and power breakdown (C35, 3.3V). . . . . . . . . . . . 98
4.6 Subject trial results: a) subject number, b) change in MFCV from chip, c) change in
MFCV from MATLAB

® , d) MARD and e) MVC force. . . . . . . . . . . . . . . . . 104
4.7 Comparison between MFCV physiological values. . . . . . . . . . . . . . . . . . . . . 104

List of Tables xvii
4.8 Summary of MFCV tracking ASIC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.9 Summary of Muscle Fatigue Tracking Wearable Device. . . . . . . . . . . . . . . . . 110
5.1 Correlation coecient values between the original and smoothed iMDF and SDMF. . 121
5.2 Correlation coecient values between the original and smoothed iMDF and SDMF
with the presence of simulated motion artefacts. . . . . . . . . . . . . . . . . . . . . . 122
5.3 SDMF ASIC power breakdown (C35, 3.3V). . . . . . . . . . . . . . . . . . . . . . . . 135
5.4 Correlation coecient values between the smoothed iMDF and measured SDMF. . . 141
5.5 Summary of SDMF tracking ASIC. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.1 Examples of dynamic contractions and the eect of the onset detector and pausing
algorithm on SDMF accuracy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
6.2 Input and output SPI control sequences. . . . . . . . . . . . . . . . . . . . . . . . . . 161

xviii
List of Abbreviations
AC: Alternating Current
ADC: Analogue to Digital converter
Ag/AgCl: Silver/Silver chloride
ANN: Articial Neural Network
AP: Action Potential
ASIC: Application Specic Integrated Circuit
ATP: Adenosine Triphosphate
BLE: Bluetooth Low Energy
BMIs: Brain Machine Interfaces
BW: Bandwidth
Ca2+ : Calcium ions
CMOS: Complementary Metal-Oxide Semiconductor
CMF: Common Mode Feedback
CMRR: Common Mode Rejection Ratio
CSV: Comma Separated Value
CWD: Choi-Williams Distribution
CWT: Continuous Wavelet Transform
DAC: Digital to Analogue Converter
DC: Direct Current
DFF: D-type Flip Flop
DFT: Discrete Fourier Transform
ECG: Electrocardiography
EEG: Electroencephalogram
EEPROM: Electrically Erasable Programmable Read-Only Memory

List of Abbreviations xix
EMG: Electromyography
EOC: End of Conversion
ESD: Electrostatic Discharge
HPF: High Pass Filter
FFT: Fast Fourier Transform
FLF: Follow the Leader Feedback
FSM: Finite State Machine
GBW: Gain Bandwidth Product
IA: Instrumentation Amplier
IC: Integrated Circuit
ICC: Intraclass Correlation Coecient
IED: Inter Electrode Distance
iEMG: integrated EMG
iMDF: instantaneous Median Frequency
iMNF: instantaneous Mean Frequency
ISI: Inter Spike Interval
JASA: Joint Analysis of myoelectric Spectrum and Amplitude
K+ : Potassium ions
LPF: Low Pass Filter
LSB: Least Signicant Bit
MARD: Mean Absolute Relative Dierence
MAV: Mean Absolute Value
ME: Myoelectric
MFCV: Muscle Fibre Conduction Velocity
MFM: Muscle Fatigue Monitoring
MMF: Multivariable Mapping Functions
MOSFET: Metal-Oxide Semiconductor Field Eect Transistor
MSP: Mean Spike Length
MU: Motor Unit
MUAP: Motor Unit Action Potential

List of Abbreviations xx
MVC: Maximum Voluntary Contraction
Na+ : Sodium ions
NMOS: n-channel Metal-Oxide semiconductor
NSEM: Normalised Standard Error of the Mean
opamp: Operational Amplier
OTA: Operational Transconductance Amplier
PCA: Principal Component Analysis
PCB: Printed Circuit Board
PGA: Programmable Gain Amplier
PLD: Pulse Length Detector
PMOS: p-channel Metal-Oxide semiconductor
POR: Power-on-Reset
PSD: Power Spectral Density
PV: Polarization Voltage
PWM: Pulse Width Modulation
RMS: Root Mean Squared
SC: Switched-Capacitor
SDMF: Spike Duration based Median Frequency approximation
sEMG: surface Electromyography
SENIAM: Surface Electromyography for Noninvasive Assessment of Muscles
SNR: Signal to Noise Ratio
SoC: System-on-Chip
SOIC: Small Outline Integrated Circuit
SPI: Serial Peripheral Interface
STFT: Short Time Fourier Transform
SW: Spike Width
T/A: Turns per unit time
TC: Transconductance
TI: Transimpedance
THD: Total Harmonic Distortion

List of Abbreviations xxi
UART: Universal Asynchronous Receiver/Transmitter
USB: Universal Serial Bus
VCM: Common Mode Voltage
VCO: Voltage Controlled Oscillator
VHDL: Very high speed integrated circuit Hardware Description Language
WVD: Winger-Ville Distribution
ZCR: Zero Crossing Rate

1
Chapter 1
Introduction
1.1 Motivation for Muscle Fatigue Monitors
E
LECTROMYOGRAPHY (EMG) is a technique used to evaluate the electrical activity of
muscles. Etienne-Jules Marey was the rst to record the electrical activity of muscles in
1890, introducing the term electromyography [1]. Muscle activity is controlled by the nervous
system and the resulting muscle contractions give rise to electrical currents which in turn form the
EMG signal. The EMG signal appears to be random, yet it is a rich source of information which
can facilitate such an insight into our muscles and especially their activation and fatigue level.
Muscles are vital for strength, balance, heat, posture and movement. Feedback driven from
the way our muscles behave and fatigue has the potential to improve the quality of life of millions
of people suering from muscular related disorders and set the foundations for ground breaking
rehabilitation technologies. To date, there are no reliable metrics of fatigue and point-of-care au-
tonomous systems are still a dream. This work addresses the challenge of reliably and eciently
estimating a muscle's fatigue state and the implementation of integrated systems to detect it.
There exist three types of fatigue; localised muscle fatigue, central fatigue (nervous system)
and peripheral fatigue. Localised muscle fatigue analysis, has numerous applications. The demand
for EMG analysis and biofeedback in rehabilitation (surgery recovery, stroke recovery, spinal cord
injuries, brain injuries) and its application in conjunction with Electroencephalogram (EEG) for
identifying neuropathies, brain disorders and assisting Brain Machine Interfaces (BMIs) is rapidly
growing in translational healthcare. EMG analysis is advantageous in these areas by using measured
results from the body to change the way we behave, improve our performance and achieve better
compliance to rehabilitation through the process of biofeedback.

1.1 Motivation for Muscle Fatigue Monitors 2
EMG analysis has been widely applied in ergonomics to isolate, understand and treat muscle
pain; fatigue analysis has been leveraged to study the eects of exposures at work. In heavy work
environments, mechanical overstress (fatigue) from prolonged low-level contractions can be harmful,
causing injuries and leading to muscular disorders [2]. Another example involves long distance truck
drivers, where excessive fatigue in the shoulder muscles can lead to driver loss of focus. In pursuit
of identifying the underlying causes of back and neck pain, a condition aecting almost 80% of
all people, EMG has been used to assess changes in muscle fatigue, recruitment and coordination
[3, 4]. EMG and fatigue analysis plays a vital role in biofeedback rehabilitation. For example,
knee rehabilitation after surgery due to osteoarthritis focuses on maintaining a balance between the
two large muscles that hold the patella (knee cap) in place, Vastus Lateralis and Vastus Medialis.
Careful tracking of muscle fatigue of these two muscles can provide essential adjustments to the
rehabilitation procedure and avoid patellar dysfunction [5]. Finally, strength training and movement
analysis in sports science have greatly beneted from EMG feedback. Monitoring muscle activity
and adjusting training regimes can maximise eectiveness and avoid injuries due to excessive fatigue.
Modern fatigue analysis tools involve collecting large amounts of EMG data, using either
wireless or wired approaches. Decoding the EMG and extracting specic information is compu-
tationally demanding and is performed in workstations, also involving advanced signal processing
techniques and dedicated personnel continuously monitoring the process. Using wires introduces
noise sources, which deteriorate the quality of the EMG signal and limit the mobility of the user
in various exercises. Moreover, these techniques limit considerably the possible applications based
on muscle fatigue, because they reduce the user's degree of freedom and the number of available
measuring sites. On the other hand, wireless approaches continuously digitise and transmit the
raw EMG signal. To increase eectiveness of biofeedback, numerous sensors are needed to monitor
several muscle groups. However, the sheer amount of transmitted data and processing requirements
form a bottleneck and challenge biofeedback feasibility.
Consider the following example of a wireless device employing Bluetooth transmission. The
theoretical limit of Bluetooth Low Energy throughput is 84 kbps. Assuming a suitable for EMG 10
bit Analogue to Digital Conversion at 1.5 kHz, the maximum number of EMG channels that can be
transmitted with one device is n=5. On the other hand, muscle fatigue evolves very slowly (sub-Hz).
As a result, transmitting fatigue relevant information only would signicantly compress the EMG
content. Therefore, increasing the number of channels while reducing power consumption can be
1.1 Motivation for Muscle Fatigue Monitors 3
Current Technologies EMG Fatigue Analysis Applications Future of Biofeedback

Wireless Wired
Biomechanics
Strength training
Orthopedic research Active therapy training
Posture analysis Surgery rehabilitation
Risk prevention
Design improvements
product certification
www.athos.com
www.delsys.com
+ Wired systems are accurate, ideal for clinical use CMOS Technology + Local detection and processing
+ Portable systems extend user mobility + Information driven system
-- Post processing of EMG + Low power and small form factor + Scalable, integrated electronics
-- Gap between recording systems and analysis tools + Embeded custom algorithms + Multimodal, real-time monitoring
-- Limmited application range + Combiled EMG detection and processing + Continious, multi-site monitoring
-- Lack of real-time feedback + Real-time biofeedback feasibility + Real-time biofeedback
Figure 1.1: CMOS technology can unlock the potential of future EMG biofeedback systems by
overcoming the limitations current technologies are facing. Integration
Limited of detection and processing
power consumption.
stages will signicantly reduce data transmission and allow
Localseveral
processingEMG channels
to reduce transmissiontopower.
be processed
Extended degree of freedom and applications.
concurrently, thus increasing the insigh and eectiveness of biofeedback in the
Information driven system. future.
Multi
node, real time monitoring.
Bortec Electronics
No need for post processing of data.
AD Instruments Delsys Non obstructive monitoring.
Cadwell Biometrics
Myotronics
achieved by introducing localised processing techniques and Real
Myon
thus
timereducing
feedback. the data rate. Eorts to There is a gap between mod
Motion Lab Systems Monnit analysis tools/algorithms an
Noraxon date back over three decades. At that time, most of the designs platforms capable of analysi
create a portable fatigue monitor
fatigue.
involved either custom analogue circuits or microcontroller based, digital implementations [612].
Due to the limited resources of the time, the designs oered limited tracking capabilities and involved
bulky equipment. Even though modern microcontrollers are becoming more power ecient and
more capable, microcontroller based systems are still limited by the number of EMG channels that
can be processed in parallel and range of processing methods that can allow ecient computation.
Whilst the current approaches oer invaluable insight, real-time biofeedback outside a controlled
environment remains a critical barrier. Consequently, there is a clear need for a portable, compact
and wearable system, capable of unsupervised EMG fatigue analysis. Fig. 1.1 depicts the limitations
of the current approaches for EMG muscle fatigue analysis along with the application areas that
will benet in the future from real-time biofeedback systems. Most of the EMG applications shown
in Fig. 1.1 are limited to clinical/laboratory environments and have not made their way in our daily
lives due to the lack of real-time biofeedback systems. However, with the use of CMOS technology,
EMG detection and processing can be eciently combined to create a miniaturised EMG system
which can be easily scaled to monitor several muscles concurrently and eciently, allowing for real-
time, continuous operation. Thus, CMOS technology can be exploited to unlock the feasibility of
biofeedback in the future.

1.2 Research Objectives 4
1.2 Research Objectives
This research is aimed in developing a muscle fatigue monitoring System-on-Chip (SoC) capable of
ecient independent processing and reduced data transmission. This aims to deliver an autonomous,
low power and portable surface EMG (sEMG) monitoring platform to increase the accessible muscle
sites and activity range, unlocking the potential of biofeedback applications.
CMOS technology facilitates the integration of amplication and processing stages to achieve
complete miniaturisation and reduced power consumption. Furthermore, it enables the develop-
ment of application-specic algorithms to increase the versatility and eciency of the system.
Subsequently, this work is targeted towards providing some insight into answering the following
questions:
1. Which sEMG attributes are most accurate and suitable for muscle fatigue monitoring?
2. What SoC topology and on-chip processing techniques are most suited to tracking muscle
fatigue metrics with low power and reduced transmission requirements?
3. What are the limiting factors for real-life unsupervised operation and how can a system in
CMOS technology adapt to these?
1.3 Thesis Organisation
This thesis is organised in a chronological order, with each chapter building upon the work presented
in preceding chapters.
Chapter 2 provides an overview of the mechanics of muscle contraction, sEMG formation and
their dening role in the interpretation and analysis of muscle fatigue. The most helpful sEMG signal
variables along with the most promising signal processing techniques for muscle fatigue tracking are
identied. Limitations of current approaches and further opportunities for research are highlighted.
Following this, in Chapter 3 the properties of the sEMG power spectrum during fatigue, the
benets of instantaneous Median Frequency (iMDF) for spectral tracking and possible circuit im-
plementations were thoroughly investigated. Moreover, two iMDF tracking systems were presented
along with their strengths and weaknesses, as one was a completely analogue implementation and
the other was based on a mixed-signal approach.
In Chapter 4, a novel method for estimating the muscle bre conduction velocity based on a
bit-stream cross-correlator is presented. This approach simplies the sEMG signal without any loss
1.3 Thesis Organisation 5
of information, thus reducing computational complexity and minimising power consumption. The
developed technology is evaluated in a clinical study and embedded in a wireless node operating
with commercially available wet electrodes.
Chapter 5 introduces a new metric similar to iMDF combining the advantages of the bit-
stream approach developed in Chapter 4. This was achieved by exploiting a link between zero
crossing and median frequency. The developed technology was tested with human subjects, demon-
strating that the validity of the proposed fatigue metric. A second wireless node making use of dry
electrodes is designed to accommodate the new microchip, further reducing the size and weight of
the previously developed node.
Chapter 6 builds upon the systems developed in Chapters 4 and 5 to deliver the rst SoC
oering complete insight into the underlying mechanisms and physiological changes during muscle
fatigue through sEMG analysis. This is achieved by exploiting both amplitude and spectrum pa-
rameters of the sEMG signal. The new platform was designed to operate under both static and
dynamic contractions to maximise impact in real-life applications.
Chapter 7 summarises the accomplished work, highlights the contributions of this PhD thesis
and nally provides future directions for this work.

REFERENCES 6
References
[1] M. Reaz, M. Hussain, and F. Mohd-Yasin, Techniques of emg signal analysis: detection,
processing, classication and applications, Biological procedures online, vol. 8, no. 1, pp. 11
35, 2006.
[2] G. Caer, D. Heinecke, and R. Hinterthan, Surface emg and load level during long-lasting
static contractions of low intensity, International Journal of Industrial Ergonomics, vol. 12,
no. 1-2, pp. 7783, 1993.
[3] S. H. Roy, C. J. De Luca, L. Snyder-Mackler, M. S. Emley, R. L. Crenshaw, and J. P. Lyons,
Fatigue, recovery, and low back pain in varsity rowers, Med Sci Sports Exerc, vol. 22, no. 4,
pp. 4639, 1990.
[4] R. Merletti and P. A. Parker, Electromyography: physiology, engineering, and non-invasive

applications, vol. 11. John Wiley & Sons, 2004.
[5] M. Cifrek, S. Tonkovi¢, and V. Medved, Measurement and analysis of surface myoelectric
signals during fatigued cyclic dynamic contractions, Measurement, vol. 27, no. 2, pp. 8592,
2000.
[6] J. Petrofsky, Filter bank analyser for automatic analysis of the emg, Medical and Biological
Engineering and Computing, vol. 18, no. 5, pp. 585590, 1980.
[7] L. D. Gilmore and C. J. De Luca, Muscle fatigue monitor (mfm): second generation, Biomed-
ical Engineering, IEEE Transactions on, no. 1, pp. 7578, 1985.
[8] F. B. Stulen and C. J. De Luca, Muscle fatigue monitor: a noninvasive device for observing
localized muscular fatigue, IEEE Transactions on Biomedical Engineering, no. 12, pp. 760
768, 1982.
[9] R. Merletti, D. Biey, M. Biey, G. Prato, and A. Orusa, On-line monitoring of the median
frequency of the surface emg power spectrum, Biomedical Engineering, IEEE Transactions
on, no. 1, pp. 17, 1985.
[10] A. Peyton, Circuit for monitoring the median frequency of the spectrum of the surface emg
signal, IEEE transactions on biomedical engineering, vol. 5, no. BME-34, pp. 391394, 1987.
REFERENCES 7
[11] J. Duchêne and F. Goubel, Acquisition and processing of surface emg signals with a low-cost
microprocessor based system, Journal of biomechanics, vol. 15, no. 10, pp. 791793, 1982.
[12] G. Inbar, O. Paiss, J. Allin, and H. Kranz, Monitoring surface emg spectral changes by the zero
crossing rate, Medical and Biological Engineering and Computing, vol. 24, no. 1, pp. 1018,
1986.
8
Chapter 2
Fundamentals of Electromyography and

Muscle Fatigue Analysis
2.1 Introduction
T
HIS chapter provides a basic introduction to the fundamentals of surface EMG (sEMG) for-
mation. Understanding the mechanisms that create the sEMG signal and how it propagates
along the surface will provide a signicant insight in muscle fatigue. Furthermore, the factors aect-
ing the EMG signal as well as its manifestations in muscle fatigue are presented. The limitations
of current approaches for muscle fatigue monitoring and analysis of existing sEMG methods for
fatigue assessment will guide the approach of this thesis.
2.2 Electromyography
2.2.1 Muscle contraction and EMG formation
The muscle consists of cells called muscle bres. Muscle bres group together into forming motor
units (MU) and many MUs make up a muscle. Every MU has a dedicated motor neuron that excites
it. Hence, all the muscle bres comprising the MU act as one unit, exing or relaxing together. A
diagram of the MU can be seen in Fig. 2.1a. An elastic tissue called epimysium acts as a connective
means covering the muscle and holding it in place upon the large forces generated under contraction.
Similarly, the endomysium surrounds and reinforces every muscle bre cell. Muscle bres contain
cylinders of muscle proteins that allow the muscle to contract, called myobrils. Specically, they
contain two key protein laments, one thin (actin) and one thick (myosin). At either end of the
repeating pattern of the laments there is a connective structure called the Z-line (or Z-disk).
Consider Fig. 2.1b for a detailed diagram of a muscle bre. Running along the Z-lines are small
2.2 Electromyography 9
(a) (b)
Figure 2.1: (a) Muscle anatomy showing motor unit, motor neuron and muscle bre. Figure from
[2], (b) Muscle cell anatomy. Figure from [3].
tubes called T-tubules which are part of the cell membrane (sarcolemma) that extends inside the
bre. Along the bre axis and in-between the T-tubules is a membrane called the sarcoplasmic
reticulum, which stores and releases the calcium ions (Ca

2+ ) that trigger the muscle contraction
[1].
When the muscle is relaxed, the two protein laments, actin and myosin, cannot bind to
each other because they are blocked by two other proteins, tropomyosin and troponin. To contract
a muscle, an electrical impulse from the motor neuron travels to the muscle and when it reaches
the neuron synapse, it releases a chemical message, which in turn causes an action potential in the
muscle cell (AP). The AP travels in the muscle cell though the sarcolema and enters the myobrils
through the T-tubules. Subsequently, the AP triggers the ion gates of the sarcoplasmic reticulum.
As a result, calcium ions are released and ow in the protein laments area (cytoplasm). Calcium
ions bind with troponin-tropomyosin molecules and this causes a shape change which moves them
away from the actin laments. Hence, myosin and actin are free to bond to each other, causing a
contraction. The myosin grabs on actin by forming cross-bridges and pulls actin past them, making
the sarcomere shorter. After the AP has passed, the calcium ion gates close and pumps remove
the Ca2+ from the cytoplasm. Thus, troponin-tropomyosin return to their normal shape and cover
the actin-myosin binding cites. As a result, the muscle relaxes. Consider Fig. 2.2a for a graphical
representation of the actin-myosin binding cycle. In order for this activity to take place, energy is
required. Muscles get energy from adenosine triphosphate (ATP). ATP is made in the mitochondria
inside the muscle bre cells. During an anaerobic respiration, glucose is broken down to lactic acid
and ATP. During an aerobic respiration, glucose, glycogen, fats and amino acids are broken down
with oxygen and ATP is produced. Muscles consist of two basic types of bres; fast twitch and slow
twitch bres. Fast-twitch bres are capable of developing greater forces, contracting faster and have
greater anaerobic capacity. In contrast, slow-twitch bres develop force slowly, they can maintain
contractions longer and they have higher aerobic capacity [1].
So far, the muscle contraction mechanisms are presented. However, for the sliding lament
model to take place, the contraction has to be triggered by a muscle AP. The muscle cell membrane
(sarcolema) is a semi-permeable membrane. When the muscle is relaxed, the membrane pumps
+
sodium ions (N a ) out of the cell and potassium ions (K
+ ) into the cell. Hence, the concentration
of sodium ions outside and potassium ions inside the muscle bre cell is increased. As ions carry a
charge, their concentration dierence generates an electric potential gradient. An ionic equilibrium
between the inner and outer places of a muscle cell forms a resting potential (-70 mV). Upon the
presence of an external chemical stimulus such as a neurotransmitter, the membrane's permeability
to sodium ions locally increases. Hence, N a+ rush into the cell, modifying the membrane potential
to an extent that an excitation threshold is crossed. Beyond that point, sodium ions increasingly
rush in and further raise the membrane potential. At this point, the membrane potential is approx-
imately +30 mV. In the meantime, the shift of the membrane voltage to zero, triggers the opening
of K+ channels, causing the release of potassium ions. The outow of positive charges counteracts
the potential increase due to the inow of sodium ions, so the membrane voltage starts to fall
toward zero. Hence, N a+ channels start to close and membrane potential drops even further, re-
turning to the resting potential. Consider Fig. 2.2b for an illustration of the muscle cell membrane
potential due to N a+ and (K

+) concentrations. Once the membrane threshold is exceeded, the
contraction/relaxations cycle of the muscle begins, based on the sliding lament model. This depo-
larization wave propagates along the surface of the muscle bre and leads to the release of calcium
ions in the intra-cellular space, which trigger a shortening of the muscle cell [1]. This excitation is
the muscle bre action potential. Since all muscle bres inside a MU contract simultaneously, each
MU gives a single AP (MUAP), formed by the superposition of all the muscle bre APs. As more
force is needed, more motor units are recruited and the ring rate is increased. The superposition
of many MUAPs form a bigger wave and as this wave travels, an increasing potential dierence is
measured between the surface electrodes.

Na+ channels K+ channels
Relative membrane permeability

close open
Membrane potential (mV)

+30
ion
Rel
Muscle tension
0
ract
axa
Na+
t
Na+
tion
Con
channels
open K+ Threshold
-55
-70
0 1 2 3 4 5
Time (ms)
(a) (b)
Figure 2.2: (a) The actin-myosin binding cycle. Figure from [3], (b) Muscle cell membrane potential
due to N a+ and (K + ) concentrations. Once the membrane threshold is exceeded, the contraction/re-
laxations cycle of the muscle begins, based on the sliding lament model.
2.2.2 EMG detection
There are two ways to detect and record EMG signals; by using invasive needle-electrodes inserted
in the muscle or non-invasive surface electrodes. With needle electrodes one can identify individual
MUAPs. Since the aim of this thesis is to deliver an easy to use, non-invasive system, only surface
electrodes will be used. The recorded sEMG signal is a superposition of all MUAPs from the
neighbouring MUs close to the skin electrode. The sEMG signal appears random in nature and has
an amplitude distribution in the range of 0-5 mV prior to amplication. Thus, a signal amplier is
required. Typically, a dierential amplier is used as a rst stage amplier [4]. As any biomedical
signal, sEMG is aected by noise. Noise accumulates from several sources; inherent noise of the
electronics used, ambient noise (large oset) and motion artefacts -arising from electrode interference
and electrode cables [4]. The bandwidth of the sEMG signal is between 10 and 500 Hz [5]. The
limited bandwidth of the signal makes it easier to remove motion artefacts, DC osets and high
frequency noise.
The properties of the sEMG signal are related to the biochemical and physiological changes
in muscles during fatiguing contractions. With increasing force levels, an increase in the sEMG
amplitude is observed. As mentioned earlier, one by-product of muscle energy production is lactic
acid. The lactate concentration depends on the type, size and number of muscle bres, force levels,
type of contraction and blood ow. Increasing lactate concentration is responsible for changes in
MOTOR UNIT ACTION POTENTIAL WAVEFORM
MUSCLE FIBRE DEPOLARIZATION ZONE

SPATIAL FILTERING
CONDUCTION VELOCITY OF MUSCLE FIBRES
DIAMETER OF INTRAMUSCULAR MEMBRANE

INTRAMUSCULAR pH
MUSCLE FIBRES TEMPERATURE PROPERTIES
H+ H+
GENERATED REMOVED
FIBRE
DEPTH
ELECTRODE FIBRE TYPE BLOOD FLOW CONSTANT-FORCE

CONFIGURATION
ISOMETRIC
SUSTAINED
FORCE LEVEL CONTRACTION
Figure 2.3: Factors aecting the shape, amplitude and frequency of the detected MUAP on the
skin surface during a sustained contraction. Figure from [6].
the intracellular pH. This in turn aects the Muscle Fibre Conduction Velocity (MFCV), the shape
of the MUAP and hence the shape and frequency of the detected sEMG signal. Moreover, the
electrode size, position and distance from the bres can aect the sEMG signal. Consider Fig. 2.3
for a breakdown of the factors that aect the sEMG signal.
The recorded sEMG signal is a spatially localised recording of the muscles electrical activity
using skin attached electrodes. Thus, sEMG amplitude, power spectrum and quality are highly
dependent upon muscle geometry, electrode position, force levels and skin-electrode contact [5, 7
10]. As a result, electrode conguration and location can aect the reproducibility of measurements
and is still a major barrier on clinical sEMG monitoring [8, 1114]. Moreover, the electrode location
with respect to the lateral edge of the muscle and detection area will determine the amount of
crosstalk from nearby muscles that may be detected by the electrode. As the muscle exes and
extends, the distance between the MUs and the skin electrode constantly changes, while the muscle
bres also change length. Since the tissue between MUs and the skin acts as a volume conductor,
the detected sEMG potential is inversely proportional to this distance (MU - skin distance). The
volume conductor between the surface electrode and the excited MU acts as a spatial low-pass lter.
Hence, MUs closer to the surface would generate a steeper response at higher frequencies than that
of MUs located deeper in the muscle.
One or more electrodes can be used to record the sEMG signal at a single muscle location. A
Figure 2.4: The eect of monopolar, bipolar (single dierential) and double dierential electrode
congurations on the recorded sEMG signal. Figure from [15].
monopolar conguration uses a single electrode and a reference signal, usually the body ground, for
detection and amplication. Although this conguration contains the entire information available
from the detected signal, it is very sensitive to common mode signals, thus it is not suitable for
real-life applications. A bipolar conguration uses two electrodes and is the most widely used con-
guration as it oers increased robustness to common mode variations [8]. Higher order electrode
structures, such as the double dierential conguration (three electrodes) or four electrode cong-
urations (normal double dierentiating conguration) can be used to limit the detection volume,
reduce crosstalk and increase spatial selectivity. Furthermore, literature has shown that a reduction
in the Interelectrode Distance (IED) can limit the detection volume of the electrodes [8].
The use of a monopolar conguration does not have any impact on the recorded sEMG fre-
quency. However, it exhibits poor selectivity and individual MUs cannot be distinguished. Consider
the last trace in Fig. 2.4. The monopolar conguration has a very large pick-up area (illustrated as
half circles). The use of two or more electrode congurations can increase the selectivity and poten-
tially allow for single MU excitations to be isolated and recorded. This is observed in the top three
traces in Fig. 2.4, where the rst two are a result of a double dierential conguration and the third
is the result of a bipolar conguration. As the number of detection cites is increased, the detection
area is decreased. The addition of more recording sites with a dierent weight factor allows for the
creation of spatial lters. Hence, the downside of high selectivity with surface electrodes is that the
spectral components of the sEMG are aected. The inner distance between the surface electrodes
acts as a spatial lter and determines the bandwidth of the dierential conguration. Moreover,
electrode location can change both the amplitude and frequency spectrum of the recorded sEMG
Figure 2.5: Spatial ltering eect of (a) monopolar , (b) bipolar and (c) double dierential electrode
congurations. (left) conguration topology, (middle) spatial lter impulse responce and (right) lter
output signal. Figure from [16].
signal. The use of more than one electrodes can provide more elaborate spatial ltering techniques,
as demonstrated by Reucher et al [15, 16]. Fig. 2.5 shows the dierent connection schemes and
their equivalent spatial transfer function H(fx ) and measured output signals. fx is the spatial
frequency depending on interelectrode distance D. Thus, the use of spatial ltering can aid sEMG
signal analysis. However, the small interelectrode distances limit the use of commercially available
electrodes and require custom-made, not easily available, sensing electrodes.
2.2.3 Standardised sEMG Detection Protocol
Before the 1990s most eorts to understand and use sEMG were taking place locally, hence dierent
research laboratories developed dierent methodologies for sEMG detection. However, the shape,
surface area and distance between the electrodes as well as the placement conguration and location
aect the sEMG signal. As mentioned above, these factors can alter the amplitude and frequency
content of the signal [5]. Hence, dierent sEMG studies could not share results or exchange data.
The Surface Electromyography for Noninvasive Assessment of Muscles (SENIAM) was a European
initiative in 1996 in an eort to standardize sEMG detection procedures. The goal was to provide
guidelines for key variables such as sensors, electrode placement, signal processing and modelling by
examining over 140 journal articles. The following paragraphs summarise the ndings and guidelines
[8]:
Muscle “belly” Muscle
Electrodes
Tendinous Insertion
Innervation Zone
Figure 2.6: Electrode positioning between myotendonous junction and innervation zone.
Muscle “belly” Muscle
Electrodes
Tendinous Insertion
Innervation Zone
Electrode Material
The most common material used is the silver/silver chloride pre-gelled electrode. They are quick
and easy to handle, hygienic, disposable and oer a very low skin-electrode impedance as well as a
exible mechanical layer (gel) that provides a good contact even with electrode or skin movement.
However, steel electrodes can also be used.
Electrode Shape and Size
The shape and size of the electrode refers to the conductive area of the electrodes that is in contact
with the skin. Circular electrodes are more common. SENIAM recommends that sEMG users
should always state the shape and size used in their study. Also, the size of the electrode in the
direction of the muscle bres should not exceed 10 mm. Smaller electrodes are preferred to avoid
excessive low-pass ltering.
Sensor Construction
This refers to the physical connection of the electrodes and cables to the amplier. Any movement
of the construction can create motion artefacts on the sEMG signal. It is recommended to have a
lightweight construction and to rmly x the location of the cables.
Location, Orientation and Interelectrode Distance
Electrodes should be placed between the myotendonous junction and the nearest innervation zone
(neuromuscular junction, located approximately at the middle of the muscle) on the most prominent
bulge of the muscle (muscle belly") and should be oriented perpendicularly to the length of the
muscle bres [5]. Consider Fig. 2.6 for a correct electrode placement.
IED is the centre-to-centre distance between two electrodes in a bipolar conguration. The
recommended IED should be no more than 20 mm. If the muscle is very small, then the IED should
Figure 2.7: Eect of interelectrode distance and electrode location on the sEMG amplitude. Figure
from [21].
be less that 1/4 of the muscle bre length and both electrodes should be on the same side of the
innervation zone.
The recorded sEMG amplitude is a function of IED and electrode placement location. When
electrodes are over the innervation zone, the signal is small and noisy, and carries little information
because of the cancellation eect due to the bidirectional propagation. It is also extremely sensitive
to small displacements between the electrodes [11, 1720]. Consider Fig. 2.7 for an illustration of
the eect of IED and electrode location on the sEMG amplitude. The traces on the top show sEMG
signals that were detected using a bipolar conguration and a linear array. The IED was set to 10
mm and the electrodes were 1 mm thick and 10 mm long. The signal traces on the bottom show
the bipolar sEMG signal obtained by changing the IED to 10, 20, 30, 40, and 50 mm.
Placement Procedure
Proper skin preparation is essential to provide a good, repeatable and reliable skin-electrode contact.
Firstly, hair should be removed from the placement site. An abrasive material should be used to
2.3 Myoelectric Manifestations of Muscle Fatigue 17
remove dead skin cells and medical alcohol can be used to clean the skin from dirt, oils and sweat.
The skin should have a light red colour, indicating a good skin impedance condition.
sEMG Amplication
Due to the limited bandwidth of the signal, band-pass ltering can reduce the eects of added
noise, DC osets and motion artefacts. The dierential electrode conguration was chosen as the
preferred topology for this work, as a balance between common mode immunity, selectivity and
spatial ltering side-eects. The front-end amplier and band-pass ltering recommendations are
listed bellow:
High-pass lter corner frequency ⇒ ≤ 10Hz
Low-pass lter corner frequency ⇒ ∼ 500Hz
Input referred noise ⇒ ≤ 1µVRM S (10-500Hz)
Input impedance (gel electrodes) ⇒ ≥ 100M Ω
Input impedance (dry electrodes) ⇒ ≥ 1000M Ω
Common Mode Rejection Ratio (CMRR) ⇒ ≥ 80dB)
Gain ⇒ Suitable for adequate Analogue to Digital Converter (ADC) resolution (≥ 40dB))
It is clear that high input impedance, high CMRR and low noise are mandatory features of a sEMG
amplier.
2.3 Myoelectric Manifestations of Muscle Fatigue
Although we experience fatigue everyday, dening or quantifying muscle fatigue is not a simple task.
Fatigue is interpreted as a feeling of weakening, muscle pain or performance decrease. However,
these feelings are not suitable for measurement or quantication. Fatigue itself is not a physical
value. Hence, quantifying muscle fatigue is proven to be dicult, as there is no universal index
for it. The inability to maintain a certain muscle force, to perform a certain task or to generate
the same level of Maximum Voluntary Contraction (MVC) force, sometimes associated with pain,
serves as the clinical denition of localised muscle fatigue [22].

As a result, muscle fatigue is dened by measuring physical variables such as force, level of
MVC, power produced upon contraction or angular velocity of a joint. Moreover, it can be dened by
myoelectric (ME) variables such as MU ring rates, conduction velocity, muscle activation, sEMG
amplitude and sEMG spectral estimates. For non-invasive, real-time and realistic applications
measurement of fatigue is based on the analysis of sEMG and not on physical variables. Fatigue
is better evaluated in time, showing progressive changes in the muscle during a contraction. Thus,
fatigue is assessed from the beginning to the end of a muscle contraction [8].
In 1912 Piper was the rst to observe a progressive slowing" of the sEMG during voluntary
static contractions [8, 23]. Due to the random nature of the sEMG, it is easier to quantify this
spectral compression in the frequency domain. In addition, Cobb and Forbes noted an increase of
the sEMG amplitude during static, sub-maximal, fatiguing contractions [8, 24]. These myoelectric
changes precede the mechanical manifestations of muscle fatigue. Although there are several my-
oelectric manifestations of muscle fatigue, there are many diculties for establishing a universal
fatigue index. Firstly, the EMG signal varies in terms of amplitude and frequency between muscles
(size and number of muscle bres) and people. Secondly, some sEMG attributes depend on the force
levels exerted by the muscles. For example, in sub-maximal eort, sEMG signal shifts to higher
amplitudes and lower frequencies as the muscle fatigues. In maximal eort, as seen in Fig. 2.8,
the EMG signal shifts to lower amplitude and again lower frequencies. Due to the force-sEMG
relationship [2528], a high sEMG amplitude is attributed to a strong contraction. Finally, fatigue
manifestations must be observable and reliable in static, as well as dynamic contractions. A static
(isometric) contraction refers to a muscle contraction where there is no movement in the angle of a
joint. In a dynamic (isotonic) contraction, the force generated by the muscle causes movement in the
joint. Static contractions are most commonly used for research purposes, because they can easily
be controlled and assessed. Dynamic contractions better represent real life scenarios and physical
activities. As with any signal, information in the sEMG signal is carried in the time domain and in
the frequency domain. Thus, sEMG signals can be processed in the time domain (amplitude and
timing characteristics) or in the time-frequency domain (frequency content of the ME signal).
It has been reported in literature that neighbouring MUs tend to synchronise and regularly
discharge at or near the time that other motor units discharge [29]. Researchers have observed an
increased tendency of MUAP synchronisation during muscle fatigue [30, 31]. However, the eect
of synchronisation cannot be solely attributed to muscle fatigue. MUAP synchronisation has also
Px
P(W/Hz)
P’x
During
fatigue
0 500
f (Hz)
Figure 2.8: The power spectral density of the sEMG signal during a static, sub-maximal, fatiguing
contraction.
been correlated with physical tremor [32, 33] or in cases where the muscle spindle feedback ( sensory
receptors in the muscle that detect length changes) is impaired, while in [34] it was observed that
synchronisation is more evident in subjects who frequently exert large brief forces. This phenomenon
has not been analysed as extensively as other MU properties. Although it has been observed to be
more evident during fatigue, it is not used as a myoelectric manifestation of muscle fatigue [29].
2.3.1 Processing in Time Domain
Over the years the sEMG amplitude has been studied for its use in the development of algorithms
for myoelectrically controlled limb prosthetics. Furthermore, sEMG amplitude has been used to
study muscle coordination and activation intervals, as well ass in gait analysis to determine muscle
activity. This thesis focuses on the sEMG amplitude properties as they can be indicators for MU
conduction velocities, MU recruitments and other mechanisms occurring during fatigue [6, 8].
2.3.1.1 Root Mean Square, Spike Counting, Zero Crossing Rate
Several amplitude properties of the sEMG signal can be used as indicators of muscle fatigue. The
number of spikes decreases with fatigue and ring rates of MUs increase. The Root Mean Square
(RMS) or the Mean Absolute Value (MAV) of the signal after band-pass ltering and rectication
can be used to track those changes in all contractions [35, 36]. Similarly, the integral of the RMS per
unit time (integrated EMG - iEMG) or a linear envelope of the rectied and low-pass ltered sEMG
(third order Butterworth, fc = 2.5 Hz) [37] can be used. Another amplitude measure that can be
included in the fatigue analysis suite is the signal's peak-to-peak value [38]. Similarly, mean spike
amplitude, mean spike duration, mean spike slope, mean number of peaks per spike and average
number of spikes per unit time also show evidence of correlation with muscle fatigue [39, 40].
The number of zero crossings (signal crossing the baseline) or Zero Crossing Rate (ZCR) is
another measure indicative of muscle fatigue [1]. The ZCR correlates with the spectral compression
of the sEMG signal, however it can be sensitive to the Signal to Noise Ratio (SNR) of the signal.
The number of turns per unit time (T/A), dened as potential reversals more than 100 µV of
the signal, can also aid in muscle fatigue tracking and has been used for distinguishing between
myopathies and neuropathies [41].
The aforementioned time domain processing techniques can be used to analyse sEMG in
both static and dynamic conditions. However, these methods of muscle fatigue tracking depend
the type and eort of the contraction; static/dynamic, maximal/sub-maximal or high/low force
contractions will yield dierent results and aect the accuracy of the estimation of muscle fatigue
[6, 36, 39, 40, 42]. Furthermore, these methods could prove ineective at low force levels. Finally,
in order to extract reliable myoelectric spike information, careful threshold selection is required.
2.3.1.2 Muscle Fibre Conduction Velocity
As the muscle fatigues lactic acid and K+ accumulate in the extracellular muscle space, impairing
the conduction of action potentials across the muscle membrane, thus slowing down MUAPs [36].
Furthermore, fast-twitch muscle bres dominate high force production but get easily fatigued and
drop-out. Following that, fatigue resistant, slow-twitch muscle bres take over. These have smaller
amplitudes since they are located deeper inside the muscle and exhibit slower conduction velocities,
associated with the observed spectral compression. MFCV is a measure of the travelling speed of
MUAPs along their propagation in muscle tissue. MUAPs originate from the innervation zone of
the muscle and will propagate to the tendon region. As the muscle fatigues, MFCV decreases while
the sEMG power spectrum compresses to the left [39].
Lindstrom et al. developed a mathematical model shown in Eq. 2.1, linking MFCV with
the observed EMG power spectrum [43]. In this equation, P (f ) is the EMG power spectrum,

f
v is conduction velocity and G v is the shape of the spectrum of the detected surface action
potential. However, changes in the spectral content of the sEMG signal are disproportionately
larger than decreases in MFCV. Furthermore, recoveries in frequency are more rapid than lactate
removal in the muscle. Thus, MFCV can provide a more detailed insight on muscle fatigue and
muscle recovery [44, 45]. One of the advantages of MFCV is that it is reliable under static and
dynamic contractions [44]. Nevertheless, alignment of two MUAPs is a dicult and potentially
complicated task [42, 46, 47].

1 f
P (f ) = 2 G (2.1)
v v
2.3.2 Processing in Time-Frequency Domain
The Power Spectral Density (PSD) function of the ME signal undergoes a progressive compression
towards lower frequencies and change in shape during fatigue [48, 49]. This change can be attributed
to dierent muscle bre properties. During a fatiguing contraction, force production arises from
fast twitch muscle bres which get easily fatigued. As time progresses, fatigue resistant, slow twitch
muscle bres take over. Action potentials from slow twitch muscle bres have smaller amplitudes
and slower conduction velocities in the muscle tissue, hence lower frequencies are observed. Thus,
many researchers in an eort to quantify muscle fatigue decided to analyse the frequency spectrum of
the EMG signal. The instantaneous Median Frequency (iMDF) and instantaneous Mean Frequency
(iMNF) can be used to track this PSD compression. The iMDF and iMNF are described by Eq. 2.2
and Eq. 2.3 respectively:
Z iM DF (t) Z ∞
S(t, f ) df = S(t, f ) df (2.2)
0 iM DF (t)
Z ∞
f S(t, f ) df
iM N F (t) = Z0 ∞ (2.3)
S(t, f ) df
0
Where S(t, f ) is the time-frequency spectrum. Fig. 2.9 shows the spectral compression and corre-
sponding iMDF the sEMG signal undergoes during fatigue as it progresses in time.
Both iMDF and iMNF can adequately track muscle fatigue under static and dynamic con-
tractions. These spectral measures are able to track changes at low force levels, but the detected
changes get smaller [6, 42], while performance degrades in random conditions [50]. The relative
changes in iMDF and iMNF in static and dynamic contractions are not the same and smaller slopes
are observed in dynamic contractions than in static [44]. The iMDF is less sensitive than iMNF
to modied force levels and amplitude changes, and more accurate in low intensity static and dy-
Power Spectral Density of EMG signal
x 10 −4 −4 −4 −5
4.5 8 x 10 8 x 10 2.5 x 10
4 T = 0s 7 T = 20s 7 T = 40s T = 80s
3.5 2
6 6
sEMG iMDF
Power (W/Hz)
3
5 5 1.5
2.5 sEMG PSD
4 4
2
3 3 1
1.5
1 2 2
0.5
0.5 1 1
0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400 450 0 50 100 150 200 250 300 350 400 450
Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz)
Figure 2.9: Progression of the spectral compresion of the sEMG signal during fatigue. The iMDF
is marked in red.
namic contractions. Hence, iMDF is generally preferred over iMNF [1, 4, 45, 51] and researchers
have established it as the main method of tracking muscle fatigue in conjunction with other sEMG
attributes [6, 36, 38, 40, 52, 53] instead of monitoring output force levels. Lastly, the conduction
velocity is directly related to the diameter of the muscle bres and linked to the observed spectral
compression of the sEMG during fatigue. Hence, iMDF can also reect the size of the muscle -
bres indicating growth, or loss [54] while oering a better noise immunity than MFCV in fatigue
measurements.
2.3.2.1 Short Time Fourier Transform
In order to evaluate iMDF and iMNF, it is necessary to calculate the time-frequency spectrum of the
signal. This can be achieved by applying Short Time Fourier Transform (STFT) which is currently
the most widely used method. STFT is eective under the assumption that the signal is stationary,
meaning that the statistical properties of a process do not change in time. Although the sEMG
signal is a non-stationary process, under constant contraction forces and short time windows (≤ 2s)
the sEMG signal can be considered stationary. Thus, by dividing the sEMG recording in small
segments and estimating the power spectrum for each one, it is possible to monitor the progressive
spectral changes during fatigue. However, in dynamic contractions, changes in force, muscle length
and electrode location have an eect on the sEMG signal [20, 55]. Under these conditions, sEMG
is considered to be a non-stationary signal, hence aecting the time and/or frequency resolution of
the STFT [4, 51].
2.3.2.2 Continuous Wavelet Transform
In another attempt to track spectral variation with time and nd iMDF/iMNF, Continuous Wavelet
Transform (CWT) can be used. CWT is a mathematical tool for local analysis of non-stationary
and fast transient signals [4]. CWT in sEMG signal analysis works under the assumption that the
EMG signal can be considered as the sum of scaled delayed versions of a single prototype. CWT
decomposes the signal into several multiresolution components according to a basis function called
mother wavelet. CWT is described by Eq. 2.4 where x(t) is the sEMG signal and
∗ (t)
ψa,τ is the
mother wavelet. The mother wavelet is described by Eq. 2.5, where α is the wavelet scale [4].
WT can be implemented as a series of band-pass lters with their cuto frequencies representing
the wavelet scales and the impulse of the lters matching the mother wavelet. The relationship
with frequency is shown in Eq. 2.6, where fc is the centre frequency of a wavelet in Hz, fa is
the pseudo-frequency corresponding to scale a in Hz and ∆ is the sampling period. WT oers
good time but poor frequency resolution at high frequencies, and good frequency but poor time
resolution at low frequencies [51]. Hence, CWT could replace STFT for estimating iMDF or iMNF
[6, 51, 52]. The CWT fatigue tracking accuracy can be increased by selecting a mother wavelet
that matches the shape of a typical MUAP. It has been shown that the second order derivative
of a Gaussian distribution, also known as the Mexican hat wavelet, matches closely the MUAP
shape [51]. However, the type of wavelet and scale used play an important role in sEMG analysis,
as there is no single function that is eective for all cases [56]. Finally, the accuracy of CWT in
tracking muscle fatigue compared to STFT is controversial, as some researchers have shown CWT
to be better than STFT in tracking EMG spectral changes [6], while others concluded that both
techniques provide similar information about muscle fatigue [48, 5761]. However, the use of CWT
in sEMG analysis extends beyond muscle fatigue tracking.
Z
∗
CW Tψx (ατ ) = x(t)ψa,τ (t)δt (2.4)

1 t−τ
ψa,τ (t) = √ ψ (2.5)
a a
fc
fa = (2.6)
a·∆
2.3.2.3 Other Time-Frequency Methods
During dynamic contractions the sEMG signal becomes highly non-stationary and its spectral prop-
erties change making it dicult to track with classic spectral techniques. Researchers have pro-
posed dierent methods for time-frequency analysis based on Cohen's class distributions such as
the Wigner-Ville distribution (WVD) and the Choi-Williams distribution (CWD) [6, 6, 48]. These
methods overcomes some limitations of the spectrogram and provides a high resolution representa-
tion in both time and frequency for non-stationary signals. Although these methods are believed
to outperform STFT, there are still no modelling studies showing real improvements in practical
applications [8].
Since all time-frequency approaches aim to show the spectral compression of the signal dur-
ing fatigue, one could quantify fatigue as the ratio of low to high frequencies present, either by
using band-pass lters or selected wavelet scales [52, 62]. This approach can detect a change in
the frequency spectrum of the EMG signal but is not indicative of muscle fatigue, as it cannot
dierentiate fatiguing and non-fatiguing contractions. Careful calibration and selection of cuto
frequencies for every test muscle and subject is required.
2.3.3 Fatigue Index
As observed in the literature, there exist several methods for tracking muscle with varying levels
of accuracy, reliability and adaptability. However, it is very hard to nd one that is reliable under
dierent types of contractions, force levels and a broad subject sample. The myoelectric manifes-
tations of muscle fatigue depend on several factors and involve dierent physiological processes,
all progressing simultaneously. Hence, researchers have tried to combine several fatigue estimation
methods together in an eort to produce a universal and reliable fatigue index. Fig. 2.10 shows the
progression of dierent sEMG variables during a fatiguing contraction as evolved in time.
Recent fatigue monitoring algorithms include Multivariable Mapping Functions (MMF) [50,
64, 65], Principal Component Analysis (PCA) [37, 66] and Articial Neural Networks (ANN) [67].
With the aid of mapping functions, several fatigue tracking methods can be combined together,
such as ZCR, mean absolute value and spectrum estimators [50]. Using an ANN, the network can
learn to map a set of inputs to a set of outputs and it is possible to discover patterns in data which
are not easily detected by other methods.
Although these methods can potentially yield a better fatigue index, they are computation-
ally expensive and are usually congured for each user specically. However, it is clear that one
absolute" metric is not enough to robustly track muscle fatigue. Since the sEMG amplitude and
spectrum depend both on fatigue and force, there is a simplied method to easily evaluate the
Figure 2.10: The progression of dierent sEMG variables in the time and frequency domian during
a fatiguing contraction. Figure from [63].
occurrence of fatigue. Joint Analysis of myoelectric Spectrum and Amplitude (JASA) takes into
consideration changes from both signal processing domains and uses simple logic to form a fatigue
index [6, 68]. Thus, it provides a better insight for muscle fatigue/recovery and force increase/de-
crease [8, 69, 70]. At the same time, JASA is in accordance with the aim of this work to design
a compact, low power, user friendly fatigue monitor, where fatigue estimation takes place locally
on the Application Specic Integrated Circuit (ASIC) and not o-line. The case logic for JASA is
shown bellow:
1. If amplitude increases and spectrum shifts to the right ⇒ Force increase.
2. If amplitude decreases and spectrum shifts to the left ⇒ Force decrease.
3. If amplitude increases and spectrum shifts to the left ⇒ Muscle fatigue.
4. If amplitude decreases and spectrum shifts to the right ⇒ Muscle recovery.
However, the second case can also be attributed to muscle fatigue, if the muscle monitored is under
MVC. Consider Fig. 2.11 for the JASA decision case diagram.
2.4 Current Approaches for Muscle Fatigue Monitoring 26
Amplitude Increase
Muscle Force
Fatigue Increase
Spectral Shift Spectral Shift
Left Right
Force Muscle
Decrease Recovery
Amplitude Decrease
Figure 2.11: Joint Analysis of myoelectric Spectrum and Amplitude decision case algorithm.
2.3.4 Repeatability of sEMG Muscle Fatigue Estimates
The issue of repeatability is of considerable importance for the clinical and repeatable assessment
of sEMG variables. There are two issues to be considered when assessing repeatability of sEMG
measurements and variables; rstly, ensure that data from dierent laboratories and operators are
comparable and secondly, dene the minimum observable variable change that can be attributed to
physiological variations [8]. In order to address the rst issue, SENIAM was formed in an eort to
standardize sEMG detection procedures. The latter issue of repeatability can be addressed using
two complementary aspects. The rst aspects concerns the reliability and constancy within days and
the variability due to repeated trials and electrode positioning. The measure precision is described
by the Normalised Standard Error of the Mean (NSEM). The second aspect concerns the capability
of the variable to distinguish between dierenc subjects and is best described using the Intraclass
Correlation Coecient (ICC) [8]. However, when the within-subject variability is more than the
between-subject variability, then ICC becomes meaningless [8]. Several researchers have addressed
the issue of repeatability of sEMG variables. A comparison between the state of the art on the issue
of repeatability in dierent muscles is shown in Table 2.1.
2.4 Current Approaches for Muscle Fatigue Monitoring
Several researchers proposed muscle fatigue monitor designs in the past. Eorts to create a portable
fatigue monitor date back over three decades. At that time, most of the designs involved either
custom analogue circuits [7882] or a microcontroller based, digital implementation [1, 83]. Due to
Table 2.1: Repeatability of Initial Values of Mean Power Spectral Frequency (MNF),
Average Rectied Value (ARV), and Muscle Fibre Conduction Velocity (MFCV) during
Isometric Voluntary and Electrically Elicited Contractions in Dierent Muscles. Table
recreated from [8].
Muscle MNF ARV MFCV

ICC NSEM ICC NSEM ICC NSEM N Type
(%) (%) (%) (%) (%) (%)
Sternocleidomastoid 43.9 3.0 70.6 16.8 42.5 2.8 9 [71]
Anterior scalenus 74.3 7.15 25.7 13.6 28.0 2.6 9 [71]
Biceps brachii 71.8 4.4 77.7 18.3 <0 1.5 10 [72]
left/right
Longissimus 87.0 73.7 10 [73]
mean
left/right
Multidus 87.5 49.5 10 [73]
mean
left/right
Iliocostalis 88.5 71.0 10 [73]
mean
Rectus femoris 88.7 73.7 18 [74]
left/right
Vastus lateralis 81.5 6.8 80.1 16.3 61.4 7.3 9 [75]
mean
left/right
Vastus medialis 70.1 4.3 73.9 14.7 <0 5.3 9 [75]
mean
electrically
Vastus medialis 85.2 2.3 21.6 7.4 35.7 3.1 9 [76]
elicited
electrically
Tibialis anterior 86.5 78.0 11.5 10 [77]
elicited
the limited resources of the time, the designs only took into consideration one method of muscle
fatigue monitoring and involved bulky equipment and wired approaches.
As computers became more powerful, researchers started exploiting them in order to process
the recorded sEMG in real-time or after collecting a large dataset and measure muscle fatigue [84
88]. As mentioned in Section 2.3.3, with the use of computers, researchers could combine several
sEMG analysis tools and create complex fatigue indexes. Most of the literature used readily available
commercial systems for recording sEMG and a dedicated computer for analysis. Some of the major
commercial sEMG acquisition systems are shown in Table 2.2. These consist of multichannel wired
systems controlled by a computer or wireless devices that temporarily store or transmit the sEMG
to a workstation for analysis. The limitation with wired systems is the physical boundaries placed
by the electrode leads, limiting user freedom of movement and their use in real-life applications. At
the same time, such systems are easier to use and much cheaper than wireless devices.
At the same time, microcontrollers are becoming more power ecient and more capable. In
the case of specialised applications, such robotics or prosthetic limb control, some systems employ
microcontrollers in order to achieve a real-time biofeedback. However, the sEMG signal can only
convey information for the muscle being monitored. In order to track fatigue on dierent muscles,
multiple recording sites are needed. As the number of monitored muscles increases, a bottleneck is
formed by the amount of data that needs to be processed in parallel and in real-time (microcontroller
applications) or transmitted for monitoring (wireless applications). The current solutions investi-
gated in the literature involve the use of several microcontrollers or several transmitting devices to
expand to multiple potential recording sites.
In the past decade researchers have focused their eorts in integrating multiple sEMG sensors
into clothing [8996]. The intended applications for these integrated sensors vary and include, but
not limited to, stress monitoring, fatigue analysis, performance tracking, musculoskeletal disorder
monitoring, injury prevention and prosthetic limb control. By expanding the monitored areas,
the eectiveness of the biofeedback or analysis tool is increased. Integrated sensors show great
potential, however research eorts are concentrated on the integration of the sensors alone and not
the processing mechanism. Collection of the sEMG recoding is performed by external ampliers
and processed in workstations,limiting their use in real-time, unobtrusive applications.
Hence, it is sensible to design an ASIC not only capable of amplifying but also processing the
sEMG signal to measure and monitor muscle fatigue in real-time. Thus, reducing the bandwidth
Table 2.2: Major commercial EMG acquisition systems.

Wireless Systems
DelSys www.delsys.com/
Biometrics www.biometricsltd.com/
Myon www.myon.ch/
Monnit www.monnit.com/
Noraxon www.noraxon.com/
Wired Systems
Bortec Electronics www.bortec.ca/
AD Instruments www.adinstruments.com/
Cadwell www.cadwell.com/
Myotronics www.myotronics.com/
Motion Lab Systems www.motion-labs.com/
and power requirements and increasing the number of recording sites. As a result, researchers or
users can focus on the development of biofeedback applications and not on data collection.
2.4.1 Approach in this Thesis
Although sEMG analysis is a helpful and eective tool for the assessment of muscle fatigue, the signal
changes and their relationship with the physical phenomena taking place during fatigue are not yet
fully understood. Based on the literature review it is clear that both the sEMG amplitude and
spectrum oer valuable information about muscle fatigue. Furthermore, both time and frequency
methods are required, working in tandem to fully asses and reliably track muscle fatigue.
A comparison amongst the presented methods for muscle fatigue tracking is shown in Ta-
ble 2.3, along with their advantages and disadvantages from an ASIC implementation point of view
and the type of study they were used. It is concluded that iMDF is more robust than iMNF as a
fatigue index and it is recognised by many researchers as the golden standard" for muscle fatigue
monitoring. In most of the literature, a myoelectric variable was considered to be a valid fatigue
tracking method if it correlated well with the changes in iMDF in a fatiguing contraction. How-
ever, iMDF alone does not provide a complete understanding of the changes a muscle is undergoing
during a fatiguing contraction. MFCV is a powerful tool oering a better insight about fatigue and
recovery as it depends on the concentration of metabolic byproducts in the muscle. Furthermore,
sEMG MAV or RMS is a helpful tool in assessing the type and strength of the muscle contraction.
Other amplitude parameters, such as spike density, turns and ZCR can oer a look into the eects
of the force-EMG relationship to the observed changes that are interpreted as muscle fatigue.
Another key conclusion are the physical underling dierences the muscle undergoes during
static and dynamic contractions and how these are manifested as sEMG variable changes. In
more detail, intramuscular pressure during a static contraction prevents blood ow and metabolic
byproducts (lactic acid) accumulate in the muscle. The increase of lactic acid decreases the pH in
the muscle and inhibits the excitability of the muscle membrane. Similarly, high concentrations of
potassium ions during a strong, sustained, static contraction inuence the excitability of muscle bre
membrane and aect the MFCV [9799]. In dynamic contractions though, the maintained blood
ow by the stretching and shortening of the muscle facilitates a pumping action that removes the
metabolic byporducts faster. Blood ow also changes the intramuscular pH and the concentration
of K +. Thus, MFCV is enhanced by the progressive removal K+ and lactic acid. Lastly, the oxygen
supply in dynamic contractions is enhanced and the metabolic function in the muscle is improved.
However, MFCV is not the only parameter aected by the type of muscle contraction. It
is observed that in dynamic contractions, the sEMG RMS undergoes little to no change, while the
iMDF is aected signicantly more [100]. On the other hand, MFCV can decrease signicantly in
fatiguing static contractions but not in dynamic contractions [44].
Moreover, force and temperature also aect MFCV, iMDF and sEMG amplitude. At very
low contraction levels (≤ 20% MVC) when the number of action potentials is not big enough
to get sucient summation, standard methods of sEMG analysis can not be used [20]. In long-
lasting, ultra-low contraction levels (≤ 10% MVC) MFCV no longer reects the changes observed
in the power spectrum of the sEMG signal [101], as the changes in iMDF are bigger that those in
MFCV. Lastly, the rate of change of these variable, sometimes overlooked by researchers, can oer
a signicant insight in the physiological changes in the muscle.
Hence, iMDF, MFCV, and several sEMG amplitude variables such as ZCR, MAV and spike
counting are the most promising candidates for fatigue estimation in custom silicon. Moreover, the
relationship of the spectrum compression and the aforementioned sEMG amplitude variables can be
further investigated, as they oer a link between force, tupe of contraction and measured fatigue. It
is concluded that the methods of iMDF, MFCV, mean sEMG amplitude, ZCR and mean number of
spikes will take part in the eort to create a standalone Muscle Fatigue Monitoring (MFM) ASIC.
Table 2.3: Comparison amongst the presented methods for muscle fatigue tracking.
Analysis Method Type of study Disadvantages

Static Dynamic Advantages
Good correlation with fatigue. Sensitive to SNR.
Zero Crossing Rate [1, 36, Sensitive to sEMG changes. Sensitive to sEMG amplitude.
88] Hardware implementable. Threshold selection required.
Hardware implementable. Threshold selection required.
Spike and Peak Counting [39, 40,
Low power approach. Depends on force levels.
102]
Changes in slope depend
Indicative of force levels.
Mean Absolute Value (RMS) [6, 36, [44, 45, on type of contraction.
Hardware implementable.
39, 42, 100, 104, Insensitive at low force levels.
44, 49, 105]
103]
Reliable at low contraction levels.
Reliable on static
Small changes at low force levels.
Instantaneous Median Frequency [4, 6, 36, [44, 48, and dynamic contractions.
Challenging hardware implementation.
42, 44, 100, 105, Wide range of possible implementations.
106] 107, 108] Preferred over iMNF.
µV level potential reversals.
Applications extend beyond
Turns per Amplitude [41] Very sensitive to SNR.
muscle fatigue analysis.
2.4 Current Approaches for Muscle Fatigue Monitoring
Risky hardware implementation.

Better insight on muscle fatigue. More than one channels required.
Muscle Fibre Conduction Velocity [43, 44, [44, 47, Insight on muscle recovery. Susceptible to motion artefacts.
46, 101, 115] Insight on metabolic byproducts. Aected by electrode spatial selectivity.
109114]
Performs similarly to STFT.
Many scales are required.
Wavelet Transform [52, 57, [118] Closely separated corner frequencies.
Reliable under static
116, 117] High selectivity lters required.
and dynamic contractions.
Unable to distinguish
non-fatiguing contractions.
High over Low Frequency Ratio [52] Less lters required.
Not very sensitive.
Relaxed requirements on lter selectivity.
Depends on subject, calibration needed.
31
2.5 Summary 32
2.5 Summary
This chapter provided an overview of the mechanics of muscle contraction, sEMG formation and
their dening role in the interpretation and analysis of muscle fatigue. Although the sEMG mani-
festations of muscle fatigue are have been thoroughly explored, the physiological changes the muscle
undergoes during a fatiguing eort are highly complex and not fully understood yet. Part of the
diculty lies on the nature of sEMG, as it is the superposition of all the underlying motor units.
If detected using a monopolar conguration, due to the very small selectivity of the conguration,
it appears random. Increasing the number of electrodes increases selectivity and reduces crosstalk,
but at the same time alters the spectrum and amplitude of the recorded sEMG signal.
Interestingly, the sEMG signal carries a lot of information about the muscle. Several sEMG
attributes change progressively as the muscle fatigues. However, the sEMG amplitude and spectrum
also depend on muscle type, electrode location, electrode movement and selectivity, force levels, type
of contraction and can vary from person to person. Thus, establishing a single method or index for
fatigue tracking is impossible. Several methods need to be combined in order to achieve that.
Literature has identied the most helpful sEMG signal variables along with the most promis-
ing signal processing techniques for muscle fatigue tracking. Processing in both time domain and
frequency domain is imperative. The most important metrics are found to be the MFCV and the
iMDF of the sEMG signal. Moreover, mean amplitude, spike counting, spike length and other
sEMG amplitude statistics oer an unique look in the type of contraction, force levels and at the
same time attribute in the understating of the origins and manifestations of muscle fatigue.
Currently, a gap exists between sEMG signal processing techniques for fatigue analysis and
methods of sEMG collection and processing. Custom signal processing methods adapted for fatigue
monitoring and implemented in low power CMOS design would bridge that gap.
REFERENCES 33
References
[1] G. Inbar, O. Paiss, J. Allin, and H. Kranz, Monitoring surface emg spectral changes by
the zero crossing rate, Medical and Biological Engineering and Computing, vol. 24, no. 1,
pp. 1018, 1986.
[2] M. Z. Jamal, Signal acquisition using surface emg and circuit design considerations for robotic
prosthesis, 2012.
[3] H. Lodish, Molecular cell biology. Macmillan, 2008.
[4] M. B. Reaz, M. Hussain, and F. Mohd-Yasin, Techniques of emg signal analysis: detection,
processing, classication and applications, Biological procedures online, vol. 8, no. 1, pp. 11
35, 2006.
[5] C. J. De Luca, The use of surface electromyography in biomechanics, Journal of applied

biomechanics, vol. 13, pp. 135163, 1997.
[6] M. Cifrek, V. Medved, S. Tonkovi¢, and S. Ostoji¢, Surface emg based muscle fatigue evalu-
ation in biomechanics., Clinical biomechanics (Bristol, Avon), vol. 24, no. 4, p. 327, 2009.
[7] A. Cechetto, P. Parker, and R. Scott, The eects of four time-varying factors on the mean
frequency of a myoelectric signal, Journal of Electromyography and Kinesiology, vol. 11, no. 5,
pp. 347354, 2001.
[8] R. Merletti and P. A. Parker, Electromyography: physiology, engineering, and non-invasive

applications, vol. 11. John Wiley & Sons, 2004.
[9] D. Farina, W. Jensen, and M. Akay, Introduction to neural engineering for motor rehabilita-
tion, vol. 40. John Wiley & Sons, 2013.
[10] Y. Blanc and U. Dimanico, Electrode placement in surface electromyography (semg)minimal
crosstalk area(mca), Open Rehabil. J, vol. 3, pp. 110126, 2010.
[11] E. Zuniga, X. Truong, and D. Simons, Eects of skin electrode position on averaged elec-
tromyographic potentials., Archives of physical medicine and rehabilitation, vol. 51, no. 5,
pp. 264272, 1970.

REFERENCES 34
[12] J. H. Viitasalo and P. V. Komi, Signal characteristics of emg with special reference to re-
producibility of measurements, Acta Physiologica Scandinavica, vol. 93, no. 4, pp. 531539,
1975.
[13] A. Rainoldi, M. Nazzaro, R. Merletti, D. Farina, I. Caruso, and S. Gaudenti, Geometrical fac-
tors in surface emg of the vastus medialis and lateralis muscles, Journal of Electromyography
and Kinesiology, vol. 10, no. 5, pp. 327336, 2000.
[14] D. Farina, R. Merletti, M. Nazzaro, and I. Caruso, Eect of joint angle on emg variables in
leg and thigh muscles, IEEE engineering in medicine and biology magazine, vol. 20, no. 6,
pp. 6271, 2001.
[15] H. Reucher, G. Rau, and J. Silny, Spatial ltering of noninvasive multielectrode emg: Part i-
introduction to measuring technique and applications, Biomedical Engineering, IEEE Trans-

actions on, no. 2, pp. 98105, 1987.
[16] H. Reucher, J. Silny, and G. Rau, Spatial ltering of noninvasive multielectrode emg: Part
ii-lter performance in theory and modeling, Biomedical Engineering, IEEE Transactions on,
no. 2, pp. 106113, 1987.
[17] T. Masuda, H. Miyano, and T. Sadoyama, The position of innervation zones in the biceps
brachii investigated by surface electromyography, IEEE transactions on biomedical engineer-

ing, no. 1, pp. 3642, 1985.
[18] C. Jensen, O. Vasseljen, and R. H. Westgaard, The inuence of electrode position on bipolar
surface electromyogram recordings of the upper trapezius muscle, European journal of applied
physiology and occupational physiology, vol. 67, no. 3, pp. 266273, 1993.
[19] Z. C. Lateva, N. A. Dimitrova, and G. V. Dimitrov, Eect of recording electrode position
along a muscle bre on surface potential power spectrum, Journal of Electromyography and
Kinesiology, vol. 3, no. 4, pp. 195204, 1993.
[20] S. H. Roy, C. J. De Luca, and J. Schneider, Eects of electrode location on myoelectric
conduction velocity and median frequency estimates, Journal of Applied Physiology, vol. 61,
no. 4, pp. 15101517, 1986.
[21] R. Merletti, A. Rainoldi, and D. Farina, Surface electromyography for noninvasive character-
ization of muscle, Exercise and sport sciences reviews, vol. 29, no. 1, pp. 2025, 2001.
REFERENCES 35
[22] D. B. Chan, Localized muscle fatigue-denition and measurement, Journal of Occupational

and Environmental Medicine, vol. 15, no. 4, pp. 346354, 1973.
[23] H. Piper, Elektrophysiologie menschlicher muskeln springer verlag, 1912.
[24] S. Cobb and A. Forbes, Electromyographic studies of muscular fatigue in man, American
Journal of PhysiologyLegacy Content, vol. 65, no. 2, pp. 234251, 1923.
[25] E. Kuroda, V. Klissouras, and J. Milsum, Electrical and metabolic activities and fatigue in
human isometric contraction., Journal of applied physiology, vol. 29, no. 3, pp. 358367, 1970.
[26] B. Bigland and O. Lippold, The relation between force, velocity and integrated electrical
activity in human muscles, The Journal of physiology, vol. 123, no. 1, p. 214, 1954.
[27] R. Harding and R. Sen, Evaluation of total muscular activity by quantication of electromyo-
grams through a summing amplier, Medical and biological engineering, vol. 8, no. 4, pp. 343
356, 1970.
[28] P. Komi, Relationship between muscle tension, emg and velocity of contraction under con-
centric and eccentric work, in New Concepts of the Motor Unit, Neuromuscular Disorders,
Electromyographic Kinesiology, pp. 596606, Karger Publishers, 1973.
[29] C. J. De Luca, Physiology and mathematics of myoelectric signals, IEEE Transactions on

Biomedical Engineering, no. 6, pp. 313325, 1979.
[30] O. Lippold, J. Redfearn, and J. Vu£o, The rhythmical activity of groups of motor units in
the voluntary contraction of muscle, The Journal of physiology, vol. 137, no. 3, p. 473, 1957.
[31] O. Lippold, J. Redfearn, and J. Vu£o, The electromyography of fatigue, Ergonomics, vol. 3,
no. 2, pp. 121131, 1960.
[32] F. Buchthal and A. Madsen, Synchronous activity in normal and atrophic muscle, Electroen-
cephalography and clinical neurophysiology, vol. 2, no. 1-4, pp. 425444, 1950.
[33] V. Dietz, E. Bischofberger, C. Wita, and H.-J. Freund, Correlation between the discharges of
two simultaneously recorded motor units and physiological tremor, Electroencephalography

and clinical neurophysiology, vol. 40, no. 1, pp. 97105, 1976.
REFERENCES 36
[34] H. Milner-Brown and R. Lee, Synchronization of human motor units: possible roles of exercise
and supraspinal reexes, Electroencephalography and clinical neurophysiology, vol. 38, no. 3,
pp. 245254, 1975.
[35] F. Bai, T. M. Lubecki, C.-M. Chew, and C.-L. Teo, Novel time-frequency approach for muscle
fatigue detection based on semg, in Biomedical Circuits and Systems Conference (BioCAS),
2012 IEEE, pp. 364367, IEEE, 2012.
[36] A. Fuglsang-Frederiksen, The utility of interference pattern analysis, Muscle & nerve, vol. 23,
no. 1, pp. 1836, 2000.
[37] R. Chattopadhyay, M. Jesunathadas, B. Poston, M. Santello, J. Ye, and S. Panchanathan,
A subject-independent method for automatically grading electromyographic features dur-
ing a fatiguing contraction, Biomedical Engineering, IEEE Transactions on, vol. 59, no. 6,
pp. 17491757, 2012.
[38] M. Yochum, T. Bakir, R. Lepers, and S. Binczak, Estimation of muscular fatigue under
electromyostimulation using cwt, Biomedical Engineering, IEEE Transactions on, vol. 59,
no. 12, pp. 33723378, 2012.
[39] D. A. Gabriel, J. R. Basford, and K.-N. An, Assessing fatigue with electromyographic spike
parameters, Engineering in Medicine and Biology Magazine, IEEE, vol. 20, no. 6, pp. 9096,
2001.
[40] O. Dayan, I. Spulber, A. Eftekhar, P. Georgiou, J. Bergmann, and A. McGregor, Applying
emg spike and peak counting for a real-time muscle fatigue monitoring system, in Biomedical
Circuits and Systems Conference (BioCAS), 2012 IEEE, pp. 4144, IEEE, 2012.
[41] J. Finsterer and B. Mamoli, Turn/amplitude parameter changes during sustained eort,
Electroencephalography and Clinical Neurophysiology/Electromyography and Motor Control,

vol. 101, no. 5, pp. 438445, 1996.
[42] R. Merletti, L. Lo Conte, and C. Orizio, Indices of muscle fatigue, Journal of Electromyog-
raphy and Kinesiology, vol. 1, no. 1, pp. 2033, 1991.
[43] D. Farina and R. Merletti, Methods for estimating muscle bre conduction velocity from
surface electromyographic signals, Medical and biological Engineering and Computing, vol. 42,
no. 4, pp. 432445, 2004.
REFERENCES 37
[44] K. Masuda, T. Masuda, T. Sadoyama, M. Inaki, and S. Katsuta, Changes in surface emg
parameters during static and dynamic fatiguing contractions, Journal of Electromyography

[45] J. Potvin and L. Bent, A validation of techniques using surface emg signals from dynamic
contractions to quantify muscle fatigue during repetitive tasks, Journal of Electromyography

[46] D. Farina and R. Merletti, A novel approach for estimating muscle ber conduction veloc-
ity by spatial and temporal ltering of surface emg signals, Biomedical Engineering, IEEE
Transactions on, vol. 50, no. 12, pp. 13401351, 2003.
[47] D. Farina, M. Pozzo, E. Merlo, A. Bottin, and R. Merletti, Assessment of average muscle
ber conduction velocity from surface emg signals during fatiguing dynamic contractions,
Biomedical Engineering, IEEE Transactions on, vol. 51, no. 8, pp. 13831393, 2004.
[48] S. Karlsson, J. Yu, and M. Akay, Time-frequency analysis of myoelectric signals during
dynamic contractions: a comparative study, Biomedical Engineering, IEEE Transactions on,

vol. 47, no. 2, pp. 228238, 2000.
[49] H. Christensen, Muscle activity and fatigue in the shoulder muscles of assembly-plant em-
ployees, Scandinavian journal of work, environment & health, pp. 582587, 1986.
[50] D. T. MacIsaac, P. A. Parker, K. B. Englehart, and D. R. Rogers, Fatigue estimation with
a multivariable myoelectric mapping function, Biomedical Engineering, IEEE Transactions

on, vol. 53, no. 4, pp. 694700, 2006.
[51] F. Laterza and G. Olmo, Analysis of emg signals by means of the matched wavelet transform,
Electronics letters, vol. 33, no. 5, pp. 357359, 1997.
[52] P. J. Sparto, M. Parnianpour, E. A. Barria, and J. M. Jagadeesh, Wavelet and short-time
fourier transform analysis of electromyography for detection of back muscle fatigue, Rehabil-
itation Engineering, IEEE Transactions on, vol. 8, no. 3, pp. 433436, 2000.
[53] T. V. Camata, J. L. Dantas, T. Abrão, M. A. Brunetto, A. C. Moraes, and L. R. Altimari,
Fourier and wavelet spectral analysis of emg signals in supramaximal constant load dynamic
exercise, in Engineering in Medicine and Biology Society (EMBC), 2010 Annual International
Conference of the IEEE, pp. 13641367, IEEE, 2010.
REFERENCES 38
[54] C. J. De Luca, M. A. Sabbahi, and S. H. Roy, Median frequency of the myoelectric signal,
European journal of applied physiology and occupational physiology, vol. 55, no. 5, pp. 457464,
1986.
[55] L. Arendt-Nielsen, N. Gantchev, and T. Sinkjær, The inuence of muscle length on muscle
bre conduction velocity and development of muscle fatigue, Electroencephalography and

Clinical Neurophysiology/Evoked Potentials Section, vol. 85, no. 3, pp. 166172, 1992.
[56] D. K. Kumar, N. D. Pah, and A. Bradley, Wavelet analysis of surface electromyography,
Neural Systems and Rehabilitation Engineering, IEEE Transactions on, vol. 11, no. 4, pp. 400
406, 2003.
[57] J. L. Dantas, T. V. Camata, M. A. Brunetto, A. C. Moraes, T. Abrao, and L. R. Altimari,
Fourier and wavelet spectral analysis of emg signals in isometric and dynamic maximal eort
exercise, in Engineering in Medicine and Biology Society (EMBC), 2010 Annual International
Conference of the IEEE, pp. 59795982, IEEE, 2010.
[58] P. J. Sparto, M. Parnianpour, E. A. Barria, and J. M. Jagadeesh, Wavelet analysis of elec-
tromyography for back muscle fatigue detection during isokinetic constant-torque exertions,
Spine, vol. 24, no. 17, p. 1791, 1999.
[59] I. Hostens, J. Seghers, A. Spaepen, and H. Ramon, Validation of the wavelet spectral esti-
mation technique in biceps brachii and brachioradialis fatigue assessment during prolonged
low-level static and dynamic contractions, Journal of Electromyography and Kinesiology,

vol. 14, no. 2, pp. 205215, 2004.
[60] I. Kremenic, B. Glace, and M. McHugh, Fourier-vs. wavelet-based time-frequency analysis of
fatiguing quadriceps contractions, Medicine & Science in Sports & Exercise, vol. 34, no. 5,
p. S261, 2002.
[61] T. W. Beck, T. J. Housh, G. O. Johnson, J. P. Weir, J. T. Cramer, J. W. Coburn, and
M. H. Malek, Comparison of fourier and wavelet transform procedures for examining the
mechanomyographic and electromyographic frequency domain responses during fatiguing
isokinetic muscle actions of the biceps brachii, Journal of Electromyography and Kinesiol-
ogy, vol. 15, no. 2, pp. 190199, 2005.
REFERENCES 39
[62] J. H. Viitasalo and P. V. Komi, Signal characteristics of emg during fatigue, European
journal of applied physiology and occupational physiology, vol. 37, no. 2, pp. 111121, 1977.
[63] R. Merletti and L. R. L. Conte, Surface emg signal processing during isometric contractions,
Journal of Electromyography and Kinesiology, vol. 7, no. 4, pp. 241250, 1997.
[64] M. Gonzalez-Izal, A. Malanda, I. Rodríguez-Carreño, I. Navarro-Amézqueta, E. Gorostiaga,
D. Farina, D. Falla, and M. Izquierdo, Linear vs. non-linear mapping of peak power using
surface emg features during dynamic fatiguing contractions, Journal of biomechanics, vol. 43,
no. 13, pp. 25892594, 2010.
[65] D. R. Rogers and D. T. MacIsaac, Training a multivariable myoelectric mapping function to
estimate fatigue, Journal of Electromyography and Kinesiology, vol. 20, no. 5, pp. 953960,
2010.
[66] D. R. Rogers and D. T. MacIsaac, Emg-based muscle fatigue assessment during dynamic con-
tractions using principal component analysis, Journal of Electromyography and Kinesiology,

vol. 21, no. 5, pp. 811818, 2011.
[67] D. Moshou, I. Hostens, G. Papaioannou, and H. Ramon, Dynamic muscle fatigue detection
using self-organizing maps, Applied Soft Computing, vol. 5, no. 4, pp. 391398, 2005.
[68] I. Hostens and H. Ramon, Assessment of muscle fatigue in low level monotonous task per-
formance during car driving, Journal of Electromyography and Kinesiology, vol. 15, no. 3,
pp. 266274, 2005.
[69] A. Luttmann, M. Jäger, and W. Laurig, Electromyographical indication of muscular fatigue
in occupational eld studies, International journal of Industrial ergonomics, vol. 25, no. 6,
pp. 645660, 2000.
[70] A. Luttmann, M. Jäger, J. Sökeland, and W. Laurig, Joint analysis of spectrum and amplitude
(jasa) of electromyograms applied for the indication of muscular fatigue among surgeons in
urology, in Proceedings of the XI th Annual International Occupational Ergonomics and Safety

Conference, International Society for Occupational Ergonomics and Safety, Cincinnati, 1996.
[71] D. Falla, P. Dall'Alba, A. Rainoldi, R. Merletti, and G. Jull, Repeatability of surface emg
variables in the sternocleidomastoid and anterior scalene muscles, European journal of applied
physiology, vol. 87, no. 6, pp. 542549, 2002.
REFERENCES 40
[72] A. Rainoldi, G. Galardi, L. Maderna, G. Comi, L. L. Conte, and R. Merletti, Repeatability
of surface emg variables during voluntary isometric contractions of the biceps brachii muscle,
[73] J. H. van Dieën and P. Heijblom, Reproducibility of isometric trunk extension torque, trunk
extensor endurance, and related electromyographic parameters in the context of their clinical
applicability, Journal of Orthopaedic Research, vol. 14, no. 1, pp. 139143, 1996.
[74] J. Kollmitzer, G. R. Ebenbichler, and A. Kopf, Reliability of surface electromyographic mea-
surements, Clinical Neurophysiology, vol. 110, no. 4, pp. 725734, 1999.
[75] A. Rainoldi, J. Bullock-Saxton, F. Cavarretta, and N. Hogan, Repeatability of maximal
voluntary force and of surface emg variables during voluntary isometric contraction of quadri-
ceps muscles in healthy subjects, Journal of Electromyography and Kinesiology, vol. 11, no. 6,
pp. 425438, 2001.
[76] R. Merletti, A. Fiorito, L. Lo Conte, and C. Cisari, Repeatability of electrically evoked emg
signals in the human vastus medialis muscle, Muscle & nerve, vol. 21, no. 2, pp. 184193,
1998.
[77] R. Merletti, L. L. Conte, and D. Sathyan, Repeatability of electrically-evoked myoelectric
signals in the human tibialis anterior muscle, Journal of Electromyography and Kinesiology,
vol. 5, no. 2, pp. 6780, 1995.
[78] J. Petrofsky, Filter bank analyser for automatic analysis of the emg, Medical and Biological
[79] L. D. Gilmore and C. J. De Luca, Muscle fatigue monitor (mfm): second generation, Biomed-
ical Engineering, IEEE Transactions on, no. 1, pp. 7578, 1985.
768, 1982.
on, no. 1, pp. 17, 1985.
REFERENCES 41
[83] J. Duchêne and F. Goubel, Acquisition and processing of surface emg signals with a low-cost
microprocessor based system, Journal of biomechanics, vol. 15, no. 10, pp. 791793, 1982.
[84] S. Karlsson, B. Erlandson, and B. Gerdle, A personal computer-based system for real-time
analysis of surface emg signals during static and dynamic contractions, Journal of Elec-
tromyography and Kinesiology, vol. 4, no. 3, pp. 170180, 1994.
[85] B. Karagözoglu, W. H. Sindi, and A. A. Al-Omari, Design and development of a practical
muscle fatigue monitor, in GCC Conference & Exhibition, 2009 5th IEEE, pp. 17, IEEE,
2009.
[86] S. Gutiérrez, E. Cardiel, and P. Hernández, A muscle fatigue monitor based on the surface
electromyography signals and frequency analysis, in 2016 Global Medical Engineering Physics
Exchanges/Pan American Health Care Exchanges (GMEPE/PAHCE), pp. 16, IEEE, 2016.
[87] M. M. Jankovi¢ and D. B. Popovi¢, An emg system for studying motor control strategies
and fatigue, in Neural Network Applications in Electrical Engineering (NEUREL), 2010 10th
Symposium on, pp. 1518, IEEE, 2010.
[88] G. Hägg, Electromyographic fatigue analysis based on the number of zero crossings, Pügers
Archiv European Journal of Physiology, vol. 391, no. 1, pp. 7880, 1981.
[89] R. Manero, J. Grewal, B. Michael, A. Shafti, K. Althoefer, J. L. Fernandez, M. Howard,
et al., Wearable embroidered muscle activity sensing device for the human upper leg, arXiv
preprint arXiv:1602.04841, 2016.
[90] N. Lintu, J. Holopainen, and O. Hänninen, Usability of textile-integrated electrodes for emg
measurements, University of Kuopio, Department of Physiology, Laboratory of Clothing Phys-

iology, Kuopio, Finland, 2005.
[91] T. Linz, L. Gourmelon, and G. Langereis, Contactless emg sensors embroidered onto textile,
in 4th International Workshop on Wearable and Implantable Body Sensor Networks (BSN
2007), pp. 2934, Springer, 2007.
REFERENCES 42
[92] Y.-C. Ma, Y.-P. Chao, and T.-Y. Tsai, Smart-clothesprototyping of a health monitoring
platform, in 2013 IEEE Third International Conference on Consumer Electronics¾¾ Berlin

(ICCE-Berlin), pp. 6063, IEEE, 2013.
[93] T. Finni, M. Hu, P. Kettunen, T. Vilavuo, and S. Cheng, Measurement of emg activity with
textile electrodes embedded into clothing, Physiological measurement, vol. 28, no. 11, p. 1405,
2007.
[94] J. Taelman, T. Adriaensen, C. van der Horst, T. Linz, and A. Spaepen, Textile integrated
contactless emg sensing for stress analysis, in 2007 29th Annual International Conference of
the IEEE Engineering in Medicine and Biology Society, pp. 39663969, IEEE, 2007.
[95] D. Farina, T. Lorrain, F. Negro, and N. Jiang, High-density emg e-textile systems for the con-
trol of active prostheses, in 2010 Annual International Conference of the IEEE Engineering
in Medicine and Biology, pp. 35913593, IEEE, 2010.
[96] G. Li, Y. Geng, D. Tao, and P. Zhou, Performance of electromyography recorded using
textile electrodes in classifying arm movements, in 2011 Annual International Conference of

the IEEE Engineering in Medicine and Biology Society, pp. 42434246, IEEE, 2011.
[97] J. T. Mortimer, R. Magnusson, and I. Petersen, Conduction velocity in ischemic muscle: eect
on emg frequency spectrum, American Journal of PhysiologyLegacy Content, vol. 219, no. 5,
pp. 13241329, 1970.
[98] L. Arendt-Nielsen, K. R. Mills, and A. Forster, Changes in muscle ber conduction velocity,
mean power frequency, and mean emg voltage during prolonged submaximal contractions,
Muscle & nerve, vol. 12, no. 6, pp. 493497, 1989.
[99] N. Radicheva, L. Gerilovsky, and A. Gydikov, Changes in the muscle bre extracellular
action potentials in long-lasting (fatiguing) activity, European journal of applied physiology

and occupational physiology, vol. 55, no. 5, pp. 545552, 1986.
[100] V. Singh, D. Kumar, D. Djuwari, B. Polus, S. Fraser, J. Hawley, and S. Giudice, Strategies
to identify muscle fatigue from semg during cycling, in Intelligent Sensors, Sensor Networks
and Information Processing Conference, 2004. Proceedings of the 2004, pp. 547551, IEEE,
2004.
REFERENCES 43
[101] G. Caer, D. Heinecke, and R. Hinterthan, Surface emg and load level during long-lasting
static contractions of low intensity, International Journal of Industrial Ergonomics, vol. 12,
no. 1-2, pp. 7783, 1993.
[102] B. Bigland-Ritchie, R. Johansson, O. C. Lippold, S. Smith, and J. J. Woods, Changes in
motoneurone ring rates during sustained maximal voluntary contractions., The Journal of
Physiology, vol. 340, no. 1, pp. 335346, 1983.
[103] M. HAOBERG and B. Jonsson, The amplitude distribution of the myoelectric signal in an
ergonomic study of the deltoid muscle, Ergonomics, vol. 18, no. 3, pp. 311319, 1975.
[104] A. Holtermann and K. Roeleveld, Emg amplitude distribution changes over the upper trapez-
ius muscle are similar in sustained and ramp contractions, Acta physiologica, vol. 186, no. 2,
pp. 159168, 2006.
[105] M. Gonzalez-Izal, A. Malanda, I. Navarro-Amezqueta, E. Gorostiaga, F. Mallor, J. Ibanez, and
M. Izquierdo, Emg spectral indices and muscle power fatigue during dynamic contractions,
[106] R. Merletti, M. A. Sabbahi, and C. J. De Luca, Median frequency of the myoelectric signal,
1984.
[107] J. B. Dingwell, J. E. Joubert, F. Diefenthaeler, J. D. Trinity, et al., Changes in muscle activity

and kinematics of highly trained cyclists during fatigue, IEEE Transactions on Biomedical
Engineering, vol. 55, no. 11, pp. 26662674, 2008.
[108] M. Cifrek, S. Tonkovi¢, and V. Medved, Measurement and analysis of surface myoelectric
signals during fatigued cyclic dynamic contractions, Measurement, vol. 27, no. 2, pp. 8592,
2000.
[109] D. Farina, L. Arendt-Nielsen, R. Merletti, and T. Graven-Nielsen, Assessment of single motor
unit conduction velocity during sustained contractions of the tibialis anterior muscle with
advanced spike triggered averaging, Journal of neuroscience methods, vol. 115, no. 1, pp. 1
12, 2002.
[110] M. Zwarts, T. Van Weerden, and H. Haenen, Relationship between average muscle bre
conduction velocity and emg power spectra during isometric contraction, recovery and applied
REFERENCES 44
ischemia, European journal of applied physiology and occupational physiology, vol. 56, no. 2,
pp. 212216, 1987.
[111] M. M. Lowery, C. L. Vaughan, P. J. Nolan, and M. J. O'Malley, Spectral compression of the
electromyographic signal due to decreasing muscle ber conduction velocity, Rehabilitation

Engineering, IEEE Transactions on, vol. 8, no. 3, pp. 353361, 2000.
[112] D. Farina, E. Fortunato, and R. Merletti, Noninvasive estimation of motor unit conduction
velocity distribution using linear electrode arrays, IEEE Transactions on Biomedical Engi-
neering, vol. 47, no. 3, pp. 380388, 2000.
[113] S. W. Davies and P. A. Parker, Estimation of myoelectrc conduction yelocity distribution,
IEEE Transactions on Biomedical Engineering, vol. 5, no. BME-34, pp. 365374, 1987.
[114] B. Bigland-Ritchie, E. Donovan, and C. Roussos, Conduction velocity and emg power spec-
trum changes in fatigue of sustained maximal eorts, Journal of Applied Physiology, vol. 51,
no. 5, pp. 13001305, 1981.
[115] D. Farina and D. Falla, Estimation of muscle ber conduction velocity from two-dimensional
surface emg recordings in dynamic tasks, Biomedical Signal Processing and Control, vol. 3,
no. 2, pp. 138144, 2008.
[116] X. Zhang, Y. Wang, and R. P. Han, Wavelet transform theory and its application in emg signal
processing, in Fuzzy Systems and Knowledge Discovery (FSKD), 2010 Seventh International
Conference on, vol. 5, pp. 22342238, IEEE, 2010.
[117] M. Pope, A. Aleksiev, N. Panagiotacopulos, J. Lee, D. Wilder, K. Friesen, W. Stielau, and
V. Goel, Evaluation of low back muscle surface emg signals using wavelets, Clinical Biome-
chanics, vol. 15, no. 8, pp. 567573, 2000.
[118] M. González-Izal, I. Rodríguez-Carreño, F. Mallor-Giménez, A. Malanda, and M. Izquierdo,
New wavelet indices to assess muscle fatigue during dynamic contractions, World Academy
of Science Engineering and Technology, vol. 55, p. 480, 2009.
45
Chapter 3
An ASIC for sEMG Spectrum

Monitoring using a Median Frequency
Estimator
3.1 Introduction
T
HE Power Spectral Density (PSD) of the EMG signal undergoes a progressive compression
towards lower frequencies and change of shape during fatigue [1]. As presented in Chap-
ter 2, the instantaneous median frequency (iMDF) is an ideal observation tool to track the PSD
compression the surface EMG (sEMG) signal undergoes during fatigue. In sub-maximal eorts
the sEMG signal shifts to higher amplitudes and lower frequencies as the muscle fatigues while in
maximal eorts, as seen in Fig. 3.1a, the sEMG signal shifts to lower amplitude and again lower
frequencies. Fig. 3.1b shows the iMDF decreasing during a voluntary static fatiguing contraction
and was calculated using the MATLAB

® medf req() function.
This chapter presents two real-time systems in CMOS for tracking the iMDF from surface
EMG signals. The rst implementation has been designed in full custom analogue to achieve very
low power consumption, while minimizing computational complexity. The second implementation
explored a mixed signal approach in order to increase accuracy and power eciency.
3.2 Methods for real-time iMDF tracking systems
The iMDF can be described as the frequency at which the PSD of the signal is divided into two
equal parts in a specied time window. iMDF is described by Eq. 3.1:

3.2 Methods for real-time iMDF tracking systems 46
sEMG Amplitude (V)

0.5
-0.5
Px
P(W/Hz)
-1
0 20 40 60 80 100 120 140 160
P’x 100
During
iMDF (Hz)
fatigue 50
0
2 2.25 0 500 0 20 40 60 80 100 120 140 160
f (Hz) Time (s)
(a) (b)
Figure 3.1: (a) Power Spectral Density function of EMG signal during fatigue, (b) top: sEMG
signal from static fatiguing cobntraction, bottom: iMDF of the recorded sEMG signal.
Z iM DF (t) Z ∞
S(t, f ) df = S(t, f ) df (3.1)
0 iM DF (t)
where S(t, f ) is the time-frequency spectrum. The most common and less complicated method to
estimate the iMDF of an sEMG recording is processing it using a computer or a microcontroller.
The signal is sampled, stored and split into processing windows (epochs). The PSD is calculated
using Discrete Fourier Transform (DFT) of every epoch. Finally, integration of the PSD will result
in an iMDF measurement. As the aim of this work is to design a miniaturised muscle fatigue
monitor, dierent analogue and mixed signal approaches for spectrum estimation and subsequently
iMDF tracking are explored.
Continuous, real-time iMDF tracking can be achieved using either closed-loop or open-loop
systems. An open loop approach would require the use of several high Q-factor lters, evenly
separated across the sEMG spectrum. By isolating a specied frequency band with each lter,
they can provide a crude approximation of the spectral density in each band. The power output of
each lter can subsequently be processed, integrated and averaged in order to estimate the iMDF,
similar to [2]. Although open-loop approaches oer fast, stable tracking, accuracy depends on the
number of lters used to split the sEMG spectrum into processing windows. More importantly,
spectral overlap of the lters and subsequent need for calibration of each corner frequency reduce
the eective accuracy and tracking properties of the system.
Closed-loop approaches for iMDF tracking involve partial power estimation of the sEMG
3.2 Methods for real-time iMDF tracking systems 47
P(W/Hz)
Power of original sEMG signal
Power of low-pass filtered signal

PEMG Power of high-pass filtered signal
Plow = Phigh
iMDF
0 500
f (Hz)
Figure 3.2: Partitioning the sEMG spectrum with the use of two tunable lters (high-pass and
low-pass) with the same corner frequency.
spectrum and continuous integration for tracking. Systems proposed in [24] use tunable lters to
selectively estimate the power of a portion of the sEMG signal. The estimated power is equal to half
the total power at a frequency equal to iMDF. Partitioning the sEMG spectrum can be accomplished
by suing one tunable lter (high-pass or low-pass) or two tunable lters (high-pass and low-pass)
with the same corner frequency. Consider Fig. 3.2 for a graphical representation of the method
described. Algorithms reported in literature follow the a system architecture similar to the one
reported in [3] and shown in Fig. 3.3, with small variations. Nevertheless, high order tunable lters
are hard to design and accurately control. Thus, past implementations have reduced the lter order
with an impact of tracking accuracy due to spectral overlap" or spectral leakage" of the lters.
Use of smaller order lters would result in small dierences between the two feedback branches, thus
increasing the uncertainty due to noise and mismatch of the subtracter and integrator. The solution
involves a dierent approach where a high order brick wall" lter with a xed corner frequency
is used. In order to partition the sEMG spectrum, the baseband signal is up-modulated using a
tunable oscillator. This approach would result to a more accurate control and selective spectrum
integration. This approach was proposed by Merletti et al[5] and was implemented in a digital
INTEL 2920 microcontroller. This system architecture can be adapted for low power Application
Specic Integrated Circuit (ASIC) implementation and integration with other sEMG processing
tools to form a standalone muscle fatigue monitor.

3.3 Proposed IC iMDF Tracking System 48
Control Voltage Output

(proportinal to iMDF)
V+ Voltage
A Automatic Integrator
Controlled
V- Gain Control
Filter
High Low
Pass Pass
2
X
2
X
Figure 3.3: System architecture for the estimation of iMDF, utilising tunable lters and a feedback
controller. Recreated from [3].
Band Pass Filter Original sEMG spectrum
Modulated sEMG spectrum

fc (low)-fVCO Power of filtered signal
=
P(W/Hz)
iMDF
Px 1 Px P= 18 Px
4
500 fVCO fc (centre) 3 fVCO
f (Hz)
Figure 3.4: iMDF estimation algorithm: The power of the up-modulated sEMG signal is partitioned
using a high order bandpass lter. The dierence in power of the original singal and the up-modulated
one is integrated and used to control the VCO frequency.
3.3 Proposed IC iMDF Tracking System
The iMDF is computed by progressively up-modulating the EMG signal to higher frequencies track-
ing its PSD [5]. Consider Fig. 3.4 for a graphical representation of the algorithm described. The
up-modulation is accomplished using a Voltage Controlled Oscillator (VCO) and a xed frequency
lter, therefore monitoring the frequency content of the signal. The instantaneous power of the sig-
nal is integrated in order to estimate the iMDF. The PSD of the up-modulated signal folds around
the VCO's carrier frequency fV CO . The resulting PSD is given by Eq. 3.2:
1 1
Px = Sx (f − fV CO ) + Sx (f + fV CO ) (3.2)
4 4
where Sx (f ) is the original power spectrum of the signal and Px (f ) is the power spectrum of the
Anti 1/8
Aliassing Filter Squarer Vout
Differential
SC Subtracting
Amplifier VCO G2
Integrator Amplifier
Analog Front End
Band-Pass G1
Chopper Elliptic Filter Squarer
iMDF Tracker
Figure 3.5: System-level block diagram of the chosen iMDF tracking implementation.
modulated input signal. The lower side-band is not required, hence not included in any subsequent
calculations. The cuto of the high-pass lter will be at a frequency, fc , higher that that of
1
fV CO . iMDF is found as the dierence between fc and fV CO when Px = 8 Sx (f + fV CO ). The
power (i.e. square) of the ltered signal is then subtracted from one eighth of the input signal's
power. The dierence in power is integrated and used to adjust the frequency of the VCO. When
1
2 Sx (f + fV CO ) = 81 Sx , the power dierence will be zero, hence fc − fV CO will be the iMDF of the
input signal. In order to eliminate any harmonic distortion due to nonidealities of the sine wave
VCO, a square wave is used to up-modulate the sEMG signal and a bandpass lter to lter out the
odd harmonics.
Consider Fig. 3.5 for a system block diagram representational of the iMDF tracker. The
input signal is amplied and ltered to the EMG signal bandwidth (10-500 Hz). The modulation
frequency fV CO must be higher than the bandwidth of the signal (i.e. fV CO > BWEM G ) to avoid
folding of the lower side-band and aecting the power estimation.
The principle behind this system relies on subtracting the ltered and modulated sEMG
signal power from the original sEMG signal power. The higher lter order and faster rollo used,
the bigger the dierence between the two signal paths will be. Thus, the lter rollo has an impact
on the tracking accuracy of the system. An elliptic lter has a much higher rollo than any other
lter, hence it is the best choice for this design. The rollo of an elliptic lter depends on its order
and its cut-o frequency. For a band-pass elliptic lter as the cut-o moves to higher frequencies,
in order to keep the same rollo, the number of orders must be increased. MATLAB
® simulations
showed that the best performance of a high-pass elliptic lter is between 1000 and 2000 Hz, while
fV CO > BWEM G . With a cut-o frequency in that region, the lter's stop-band frequency fs
(where attenuation is -40 dB) would be less than 250 Hz bellow the cut-o. Assuming a maximum
B
1/8
Product1 Gain2
butter E F
A -K- 1
data s
From Gain1 Gain4 Ideal Integrator
Workspace ellip D
C -G-
Product Product2 Gain3
Continuous-Time
butter VCO
G
simout
To Workspace 1450
Constant
Figure 3.6: Simulink model for tracking iMDF. The signal passes through an antialliasing lter and
then proccesed by the algorithm described above. The output of the system is the dierence between
fc and fV CO . The output is low-pass ltered for a smoother response.
iMDF frequency of 150 Hz, and a lter cut-o frequency of 1450 Hz, then the starting frequency of
the VCO will be 1300 Hz. At this frequencies, assuming a maximum iMDF frequency of 150 Hz,
fV CO is still higher than the sEMG bandwidth.
The system was rst designed and validated in SIMULINK

® using retrospective data col-
lected at Charing Cross hospital.

1 Data from back muscles (LES T12, RES T12, LES L4 and RES
L4) were collected while the muscles were fatiguing, starting from a non-fatigued state, up until
exhaustion. A complete model of the system is presented in Fig. 3.6. The input signal is fed through
an anti-aliasing lter. Then, it is processed by the iMDF tracker algorithm described above. In this
model, an ideal sine wave VCO and high-pass elliptic lter were used to minimise complexity. The
output is given by fc − fV CO . However, it is low-pass ltered by a sub-hertz lter, since iMDF can
vary signicantly and change abruptly in time.
Simulations using this model revealed that a 6th order elliptic lter is sucient for accu-
rate iMDF tracking. Although increasing the lter order yields better tracking accuracy, there is a
tradeo between lter order and circuit power consumption, silicon area and implementation com-
plexity. However, careful analysis showed that the leakage" form the lter had a minor eect on
the output iMDF. It resulted in a noticeable shift the iMDF up by a xed oset of 55 Hz, which
could easily be accounted for and removed. Gain term G (Gain4) is used to counter-balance any
leakage eects when the hardware implementation is designed. The following parameter values were
used in subsequent simulations: K = 250, G = 1, Filter order = 6, fc = 1450 Hz, fvcostart = 1300
Hz, VCO sensitivity = 100 Hz

V .
1 The author would like to acknowledge Paul Strutton for collecting and providing the mentioned datasets.
3.3.0.1 iMDF Model Analysis
As in any feedback system, it is helpful to know the system's time constant. The following analysis
makes the assumption that the lter used is ideal, in order to simplify the calculations. The analysis
was based on the work of Stulen et al in [4]. Consider the following denitions:
1. fhi ⇒ Bandwidth of input sEMG. iMDF and integrator output ( note: 1/sen-
sitivity).
2. fh ⇒ fhi + fV CO .
7. K ⇒ Integrator gain.
3. A⇒ Arbitrary signal level (V /Hz ).
4. fmedi ⇒ iMDF.
8. G⇒ Feedback gain.
5. fmed ⇒ fmedi + fV CO . 9. τi ⇒ Integrator time constant.
6. Kf ⇒ Proportionality constant between 10. fV COi ⇒ Fundamental VCO frequency.
As explained in the section above, The dierence between two power measures is integrated and
used to control a VCO. The input signal power (+) is shown bellow:
fhi A2
= Px . (3.3)
8
The ltered, up-modulated signal power (-) is as follows:
A2 (fh − fmed ) = GA2 (fhi + fV CO − (fmedi + fV CO )).

(3.4)
GA2 (fhi − fmedi ) = Py .
The integrator output is also the output of the system Vout and is given by:
Z
1
Vout = K(Px − Py )dt.
τi
(3.5)
dVout
τi = K(Px − Py ).
dt
VCO frequency is related to Vout by:


Vout
fV CO = + fV COi .
Kf
(3.6)
Vout = Kf (fV CO − fV COi ).
Substituting Eq. 3.3, 3.4 and 3.6 in Eq. 3.5:

d Kf (fV CO − fV COi )
fhi
τi = KA2 − G(fhi − fmedi ) . (3.7)
dt 8
Given that fc − fmed = fV CO , Eq. 3.7 becomes:

dfmedi 1
= KA2 Gfmedi +

−Kf τi − G fhi .
dt 8
(3.8)
1 dfmedi
KGA2 fmedi = KA2 (G − )fhi − Kf τi .
8 dt
Applying Laplace Transform to Eq. 3.8:
1
KGA2 Fmedi (s) = KA2 G − Fhi (s) − Kf τi sFmedi (s) + Kf τi fmed0 .
8
KA2 (G − 18 )Fhi (s) + Kf τi fmed0
Fmedi (s) = .
KGA2 + Kf τi s
(3.9)
1 Kf τi
1− Fhi (s) + fmed0
8G KGA2
Fmedi (s) = .
Kf τi
1+ s
KGA2
Where fmed0 is the starting iMDF of the system. It is observed that it behaves like a rst order
system. Finally, the systems time constant τsys is given by:
Kf τi
τsys = . (3.10)
KGA2
1
The feedback gain term G is much smaller than integrator gain K and VCO sensitivity
Kf . Thus,
from Eq. 3.10 it is observed that decreasing integrator gain and/or VCO sensitivity will increase
the time constant of the system. This will result in a slower" tracking of iMDF. Thus, the output
will appear smoother. However, the system will require more time to reach the present iMDF value
3.4 iMDF Tracker Circuit Implementation 53
of the signal. Furthermore, increasing feedback gain will shift the iMDF output signal to higher
frequencies. The system's time constant will decrease, resulting in larger iMDF variations.
3.3.0.2 iMDF Model Simulations
The retrospective data used to test the system were collected from back muscles LES T12, RES
T12, LES L4 and RES L4. The muscle were initially un-fatigued and then slowly fatigued until
exhaustion. Fig. 3.7 shows a 250 ms sEMG waveform and the corresponding signal transformations
as it goes through nodes B-F of the system model in Fig. 3.6. Fig. 3.8 shows the simulated iMDF for
these muscles. The simulation time is 151 seconds and during that time, as the muscle fatigues, the
iMDF decreases. The blue line is the estimated iMDF by the presented system. The red line is an
estimation of iMDF using Short Time Fourier Transform (STFT) in MATLAB

® . Small dierences
between the two lines can be observed in some cases. However, these dierences are 5 Hz or less.
Assuming a maximum iMDF that the system can estimate of 150 Hz, then the error would be less
than 3.5%. Moreover, red line iMDF was estimated in 10% time intervals. Due to the noisy nature
of the signal (instantaneous changes are expected), the red line is less of an absolute measurement
and more of a guide. In Fig. 3.8, iMDF under fatigue of muscle LES T12 is shown in (a), RES
T12 in (b), LES L4 in (c) and RES L4 in (d). Sub-plots (e) and (f ) show the iMDF of the muscles
LRA and RRA respectively. These muscles were either not engaged or did-not fatigue during the
exercise. Hence, iMDF changes are very small. As previously mentioned, the presented estimations
of iMDF had an average oset of 55 Hz which was removed.
3.4 iMDF Tracker Circuit Implementation
The proposed iMDF tracker was implemented in a commercially available 0.35µm CMOS technology
provided by AMS (C35B4).
VCO
The topology used is a gm − C variable oscillator. The circuit requires three Operational Transcon-
ductance Ampliers (OTAs) and two capacitors. The circuit schematic is shown in Fig. 3.9. The
condition for oscillation is the following:
gm3 − gm2 = 0, C1 = C2 = C. (3.11)

a) b) c)
0.4 0.014 0.4
0.3
A B 0.3
C
0.012
0.2 0.2
0.01
Amplitude (V)
0.1 0.1
0.008
0 0
0.006
-0.1 -0.1
-0.2 0.004
-0.2
-0.3 0.002 -0.3
-0.4 0 -0.4
0 250 0 250 0 250
Time (ms) Time (ms) Time (ms)
d) e) f)
0.014 0.015 0.03
D E F
0.012 0.01 0.025
0.01 0.02
0.005
Amplitude (V)
0.008 0.015
0
0.006 0.01
-0.005
0.004 0.005
0.002 -0.01 0
0 -0.015 -0.005
0 250 0 250 0 250
Time (ms) Time (ms) Time (ms)
Figure 3.7: A 250 ms sEMG waveform is sown in a) and the corresponding signal transformations
as it goes through nodes B-F in the system model Fig. 3.6 (b-f respectively).
a) 110 b) 110 c) 110

100 100 100
90 90
90
iMDF (Hz)
80 80
80
70 70
70
60 60
50 60 50
40 50 40
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (Msamples) Time (Msamples) Time (Msamples)

d) 110
e) 104 f ) 103
100 102 102
90 100
150.5 101
80 98
iMDF (Hz)
150
100
70 96
149.5
99
60 94
149
98
50 92
148.5
40 90 97
148
30 88 96
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2147.5 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (Msamples) Time (Msamples)
147 Time (Msamples)
0 500 1000 1500 2000 2500 3000
Figure 3.8: IMDF tracking during fatiguing contraction from muscles a) LES T12, b) RES T12, c)
LES L4, d) RES L4, e) L RA and f) R RA. Blue line is iMDF using system described. Red line serves
as a guide.
gm1
D Q
gm2 Vout
gm3 C QN
Biasing
Figure 3.9: Circuit schematic of the gm-C VCO.
The frequency of oscillation fo is given by:
1 √
fo = gm1 gm3 . (3.12)
2πC
By choosing all the transconductances to be the same, then the oscillation frequency simplies to:
gm
fo = . (3.13)
2πC
IB
Furthermore, by biasing the OTAs to operate in the weak inversion region, gm = 2VT . Thus, the
relationship between fo and biasing current IB becomes:
IB
fo = . (3.14)
4πCVT
There are several advantages for implementing the described topology. This architecture
oers a highly linear, externally tunable oscillator. The relationship between the bias current and
the oscillation frequency is shown in Fig. 3.10. Moreover, the topology oers a wide dynamic range.
Finally, the component count is kept to a minimum. The biasing current IB ranges from 140 to 180
nA. The output signal is buered and then the frequency of oscillation is divided by two. Thus,
components size are kept to a minimum. Buering is important such as the capacitive load on the
oscillator does not aect one of the conditions for oscillation, where C1 = C2 = 10 pF.
Multiplier
The multiplier used as a squarer is shown in Fig. 3.11. The circuit was adapted from [6]. It is a
current mode, four quadrant multiplier. The multiplier uses MOS translinear loops (M1 − M12 )
to achieve multiplication. The biasing current (Ibias) is 250 nA. Simple OTAs with a linear input
2600
2400
VCO output frequency (Hz)

2200
2000
1800
1600
1400
1200
1000
100 120 140 160 180 200 220 240 260 280 300
Biasing current (nA)
Figure 3.10: Frequency responce of the gm-C VCO with a bias current in the range of 100-300 nA.
Iout+ Iout-
Ibias Iy+ Ix+ Ix- Ibias Iy- Ix+ Ix-
M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 M12
3 1 1
10 6 6 M1-12 60
0.8
M13 M14 M15
Figure 3.11: Current mode, four quadrant, translinear loop multiplier.
range of 200 mVp−p and Gm = 250 nS are used at the input nodes. The dierential output current is
1
converted into a voltage using an OTA congured as a single-ended voltage resistor where R= Gm .
Subtracting Amplier and Gain Adjustment
A Harrison amplier architecture was used [7] (Fig. 3.12) for subtracting the two power paths
C1
and adjusting their gain. The gain is set by the capacitor ratio
C2 thus allowing precise gain
adjustments without the need for resistive passive components, thus keeping power consumption of
the amplier to a minimum. The amplier employed was a folded cascode operational amplier with
an error amplier for the Common Mode Feedback (CMFB). The error amplier consists of two
dierential pairs, each one connected to the output of the dierential amplier. The cancellation of
the dierential signals is carried out at the outputs of these two dierential pairs and fed back to
the current source of the input pair. This CMFB pair requires a Gain Bandwidth Product (GBW)
twice the GBW of the input pair [8]. A disadvantage of this CMFB control is that the output swing
Reset
C
Vi2+ Vf-
C C
Vi2- Vf+ Vf-
2C C
Vi2+ Vf-
Vi2+ + -
2C
Vi2- Vf+ A Vout
4C
Vi2+ Vf- Vi2- - +
4C C
Vi2- Vf+
Vf+ C
8C
Vi2+ Vf-
8C
Vi2- Vf+
Reset
Figure 3.12: Harrison-amplier architecture [7] for a variable gain, subtracting amplier.
is limited by common mode input range of the CMFB amplier (2.8 × (Vgs − V t)). The amplier
had an open loop gain of 73 dB, a -3 dB bandwidth at 28 kHz (for load capacitance of 1 pF) and a
power consumption of 7 µA.
Integrator
The integrator employed is a dierential in, dierential out switched-capacitor (SC) architecture,
with oset compensation, adapted from [9, 10]. The iMDF tracker is designed to operate continu-
ously in time. Thus, the inherit dierential pair oset of the opamp would result to the saturation
of the lossless integrator. Hence, is it imperative that the integrator is oset compensated. The
1
circuit is shown in Fig. 3.13. The dierential gain is set to
50 to expand the integrators dynamic
range. A dierential topology was chosen to minimize the eects of parasitics, common mode noise
and DC osets. A non-overlapping clock generator is used to generate the two clock phases. The
clock frequency is 80 kHz and C1,2 = 100 fF. In φ 1 , C1 is charged with Vin , while C2 is sampling
the operational amplier's (opamp) oset. At the same time C3 along with the opamp are hold-
ing the last integral value. During φ2 , through charge sharing between C1 and C3 a new integral
value is computed and passed to the output, while the oset Vos stored in C2 is subtracted from
the computed value. Hence, the opamp's oset and charge injection osets are cancelled out. A
sample-and-hold circuit at the output allows two phase continuous operation.

res
res res
Vin+ A ... D
φ1 φ2 φ1 φ2 φ1
C3, 50C1,2
C1 C2 Vout-
A B
C D
C1 C2 Vout+
φ1 φ2 φ1 φ2 φ1
C3, 50C1,2
Vin-
res res
Figure 3.13: Dierential in/out switched-capacitor integrator with oset compensation.
S S S
2 0.4 0.8 0.4
0.35 0.35 0.35 0.35
S S IN OUT
0.8 0.4 0.8 0.4
0.35 0.35 0.35 0.35
S S S
Figure 3.14: Schematic of the trasmition gate used with MOSFET capacitors for charge injection
absorbtion.
A transition gate was used as a switch (Fig. 3.14). In order to cancel out charge injection, dual
MOSFET caps were used on each side of the MOSFET switches. The introduction of the capacitors
help to minimise charge infection and suppress clock feed-through by choosing Wswitch = 2 × Wcap .
Finally, common mode charge injection is fully cancelled out though dierential operation.
Band-Pass Elliptic Filter
In order to remove any residual harmonics from the mixers used and also suppress the odd harmonics
of the square wave VCO (shaping into a sine wave), a bandpass lter is used. Using an elliptic lter,
a very high rollo can be achieved with an attenuation at the stop-band of -33 dB.
The roll-of of an elliptic lter depends on its order and its cut-o frequency. For a high-pass
elliptic lter as the cut-o moves to higher frequencies, in order to keep the same roll-o, the number
of orders must be increased. MATLAB

® simulations showed that the steepest rollo for a 6th order
(high-pass) elliptic lter occurs for cuto frequencies less than 2 kHz. With a cut-o frequency in
that region the lter's stop-band frequency, fs where attenuation is -40 dB, would be less than 250
-F3
-F2
Vin
K0 K1 TBP1 K2 TBP2 K3 TBP3
Adder 1 Biquad 1 Biquad 2 Biquad 3
B0 B1 B2 B3
Adder 2
Vout
Figure 3.15: FLF eliptic lter architecture.
Table 3.1: 6th order FLF bandpass elliptic lter parameters.

K0 0.6045 B0 0 F2 0.6577
K1 2.3573 B1 0.1697 F3 0.2275
K2 2.1860 B2 -0.2166 Q 3.5
K3 1.9963 B3 1.1300 fo 1750 Hz
Hz bellow the cut-o.
A 6th order band-pass elliptic lter with a bandwidth (BW) of 500 Hz (EMG signal BW)
was designed using the Follow the Leader Feedback (FLF) architecture depicted Fig. 3.15 (design
adapted from [11]). The FLF architecture was chosen for the design because of its good trade-o
between sensitivity/robustness and tuning ability/simplicity, compared to both simple cascading
and more complex leap-frog (ladder) architectures [12]. The FLF architecture oers a simpler
implementation and lower design complexity, as identical biquad bandpass lters are used as primary
resonators for the lter. Feedback coecients F2 and F3 are implemented as gains. Feed-forward
coecients B0 − B3 are used to realise nite zeros required for an elliptic lter implementation.
Constants K0 − K3 are the gains of the lter and each biquad. In order to nd the correct values for
each coecient a 3rd order low-pass lter is used to perform a low-pass to band-pass transformation.
A MATLAB
® script adapted from [13] was used to estimate the coecient values. Coecient values
and lter parameters are listed in Table 3.1.
The circuit diagram of the biquads used is shown in Fig. 3.16. In order to reduce the
transconductance and increase the linearity of the OTAs used (see Fig. 3.17) source degeneration
is used by adding triode PMOS transistors B1 and B2 . The adders used are shown in Fig. 3.18.
Biasing
CMF CMF
Vin+
gm0 gm3 gm1 gm2 Vout+
C2 C1 Vout-
Vin-
Adjustable Fc
Biasing
Figure 3.16: Biquad architecture circuit diagram.
2 2
20 20
2 2
20 20
2 2 2
20 20 20
Vout- 5
10
Vout+
Ibias B1
B2
V- M1 M2 V+
40 40
10 10
1 1 1 1 1 1
10 10 10 10 10 10
Figure 3.17: OTA used in biquad lter with source degeneration.
Addition is achieved in current mode and converted to voltage using an OTA congured as a single-
ended voltage resistor. Feedback gains are realised as the transconductance ratio of the dierential
pairs M1 − M6 and Gms . The designed lter has a centre frequency fo = 1750 Hz with cutos at
fc(lo) = 1450 Hz and fc(hi) = 2000 Hz.The lter has a rollo of -33 dB/octave. The centre frequency,
quality factor and gain are given by Eq. 3.15. Based on these equations and the parameters shown
in Table 3.1, the biquad circuit parameters are calculated and shown in in Table 3.2. To achieve
the required rollo, two lters where cascaded.
r
gm1 gm2
ωo =
C1 C2
s
gm1 gm2 C2
Q= (3.15)
(gm3 )2 C1
gmo
K=
gm3
K0 B3 F2 B2 F3 B1
7.2 7.2 4 2.2 0.4 1.1
40 or 21 40 or 70 40 or 70
4
40
Ibias 5 5 5
10 10 10
Vp_1 Vn_1 Vp_3 Vn_3

40 40 40 40 40 40
10 10 10 10 10 10
Vp_2 Vn_2
gms
Vb
Vout- Vout+
Figure 3.18: Current mode adder circuit diagram.
Table 3.2: Biquad circuit parameters.

Biquad1 Biquad2 Biquad3
gm1,2 (nS) 313 313 313
gm3 (nS) 178.9 193 210.8
gm0 (nS) 421.7 421.7 421.7
C1 (pF) 7.115 7.115 7.115
C2 (pF) 30 30 30
Simulation Results
The developed iMDF tracking system was simulated using Cadence Spectre (6.1.5isr10) simulator,
with foundry supplied PSP models. The system is validated using retrospective data from back
muscles used in a previous clinical study. Data from muscles LES T12, RES T12 and LES L4 were
collected while the muscles were fatiguing during a static contraction, starting from a non-fatigued
state up until exhaustion.
In order to aid simulation eciency, the VCO and gain stages were replaced by an ideal
components. The algorithm described in Fig. 3.6 assumes an ideal high pass lter to ensure optimum
performance. Simulation of the bandpass elliptic lter gain response, as well as the Monte-Carlo
extremities of the cascaded lter can be seen in Fig. 3.19a-c. Moreover, the eect of Q and centre
frequency variation on the FLF structure frequency response is shown in Fig. 3.19d. It is observed
that process and mismatch variation has a signicant eect on the lter's centre frequency. For
this reason, a programmable current was added, controlling the centre frequency and allowing for
post-production calibration of the lter. Furthermore, by centering the lter (and adjusting the
VCO) to lower frequencies using the programmable current, the lter response overshoot can be
minimised. Thus, it was decided to be adjusted during testing.

Table 3.3: Comparison between circuit simulations and MATLAB®

Muscle MARD R − value Figure
LEST12 3.8 (%) 0.96 3.20 (a)
LESL4 4.2 (%) 0.95 3.20 (b)
REST12 9.6 (%) 0.92 3.20 (c)
Since the lter used is of 6th order elliptic, the dierence between the cuto frequency and
the stop-band frequency aect the attenuation of the signal. To counterbalance this eect, a linear
gain term G is introduced (see Fig. 3.5). Although there is some signicant overshoot present in
the simulation, the gain term G can be adjusted during testing to accommodate for the change in
the passband gain of the lter. Furthermore, the centre frequency of the lter structure can ba
adjusted externally during testing by varying a selected bias current in the biquads. The eect
on the estimated iMDF is a constant residual oset which can be easily removed either on chip or
during post-processing of the results. The designed lter has a centre frequency fo = 1750 Hz with
cutos at fc(lo) = 1450 Hz and fc(hi) = 2000 Hz. The current of the gm cell controlling the VCO
frequency is used as the system output and its relationship with iMDF is given by Eq. 3.16. In
Eq. 3.16, 150 is the starting frequency dierence between the VCO and the lter's lower cut-o,
V COsen = 5.5 Hz/nA and IDC = 145 nA is the nominal current for the starting VCO frequency.
iM DF = 150 − V COsen (Iout − IDC ) (3.16)
Fig. 3.20 shows the estimated iMDF under fatigue of muscles LES T12 (a), LES L4 (b) and
RES T12 (c). The simulation time is 145 s. The blue line is the estimated iMDF by this system.
The red line is an estimation of iMDF in MATLAB

® after sub-hertz ltering. The Mean Absolute
Relative Dierence (MARD) is used to asses the relative error between the two calculated iMDF
signals. However, in order to measure the relative error, a constant oset of 55 Hz was subtracted
from the system output. The oset was Table 3.3 shows MARD and correlation comparison results.
Assuming a maximum iMDF that the system can estimate of 150 Hz, then the average error would
be less than 5 Hz . Due to the noisy nature of the EMG signal and the inheritable noisy, continuously
varying iMDF, this error is regarded as acceptable for the intended application. The system has
been designed for low-power consumption achieving 190 µW from a 3.3 V supply during continuous
operation. Power consumption details are summarised in Table 3.4.

a) b)
20
4
0 Ideal filter Ideal filter
Circuit filter 3 Circuit filter
-20 Cascade Cascade
2
-40
1
Gain (dB)
-60 0
-80 -1
-2
-100
-3
-120
-4
-140
-5
-160
100 10 2
10 4
10 6
10 8 1400 1500 1600 1700 1800 1900 2000 2100 2200
Frequency (Hz) Frequency (Hz)
c) d)
20
0 Monte-Carlo 0
Extreemes
-20 Nominal
-40 -2
-60
Gain (dB)
-80 -4
-100
15% Q variation
-120 -6 5% ωo variation
-140 15% Q + 5% ωo variation

no variation
-160 -8
-180
100 101 102 103 104 105 106 251 398 631
Figure 3.19: Simulation of the bandpass elliptic lter gain response: a) ideal lter, circuit implemen-
tation and cascaded lter response, b) closeup of a) around the 3dB point, c) Monte-Carlo extremities
of the cascaded lter and d) eect of Q and centre frequency variation on the FLF structure frequency
response. Figure from [11].
Table 3.4: System-Level power breakdown (C35, 3.3V).

Block Power
Mixer 27.7 µW ×2
VCO µW
26.4
Filter 21.1 µW ×2
Integrator 17.5 µW
Scaling Amplifier 23.1 µW
Subtracting Amplifier 23.1 µW
Total power 187.7 µW
a) 150
Cadence
100 Matlab
iMDF (Hz)
50
0
0 30 60 90 120 150
Time (sec)
b)
100
80
iMDF (Hz)
60
40
0 30 60 90 120 150
Time (sec)
c)
100
80
iMDF (Hz)
60
40
20
0 30 60 90 120 150
Time (sec)
Figure 3.20: Blue line: cadence simulation iMDF tracking during fatiguing contraction
® from muscles
a) LES T12, b) RES T12 and c) LES L4. Red line: iMDF tracking using MATLAB simulation.
3.5 Measurement Results and Reections on iMDF Tracker 65
Figure 3.21: Die micro-photograph of the iMDF tracking ASIC.
3.5 Measurement Results and Reections on iMDF Tracker
The fabricated ASIC (codename Altair) is shown in Fig. 3.21. Initial testing showed that the VCO
was functioning properly. By grounding the inputs to the system, the fundamental frequency was
measured to be 1310 Hz, which is very close to the desired one of 1300 Hz. A signal resembling
band limited white noise was constructed in MATLAB

® and applied to the inputs of the system.
The signal had a bandwidth of 250 Hz. Under correct operation, the output of the system would
settle to an iMDF value close to 125 Hz. The iMDF tracker was tested in closed loop operation
since it was implemented without the option to break the loop. However, the output of the iMDF
tracker, the integrator output, kept moving towards the power supply rails, indicating a saturated
integrator. Thus, no further measurements were conducted.
3.5.1 Design Considerations for Future Improvements
One of the most important elements of this system is the SC integrator. It tunes the VCO to
the desired frequency and its output serves as a multiple of the estimated iMDF. The integrators
feeds to an externally adjustable gm cell which changes the bias current of the VCO and hence its
frequency. A Programmable Gain Amplier (PGA) precedes the integrator. Its use is to adjust
1.7
1.7 3
1.68 3
1.68 2.5
RESET
RESET Period 2.5
1.66 Period
RESET
RESET Period
1.66 2Period
Amplitude (V)
2
Amplitude (V)
1.64
1.64 1.5
1.5
1.62
1.62 1
1
1.6 0.5
1.6 0.5
1.581.58
0 0 1000
10002000
2000 3000
3000 4000
4000 5000
5000 6000 7000 8000
6000 7000 8000 9000
900010000
10000 0 0 1000
10002000
2000
30003000
40004000
50005000
6000 6000
7000 7000
8000 8000 9000 10000
9000 10000
Time
Time(samples)
(samples) TimeTime (samples)
(samples)
(a) (b)
Figure 3.22: (a) The dierential output of the PGA durring and after a RESET period, (b) The
dierential output of the SC integrator following the PGA durring the same period.
the dierence from the two feedback paths to a value that does not saturate the integrator. It was
found that the integrator would saturate immediately after power-up and return to that state after
reset. It is believed that a DC oset is causing the integrator to saturate so fast. Consider Fig. 3.22
for the output of the subtracting PGA and the integrator. It is clear that after the reset period,
the PGA outputs start to drift away and the integrator saturates before a valid measurement is
obtained. Careful examination of the designed indicated that the possible reason for the DC oset
would be the transition gate switch used in the Harrison topology amplier. The switch had an
NMOS transistor with its bulk connected to ground. The p-n junction is reverse biased but exhibits
leakage, thus a small current ow causes a drift at the input nodes of the amplier and changes the
DC operating point. A possible solution would be to use only a PMOS MOSFET switch with its
bulk to common mode voltage.
Furthermore, a better implementation would be to remove the Harrison scaling and subtract-
ing ampliers completely. Since the output of the squarer is in current mode, then current scaling
and subtraction would be a much simpler, robust and intuitive solution. Replacing the translinear
current squarer with a voltage squarer would have a direct impact on signal distortion, since less
transconductance cells would be introduced in the feedback path. Furthermore, designing a lower
order, Butterworth architecture for the main lter would result to a simpler, more robust system.
Monte Carlo simulations indicated that the lter coecients are prone to mismatch. Since the cho-
sen lter architecture requires very precise coecients, it would be better to replace the elliptic lter
Voltage Current Spike Digital
2
X
- Current Difference
Amplified Neuron
sEMG signal Variable HPF Ingegrator & Output
+ Tap Controller
2 x2
X
Figure 3.23: Block diagram representation of the second version of the iMDF Tracking System.
structure with a more robust architecture. Although adjusting the centre frequency was possible,
the coecient values were xed. Finally, a switched capacitor CMFB architecture would be more
suitable for the SC integrator. Removing the analogue integrator and transferring the loop control
circuit to the digital domain would make a more robust and easy to adjust architecture.
Despite it's shortcomings, this implementation of a median frequency tracker showed that it
is possible to create a low power, real-time system that is useful in monitoring and analysing the
sEMG signal. Such a system is suitable for integration with other sEMG analysis tools dedicated
for muscle fatigue monitoring. Hence, contributing to the creation of a wearable, standalone tool
dedicated to sEMG and muscle fatigue analysis.
3.5.2 A novel Neuromorphic based iMDF Tracking System
Examination of the experimental and simulation results showed that minor circuit and design
changes could potentially improve the iMDF tracker and result in a functional fabricated ASIC.
Altering the PGA and subtracting amplier to operate in current mode would be sucient to im-
prove the design. However, instead of revising the iMDF tracker on a circuit level, a system level
improvement and redesign based on the design considerations and experience of the current system
was more promising. Careful analysis of the chip performance and design choices yielded several
areas that could be revised and implemented dierently

2. The improvements on the second version
reected upon the shortcomings of the rst system. The improvement eorts were concentrated in
reducing the complexity of the system while increasing its robustness to process variations.
The system was adapted to operate in baseband. The new system is presented in Fig. 3.23.
The VCO and high order elliptic lter were replaced by a tunable, second order high pass lter.
2 The circuit implementation of the second, improved version of the iMDF tracker was performed by Daiwen Sun
under the direct supervision of Ermis Koutsos, towards the end of this thesis.
M5 M6 M11 M12
10/1 10/1 10/1 10/1
Vout
M4
64/0.6
C1 M10
64/0.6
M7 M9 M8
M1
16/1
M3
64/0.6
M2
16/1
16/1 64/0.6
16/1
M14 C2
M13 VB 22/5
VB
22/5
Vin
Figure 3.24: Second order Gm -C High Pass Filter Schematic. By operating current source transistors
M13 and M14 in the weak inversion region, the cuto frequency depends on the biasing current. An
array of binary scaled current mirrors is used to control the bias current and adjust the cuto frequency
of the lter.
Thus, there was no need for a bandpass lter to lter out the odd harmonics and contribute to noise
and signal distortion. The voltage squarer output, a current signal, is fed into an integrate-and-re
neuromorphic circuit where it is converted into a stream of discrete time events [14]. The use of
silicon neurons, hybrid analogue and digital circuits, can provide a qualitative approximation rather
than an exact value with an optimised power consumption, which is essential for integrated systems
and wearable devices. Finally, the dierence and integration operation were implemented using
digital circuits in order to eliminate performance dependants on the SC integrator oset, drift and
noise. A digital controller was responsible for adjusting the bias current of the lter and alter the
centre frequency. The input signal to the system is a pre-recorded sEMG signal bandpass ltered
from 10-500 Hz to cover the full spectrum of the sEMG signal and amplied to an amplitude of 200
mVp−p .
The tunable lter is a current controlled high-pass lter where its cut-o frequency can be
adjusted in the range of 20-450 Hz to cover 90% of the sEMG spectrum where the median frequency
usually occurs. The lter is based on a second order Gm −C architecture where the cut-o frequency
is controlled by its biasing current(design adapted from [15]) and is shown in Fig. 6.2b. A scaled
current mirror is used to select the bias current for the control of the centre frequency. A current
mode circuit (Fig. 6.2a) was used to square the signal and approximate its instantaneous power.
In addition, the circuit also acts as a transconductor which converts the voltage mode signal into
current mode signal which can be used to drive the integrate-and-re neuron circuit [16].
An Axon-Hillock circuit (Fig. 6.2c) was used to convert the current mode signal into the
spike domain because of its simplicity and tunability [17]. The circuit integrates the signal power
M9
20/2
M10
20/2
VG+Vi VG-Vi VG VG
Ibias
M1 M2 M3 M4
2.8/1.2 2.8/1.2 2.8/1.2 2.8/1.2
I1 I2
Io
M5 M7
20/1.2 20/1.2
M6
20/1.2
M8
20/1.2
Figure 3.25: Analogue Current Mode Squarer Circuit. The circuit also acts as a transconductor
which converts the voltage mode signal into current mode signal which can be used to drive the
integrate-and-re neuron circuit.
over an adjustable time window which gives the average signal power. The integration window is
controlled by the digital controller, which is operating at 5 kHz. The integrated signal power is
represented as the amount of spikes red from the neuron circuit. An asynchronous counter is used
to count the spikes red during the given period (0.2 ms) and send them into the digital feedback
controller. The neuron is tuned to be linear and exhibit large sensitivity to the current change from
the inputs. This means that small changes in the input current will result in signicant dierence
in spikes red. This allows for a lower order lter to be used in the system because the neuron
can express a dierence even if the lter has less selectivity. As a result, an ADC is replaced by a
neuron circuit which acts as an interface between the analogue and digital domains. In addition, the
neuron circuit consumes much less power than a traditional ADC which reduces the overall power
consumption signicantly.
In order to improve the robustness of the system and to further simplify the design, a digital
adder is used instead of an analogue integrator. The integrated power dierence is used to control
the cut-o frequency of the high-pass lter, thus allowing the lter to select dierent portion of
the spectrum and estimate the median frequency of the input signal. The digital controller reads
the two counter values at the frequency of 5 kHz and calculate the dierence between them. the
number of pulses red is directly proportional to the signal power through both signal power paths
(ltered, un-ltered). This value will indicate how close the current cut-o frequency is, compared
to the median frequency of the input sEMG signal. There are a total of 127 possible biasing currents
available for the lter which result to 127 unique cut-o frequencies linearly spaced between 20 Hz
Iin M1 M5 M15
4/1
4/1 4/1
Vspk
M4 M10 M14 M16

4/1 4/1 4/1 4/1
M13
4/1
M6 M9
Vmem 4/1 4/1 M12
4/1
C1 vrfr M11
4/1
vsf M7 M8
4/1 4/1
M2 M3
4/1 8/1
iLeak
Figure 3.26: Integrate-and-Fire Neuron Circuit Implementation. The circuit integrates the signal
power over an adjustable time window which gives the average signal power. The integrated signal
power is represented as the amount of spikes red from the neuron circuit. The neuron is tuned to
be linear.
and 450 Hz.
Fig. 3.27 shows the estimated median frequency during an isometric fatiguing contraction for
subject 1 (a), subject 2 (b) and subject 3 (c). Although both median frequency tracking systems
were developed for the same purpose, dierent datasets were used to characterise them as the
neuromorphic design was developed nearly two years after the rst system. At that time, dierent
and more accurate datasets had been collected. The blue trace is estimated median frequency
by the system proposed in this paper and the red trace is the median frequency calculated by
MATLAB
® under 0.1 second segments after sub-hertz ltering. The MARD is used to asses the
relative error between the two calculated iMDF signals. Table 3.5 shows MARD and correlation
comparison results. The system can accurately track iMDF of an sEMG signal with an average error
of only 0.64%. Moreover, it has been optimised for low-power consumption achieving 138 µW from
a 3.3 V supply during continuous operation. Power consumption along with circuit performance
characteristics are summarised in Table 3.6. Compared with the previous implementation of the
iMDF tracker, in the proposed design the relative error has reduced by 80% while increasing process
variation robustness and design simplicity. A more detailed view on the schematics and VHDL code
used can be found in Appendix A.

a) 140
120
iMDF (Hz)
100
80
60
40
0 10 20 30 40 50 60
Time (sec)
b) 140
120
iMDF (Hz)
100
80
60
40
0 10 20 30 40 50 60 70 80 90
Time (sec)
c) 140
120 Cadence
Matlab
iMDF (Hz)
100
80
60
40
0 5 10 15 20 25 30 35 40 45 50
Time (sec)
Figure 3.27: Blue Trace: Median Frequency during fatigue using proposed system
® from (a) Subject
1, (b) Subject 2, (c) Subject 3, Red Trace: Median Frequency using MATLAB medf req() function.
Table 3.5: Results comparison between proposed system and MATLAB®

Subject MARD % P-Value R-Value Figure
1 0.76 < 0.01 0.9491 3.27 (a)
2 0.15 < 0.01 0.9521 3.27 (b)
3 1 < 0.01 0.9384 3.27 (c)
3.6 Summary 72
Table 3.6: Power Consumption Breakup (C35, 3.3V)

System Module Power
Gm − C Filter µW
1
Voltage Squarer 3 µW ×2
Neuron 20 µW ×2
Biasing and Buering 80 µW
Digital Controller 15 µW
Total 138 µW
3.6 Summary
In this chapter, the properties of the sEMG power spectrum during fatigue, the benets of iMDF
tracking and possible circuit implementations were thoroughly investigated. Moreover, two iMDF
tracking systems were presented along with their strengths and weaknesses, as one was a completely
analogue implementation and the other was based on a mixed-signal approach.
The spectral compression of the sEMG signal during fatigue is best represented by the median
frequency of the power spectrum. Several attempts were made in the past in an eort to produce
a portable muscle fatigue monitoring system based on the median estimator. However, the designs
presented in this chapter were the rst attempt to an IC based system. After establishing the
potential benets and challenges of a closed loop approach, the rst iMDF tracking system used a
xed high order lter and a VCO to approximate the power of the spectrum above and below the
median frequency. The subtraction and integration of the two signal paths were two very important
blocks in the feedback loop. The chosen design architecture for these system components proved to
be troublesome. Thus, measurements of the fabricated IC were incomplete as one of the feedback
paths was saturating the integrator.
After careful examination of the rst design choices and challenges, the system architecture
was revised and redesigned from the ground up, exploiting the full potential of the iMDF tracker.
The resulting system employed a baseband closed loop approach. A variable lter was used to
integrate the sEMG spectrum. Moreover, the lter order and complexity was signicantly reduced.
Hence, the lter response would be more robust to process variations. However, by reducing the
order of the lter, the tracking accuracy could be reduced. To combat the lter limitations, the
SC integrator was replaced by a neuromorphic circuit, integrating and converting the sEMG power
to a bit-stream. The negative feedback was implemented in the digital domain, thus increasing
the overall accuracy and robustness of the feedback loop to osets, noise and process variation.
3.6 Summary 73
The lter cut-o was adjusted using a digital controller, driven by the adders. Lastly, in order to
decrease the number of design components in each feedback path, an analogue voltage squarer was
used where its current mode output was scaled and fed to the neuromorphic integrator.
Both trackers performed well in simulations (< 5% average error) and their low power (<
0.2 mW) design makes them viable systems for wearable muscle fatigue monitoring devices. The
limitations of the rst analogue iMDF tracker led to the creation of a novel neuromorphic based,
mixed signal iMDF tracker with increased accuracy and better performance. However, since the
neuromorphic design lacks the large cascaded elliptic lter, its area is signicantly smaller. Fur-
thermore, it consumes 27% less power while the tracking accuracy in increased, with an average
tracking error of less than 1%. Finally, its increased robustness to variations due to process and
mismatch factors makes the neuromorphic tracker the better design choice of the two systems. The
revised system was not fabricated due to time constraints.

REFERENCES 74
References
[1] S. Karlsson, J. Yu, and M. Akay, Time-frequency analysis of myoelectric signals during dy-
namic contractions: a comparative study, Biomedical Engineering, IEEE Transactions on,

vol. 47, no. 2, pp. 228238, 2000.
[2] D. L. Gilmore and C. J. De Luca, Muscle fatigue monitor (mfm): Second generation, IEEE
Transactions on Biomedical Engineering, no. 1, pp. 7578, 1985.
768, 1982.
on, no. 1, pp. 17, 1985.
[6] M. Gravati, M. Valle, G. Ferri, N. Guerrini, and N. Reyes, A novel current-mode very low power
analog cmos four quadrant multiplier, in Solid-State Circuits Conference, 2005. ESSCIRC
2005. Proceedings of the 31st European, pp. 495498, IEEE, 2005.
[7] R. R. Harrison and C. Charles, A low-power low-noise cmos amplier for neural recording
applications, IEEE Journal of solid-state circuits, vol. 38, no. 6, pp. 958965, 2003.
[8] W. M. Sansen, Analog design essentials, vol. 859. Springer Science & Business Media, 2007.
[9] K. Nagaraj, J. Vlach, T. Viswanathan, and K. Singhal, Switched-capacitor integrator with
reduced sensitivity to amplier gain, Electronics Letters, vol. 21, no. 22, pp. 11031105, 1986.
[10] C. C. Enz and G. C. Temes, Circuit techniques for reducing the eects of op-amp imperfections:
autozeroing, correlated double sampling, and chopper stabilization, Proceedings of the IEEE,
vol. 84, no. 11, pp. 15841614, 1996.
[11] C. Qian, J. Parramon, and E. Sanchez-Sinencio, A micropower low-noise neural recording
front-end circuit for epileptic seizure detection, Solid-State Circuits, IEEE Journal of, vol. 46,
no. 6, pp. 13921405, 2011.
REFERENCES 75
[12] A. J. Casson and E. Rodriguez-Villegas, A review and modern approach to lc ladder synthesis,
Journal of Low Power Electronics and Applications, vol. 1, no. 1, pp. 2044, 2011.
[13] D. Hernandez-Garduno, Follow-the-leader architecture 6th order band pass elliptic lter,
December 2002.
[14] G. Indiveri, B. Linares-Barranco, T. J. Hamilton, A. Van Schaik, R. Etienne-Cummings, T. Del-
bruck, S.-C. Liu, P. Dudek, P. Häiger, S. Renaud, et al., Neuromorphic silicon neuron cir-
cuits, Frontiers in neuroscience, vol. 5, p. 73, 2011.
[15] R. L. Geiger and E. Sanchez-Sinencio, Active lter design using operational transconductance
ampliers: a tutorial, Circuits and Devices Magazine, IEEE, vol. 1, no. 2, pp. 2032, 1985.
[16] G. Giustolisti, G. Palmisano, and G. Palumbo, A novel cmos voltage squarer, in Circuits
and Systems, 1997. ISCAS'97., Proceedings of 1997 IEEE International Symposium on, vol. 1,
pp. 253256, IEEE, 1997.
[17] G. Indiveri, A low-power adaptive integrate-and-re neuron circuit, in ISCAS (4), pp. 820
823, Citeseer, 2003.

76
Chapter 4
A wearable MFM System based on the

Muscle Fibre Conduction Velocity
4.1 Introduction
S
O far, instantaneous Median Frequency (iMDF) is recognised by physiologists and researchers
as the golden standard" for muscle fatigue analysis. Nevertheless, as its name indicates, this
metric concentrates on the frequency content of the surface EMG (sEMG) signal. Although this is
helpful in analysing sEMG signals, it does not provide any information about the physical changes in
the muscle that caused that shift in frequency. As the muscle fatigues lactic acid and Potassium ions
(K
+ ) accumulates in the extracellular muscle space, impairing conduction of the action potentials
across the muscle membrane, thus slowing down the Motor Unit Action Potentials (MUAPs) [1].
Furthermore, fast twitch muscle bres dominate high force production but get easily fatigued and
drop-out. Following that, fatigue resistant muscle bres take over.
Muscle Fibre Conduction Velocity (MFCV) is a measure of the travelling speed of MUAPs
in muscle tissue and is one of the most important items which reects muscular activity [2]. As
the muscle fatigues, MFCV decreases [3]. One of the advantages of MFCV is that it is reliable
under static and dynamic contractions [4]. Furthermore, MFCV is partially related to the observed
spectral compression of the sEMG signal [5]. However, changes in the spectral content of the sEMG
signal are disproportionately larger than decreases in MFCV. Moreover, recoveries in frequency are
more rapid than lactate removal in the muscle, MFCV can provide a more detailed insight on muscle
fatigue and muscle recovery than Power Spectral Density (PSD) monitoring alone [4, 6].
This chapter will present a novel CMOS system for the measurement of MFCV. This tool
will enable users/researchers to further understand the mechanisms behind muscle fatigue. Firstly,
4.2 Methods for Conduction Velocity Estimation 77
methods for conduction velocity estimation are presented. Following that, a discussion of feasibility
provides the most suitable algorithm for MFCV estimation. Finally, CMOS implementation and
measured results are presented.
4.2 Methods for Conduction Velocity Estimation
MFCV correlates with the EMG power spectrum and represents the velocity of MUAPs in the muscle
[5]. Thus, MFCV can be calculated using one of two dierent approaches; processing one detected
signal or processing two or more detected signals along the muscle bre direction. Inspiration could
also be obtained from Acoustic Source Separation and Bearing Source Separation analysis as in
[79].
Processing One Detected Signal
Estimating MFCV by processing one detected signal works by extracting spectral information from
the signal. Algorithms using this approach include the following:
Characteristic Spectral Frequencies
Lindstrom et al provided a mathematical model linking MFCV to the sEMG power spectrum (4.1).

f
P (f ) is the EMG power spectrum, v is conduction velocity and G v is the the shape of the
spectrum of the surface detected action potential. A shift in the power spectrum can be accounted
for by a change in MFCV. Thus, relative changes in MFCV result to relative changes in sEMG
characteristic spectral frequencies. Both iMDF and instantaneous Mean Frequency (iMNF) can be
used to track these changes/frequencies. However, this method is valid under one assumption. The
number of Motor Units (MUs) does not change during the contraction and estimation of MFCV.
Furthermore, this algorithm can only provide information about relative changes in MFCV and not
absolute values [1017].

1 f
P (f ) = 2 G (4.1)
v v
Detection of Spectral Dips
By applying dierent electrode congurations we can realise spatial lter. The spatial and temporal
domains are related by the velocity of propagation. Thus, if a signal is travelling with constant speed,
the spatial lter transfer function is equivalent to a temporal lter transfer function. Assuming a
constant speed, the temporal (ft ) and spatial (fz ) transfer function frequencies depend on MFCV
according to: ft = v · fz . Any zeros introduced in the spatial lter transfer function will appear
in the equivalent (scaled) temporal transfer function. The position of these zeros, or spectral dips,
depends on MFCV. However, estimation of MFCV using this methods will introduce large variance
and spectral dips are suciently dicult to detect accurately [5, 18].
Autocorrelation
As described in the section above, sEMG spectral dips can be seen in the temporal domain. Hence,
they appear in the autocorrelation of the signal. Spinelli et al. showed that the rst dip in the
power spectrum introduced by a bipolar spatial lter corresponds to the rst minimum of the
autocorrelation function of the sEMG signal [19].
The disadvantage of these methods is that they require powerful spectral analysis tools.
Furthermore, MFCV estimates will not be very accurate and will show big variance. Finally, these
tools can only provide relative values of MFCV,which limit the applications of an MFCV estimator.
Processing More Than One Detected Signals
Another approach for nding MFCV involves processing two or more detected signals. The elec-
trodes are placed perpendicular to the underlying muscle bres. Each signal is detected at a dierent
point of the MUAP propagation path. The underlying principle is the measurement of time delay
between two myoelectric signals. Assuming two detection points, A and B, the conduction velocity
d
of the propagating MUAP will be dened as CV = δt where d is the distance between the two
detection points. Consider Eq. 4.2 for a simple model of the two detected sEMG signals, where θ
is the delay between the two signals, K is the number of detected signals and wk (n) is white, zero
mean, additive Gaussian noises of equal variance.
x1 (n) = s(n) + w1 (n)
x2 (n) = s(n − θ) + w2 (n) (4.2)
xk (n) = s(n − (k − 1)θ) + wk (n)
Algorithms processing two or more detected signals include the following:

a)
b)
Figure 4.1: Distance between reference points: a) Seven sEMG signals detected along the muscle
bre, b) Alignment of EMG segments from reference points. Figure from [20].
Distance Between Reference Points
The model in Eq. 4.2 assumes that the two detected signals are identical with the addition of
noise. Thus, the two signals would have the same shape with the introduction of a delay. As a
result, any specic reference point such as a valley, a peak or a zero can be used to align the two
signals and estimate the delay between them. Fig. 4.1a shows seven sEMG signals detected along
the muscle bre, where the increasing delay between them is observed. In Fig. 4.1b, the sEMG
segments are aligned and the delay between them is calculated [20]. However, the assumption is
not valid in real life scenarios. This is because the detected delayed signals exhibit changes in shape
and characteristic reference points are hard to locate. Furthermore, this approach is limited by the
systems time resolution.
Detecting Phase Dierence
Processing signals in frequency domain eliminates the issue of time resolution. The frequency
domain model of Eq. 4.2, is shown with the absence of noise in Eq. 4.3:
X1 (a) = S(a)
(4.3)
X2 (a) = S(a)e−j2παθ/N
where a is the discrete frequency and N the number of samples. Thus, the phase dierence of the
Fourier Transform of the two signals is given by 2παθ/N . The phase dierence is a linear function
of frequency, with slope related to the delay to be estimated [5, 15, 21]. However, it has been shown
[22] that this method is very sensitive to additive noise.

Spectral, Temporal and Spatial Matching
The time lag between the two (or more) sEMG signals is the time lag that can minimise the mean
square error between the original signal (point A) and the delayed signal (point B). The mean
square error in the time domain is given by Eq. 4.4:
N
X
et (τ ) = [x2 (n + τ ) − x1 (n)]2 (4.4)
n=1
where x1 and x2 are the two signals with a delay between them, n is the sample number and τ is
the time lag.
Minimization of the mean square error can happen in the time domain as well as in the
frequency domain. Furthermore, matching between signals ltered in the temporal and in the
spatial domains can be used to minimise the mean square error [5, 23, 24]. The delay between two
time shifted signals can be presented as the assessment of a transfer function in the time domain
that compensates for a lter applied in the spatial domain. On the other hand, the eect of the
transfer functions applied in the spatial domain can be compensated for by two lters in the time
domain, which will depend on the unknown estimated delay [23].
Cross-Correlation
The time lag at which the cross-correlation function is maximum can be used as an estimator of
delay, as was early demonstrated for surface EMG signals by Parker et al. [25]. Eq. 4.5 represents
the normalised cross-correlation function, where X is the reference EMG signal, Y is the delayed
EMG signal and Rτ is the correlation coecient normalised from 1 to -1. The time delay between
the two signals occurs at the point where the correlation coecient is at a maximum. Fig. 4.2 shows
the result of the cross-correlation function of two EMG signals detected along the muscle bre path.
The cross-correlation result (Fig. 4.2b) is at a maximum at a point equal to the time delay between
the two sEMG signals (Fig. 4.2a). It has been shown [22] that the cross-correlation method performs
comparably to spectral matching based methods, while being more computationally ecient.
N,N
X+τ N
X
(Xi − X̄) · (Yj − Ȳ ) x2 (n + τ )x1 (n)
i=0,j=i+τ n=1
Rτ = v =v (4.5)
uN N +τ
uX X uN N
t (Xi − X̄)2 · 2
uX X
2
(Yi − Ȳ ) t x1 (n) · x22 (n)
i=0 j=i+τ n=1 n=1
4.3 MFCV Tracking Algorithm 81
Figure 4.2: Cross-correlation result of two EMG signals detected along the muscle bre path. (a)
Reference and Comparison (delayed) EMG signals, (b) Cross-correlation coetient. Figure from [26].
4.3 MFCV Tracking Algorithm
Literature has shown that the cross-correlation method is an ideal candidate for estimating MFCV.
Cross-correlation is a robust and accurate measurement of similarity between signals, with a high
degree of noise immunity [5]. Cross-correlation uses two signals and estimates the delay between
them. The sEMG detection points must lie in the muscle bre path, from the innervation zone to the
myotendonous part of the muscle. The innervation zone is a small region or band of muscle tissue
wherein MUAPs originate and then propagate bidirectionally toward each tendon" [27]. Fig. 4.3
shows the position of the innervation zone of the short head of the biceps brachii. It can be seen that
the zone is close to the belly of the muscle. However, this observation does not hold true for every
muscle in our body, hence the innervation point should be located for each muscle separately [28].
Fig. 4.3 further illustrates the propagation of the MUAPs along the muscle bre in two directions
(up and down) [28].
The detection points A,B for the short head of the biceps brachii are shown in Fig. 4.4(a).
In this example the detection points are located above the innervation zone. Placing the detection
points bellow the innervation zone would result in the same observations. As explained in Chapter 2,
single dierential (bipolar) detection was used. The resulting SD detection points are shown in
Fig. 4.4(b). It was observed experimentally that bringing the two electrode sets closer and sharing
Arrows Indicate Innervation Zone

sEMG recordings, 5 ms/division
61%
68%
70%
74%
79%
Figure 4.3: The propagation of MUAPs along the muscle bre in two directions, originating from
the innervation zone. Figure from [28].
a) b) c)
A A A
B B B
Figure 4.4: Cross correlation detection points on bicepts brachii muscle. (a) single electrode,
monopolar detection, (b-c) double electrode, bipolar detection. Figure from [32].
the middle electrode, as in Fig. 4.4(c), would reduce the DC oset between the two recordings and at
the same time improve the detection of the delay between the two points. A better choice may have
been the use of double dierential detection. Double dierential detection would form a spatial
lter and would provide a signal that better isolates the peaks and valleys of the superimposed
MUAPs, which form the sEMG [2931]. However, this method would require the use of two triplets
of electrodes and more complicated front end amplication topologies. Single dierential detection
was selected in an eort to reduce electrode footprint, to maximise muscle applicability, and maintain
system simplicity.
As presented above, there are multiple ways of detecting and processing sEMG signals to
estimate MFCV. It is note worthy that every estimation method used will result to a dierent
MFCV measurement. Moreover, accuracy, sensitivity and repeatability of MFCV estimates depends
on several factors. A list of such factors is provided in Table 4.1. Valuing the eect of these factors
Table 4.1: Factors aecting MFCV estimates. Each primary factor call inuence MFCV
estimates depending on sEMG generation and detection [5].
Primary factors Factors inuencing sencitivity
electrode shape
electrode size
distance between bre and detecting electrodes
detection system
spatial lter
inter-electrode distance
electrode location
tissue inhomogeneities,
electrode-skin interface and electronics
additive noise
duration of signal epoch
bre components subcutaneous tissure layer thickness

length of bres
spread of innervation zones and tendon regions among motorunits
non-propagating potentials
cross-talk from nearby muscles
from other sources
on MFCV, an eort was made to keep all relevant parameters constant across all MFCV estimate
presented in this chapter, arising from simulations, sEMG recordings or system measurements.
Thus, all addressable factors had a uniform eect on all measurements, resulting in coherent and
reliable results.
4.3.1 Design Considerations for the CMOS Cross-Correlator
Adding to the list of advantages in MFCV estimation, cross-correlation is a strait forward operation
which can be implemented using custom CMOS logic in Application Specic Integrated Circuits
(ASICs). However, cross-correlation is a computationally intensive time-domain operation because
of the large amount of multiplications required. Cross-correlation is described by the following
equation, Eq. 4.6:
n
X
rl = x(i)y(i − l) (4.6)
i=1
Where l is the time shift between the two signals being correlated. For every point in time, all the
samples (n) in the correlation window must be multiplied and accumulated. In order to evaluate
the accuracy of the cross-correlator to be designed in CMOS, the MATLAB

® xcorr() function
was used. Simulations with a varying correlation window and sEMG sampling frequency were used
to nd the optimum parameters for the modelled cross-correlator. Simulations were conducted
using real retrospective sEMG data from the biceps brachii muscle of a test subject, under static,
non-fatiguing, contractions. All sEMG recording were made with the following parameters, unless
stated otherwise; for the recording of data an Octal BioAmp ML138 by ADInstruments was used,
along with pre-gelled electrodes provided by AMBU Neuroline (72000-S/25). The signal is bandpass
ltered 10-500 Hz. The sampling frequency was set to a high value of 10 k samples/s to allow further
post-processing (ltering, under-sampling) of the recordings. The Interelectrode Distance (IED) is
2 cm and the active electrode diameter is smaller than 10 mm. The IED is limited by the electrode's
width (20 mm).
Since the muscle contraction was a non-fatiguing one, the time delay between the two
recorded channels was constant and subsequently removed. Following that, an articial delay of 10
ms was introduced to one of the recordings to simulate an increase of the delay between the two
signals, similar to a fatiguing contraction. The delay was introduced articially in order to obtain a
linear increase in the signal's 65 s duration, thus creating a highly controlled test signal. The time
delay output of the ideal cross-correlation function xcorr() was compared to the linear t of the
added delay. In order to provide a comprehensible and representative error value between the two
datasets, the Mean Absolute Relative Dierence metric was used (MARD, in %). Table 4.2 presents
the resulting MARD between the xcorr() function and the linear t under dierent correlation win-
dows and sEMG sampling frequencies. A MARD value of 5% or less was deemed satisfactory for the
intended application. The optimum operating region of a cross-correlator used in MFCV estimation
is highlighted in green. Relaxing the operating parameters would result to a higher MARD value,
highlighted in orange.
The sEMG signal has a bandwidth of 10-500 Hz. In theory, a 1 kHz sampling rate with a 200
ms correlation window would suce for an accurate estimation of the time delay between the two
channels. As seen in Table 4.2, the accuracy of estimation is increasing with sampling frequency.
This is to be expected, since there are more data samples available to locate an accurate time delay.
With a sampling frequency of 1 kHz and a maximum detectable delay of 10 ms, there are only 10
available data samples for the estimation. However, this number increases to 80 using a sampling
frequency of 8 kHz. Moreover, accuracy is increasing with the correlation window up to the value
of 1.3 s and then starts to decrease again, resulting to an optimum window around 1 s.
Implementation of a cross-correlator that utilises all the bits of the digitised sEMG signal
Table 4.2: Error results between xcorr() and the linear t of the added time delay be-
tween the two sEMG recordings. Correlation windows and sEMG sampling frequency
are varied to obtain optimum working region.
Correlation Sampling Frequency (kHz)
Window (s) 1 2 3 4 5 6 7 8
0.1 382.90 228.32 95.65 73.40 18.90 11.00 13.82 6.03
0.2 428.68 76.46 24.52 15.74 9.62 7.67 6.03 4.70
0.3 251.63 37.14 22.17 14.84 9.11 7.13 5.18 3.72
0.4 181.21 33.74 20.36 14.01 9.34 5.83 4.77 3.51
0.5 117.33 32.60 19.54 13.41 8.73 5.69 3.62 2.88
0.6 96.09 31.84 20.39 12.33 7.57 5.10 3.10 2.58
0.7 96.66 32.41 19.66 11.82 7.87 5.29 2.73 2.03
0.8 76.41 30.25 18.29 14.26 7.96 3.54 2.83 2.31
0.9 74.28 32.45 19.54 11.99 7.72 4.00 2.90 2.23
1 75.82 30.12 18.38 11.89 7.15 3.84 2.09 1.49
1.1 70.09 29.88 18.36 12.26 6.40 3.68 1.93 2.27
1.2 78.46 29.65 19.66 10.38 8.18 2.56 2.16 1.73
1.3 72.07 29.36 17.67 10.46 6.61 2.77 1.86 2.54
1.4 73.12 29.87 17.17 10.79 4.81 3.40 2.51 2.01
1.5 76.02 30.26 19.79 9.96 4.85 2.51 1.77 5.01
Minimum Absolute Relative Dierence (%)
in an ASIC dedicated for muscle fatigue monitor would be a waste of resources, power and silicon
area. The sEMG signals can be converted to a discrete signal with the aid of a single threshold.
This work demonstrates that single threshold is an adequate alternative to digitizing the complete
sEMG signal, while retaining the necessary information for cross correlation and delay estimation.
The novelty of this research work lies in eliminating the need of cross-correlating the whole sEMG
signal, by cross-correlating a single bit approximation of it. Thus, complicated cross-correlator
architectures can be replaced by simple, bit-stream cross-correlator designs. Lande et al [33] de-
veloped a cross correlator which processes the output of a sigma delta modulator. The resulting
bit-stream from the modulator is then correlated with a template bit-stream stored in a buer.
Since a bit-stream consists of signal values of 1 (High) or 0 (Low), the complicated multiplications
required for cross-correlation are replace by AND gates. In this approach the input signal is buered
and similar to Eq. 4.6, for every time shift the buer window is correlated with a template signal.
This method vastly simplies the system, by minimizing the number of operations and transistors
needed, complexity and power consumption.
A system based on this approach seems very promising. In the case of cross correlating two
sEMG signals, the reference EMG signal (point A) is stored in one buer and the delayed sEMG
signal (point B) in another buer. Consider Fig. 4.5 for a graphical representation of the algorithm
Buffer a)
x(n) 1 2 3 ... n
Adder
temp xd(n) 1 2 3 ... n
Buffer
Figure 4.5: Lande et al. bit-stream cross correlator. The two signals are buered and latter
b) correlated
using an AND or XOR gate.
xd(n)
Correlation Stage
described above. A single bit converter is used to digitise the analogue sEMG signals and convert
x(n) Delay TheDelay

them to two digital bit-streams.
Delay
type of converter used isDelay
...
a comparator. The signal from the
second buer is shifted one sample at the time, applying the XOR function over the selected window.
Counter n
Counter 1
Counter 2
Counter 3
The result is the sum of all the zeros computed by the logic function. The same modied sEMG
signals used for Table 4.2 were converted to a bit-stream using a comparator with a threshold of 0V.
The signals were previously centred around 0 V. Table 4.3 presents the resulting MARD between the
xcorr function of the bit-streams and the linear t under dierent correlation windows and sEMG
sampling frequencies. It was observed that the operating window and resulting MARD values of the
bit-stream cross-correlation were very close to the original sEMG cross-correlation results. Hence,
a bit-stream cross-correlator is a viable solution for an MFCV tracking system in CMOS.
Table 4.4 shows the MARD results of the bit stream modulator presented in [33]. This
approach temporarily stores an sEMG bit-stream epoch in a buer. The buer length is given by
Sampling − F requency × correlation − window. By interpreting Table 4.4, the minimum buer
length would be at a sampling frequency of 8 kHz and a correlation window of 0.3 s. The resulting
buer length would be 2.4 k samples. Since two buers of the same length are needed, 4.8 k
samples need to be stored. Finally, another 2400 gates are required to perform the cross-correlation
operation. Implementing this type of bit-stream cross-correlator in CMOS would require 7200 digital
blocks (gates, memory). Such space requirements are very big for custom ASICs as silicon area is
very expensive. The MATLAB

® code for the selection of optimum window size and sampling
frequency for the bit-stream cross-correlator is shown in Appendix C.

Table 4.3: Error results between xcorr() and the linear t of the added time delay
between the two resulting bit-streams. Correlation windows and sEMG sampling fre-
quency are varied to obtain optimum working region.
Window (s) 1 2 3 4 5 6 7 8
0.1 347.15 266.14 109.34 47.26 21.38 10.71 14.72 6.38
0.2 441.93 112.31 28.91 16.38 9.46 7.45 6.51 5.29
0.3 250.85 43.76 24.03 15.04 8.63 8.40 5.87 3.74
0.4 180.13 39.26 21.55 14.66 8.66 6.17 4.80 4.09
0.5 144.55 34.47 18.36 12.86 9.35 5.98 4.22 3.23
0.6 122.71 33.39 18.52 12.61 8.70 5.81 3.41 2.72
0.7 106.43 34.25 19.18 12.12 8.84 6.04 2.73 2.27
0.8 86.58 32.65 18.06 12.77 9.18 3.68 2.83 2.64
0.9 88.20 33.27 18.82 12.62 8.01 4.13 2.94 2.22
1 85.34 31.50 17.85 12.08 7.17 3.83 2.11 1.57
1.1 80.12 29.69 18.74 12.83 7.66 3.77 1.97 2.36
1.2 83.68 31.66 18.68 10.92 7.20 2.85 2.16 1.76
1.3 87.49 32.12 18.22 10.70 4.20 3.16 1.89 2.56
1.4 86.42 30.04 16.79 10.99 4.86 3.39 2.43 2.02
1.5 78.31 32.37 17.79 10.01 4.90 2.56 1.78 5.02
Table 4.4: Error results between the cross-correlator proposed in [33] and the linear t
of the added time delay between the two resulting bit-streams. Correlation windows
and sEMG sampling frequency are varied to obtain optimum working region.
Window (s) 1 2 3 4 5 6 7 8
0.1 124.56 64.57 35.25 24.04 17.18 12.72 10.45 8.28
0.2 116.26 48.55 27.93 17.52 11.93 10.51 7.31 6.22
0.3 113.93 44.98 25.74 16.08 9.03 7.82 7.00 4.93
0.4 108.04 46.05 24.18 15.84 10.57 7.65 6.32 5.06
0.5 104.13 38.87 23.57 13.99 8.81 7.43 5.16 4.55
0.6 104.37 44.91 24.74 12.36 8.28 6.97 4.52 3.43
0.7 100.22 39.66 23.66 12.18 9.02 7.23 3.24 3.07
0.8 101.21 38.35 22.06 14.84 7.96 5.76 3.44 3.81
0.9 99.06 41.85 23.15 14.54 9.66 4.78 3.79 2.34
1 91.23 38.90 22.74 12.62 8.55 4.47 2.37 1.81
1.1 90.44 38.82 22.28 13.63 8.50 4.55 2.10 2.56
1.2 91.26 37.05 23.56 12.68 7.44 3.74 2.38 1.90
1.3 85.93 40.42 22.41 12.90 7.32 4.34 2.17 2.86
1.4 86.73 38.33 20.90 13.09 5.80 3.70 2.77 2.24
1.5 86.57 37.69 24.92 11.37 5.75 2.73 2.00 7.55
4.3.2 Proposed Bit-Stream Cross-Correlator
The aim of the MFCV cross-correlator is to estimate the time delay between the two bit-stream
signals x(n) and xd (n). x(n) is the digitised EMG signal detected at point A (see Fig. 4.4(b)) and
xd (n) is a delayed version of x(n) detected at point B. Thus, there is no need to buer xd (n) for
cross-correlation since it is already shifted in time. Instead of storing and shifting data to nd a
time lag, it is possible to continuously return the similarity of the two signals for a given time lag
by counting all the time instances where the two signals are the same. The proposed bit-stream
architecture is similar to the operation of correlator banks [34]. Several discrete time lags for the
cross-correlation output are obtained by continuously delaying the input signal and repeating the
aforementioned process. Thus, obtaining a similarity (correlation) result for every discrete time
delay. Finally, the time lag between the two signals is returned by the counter with the larger
value. This approach greatly simplies the number of transistors required. However, the time lag
resolution is limited. The time lag dynamic range depends on the number of time delays introduced
to the system. Furthermore, the resolution of the system can be varied according the sampling
frequency (T = Delay) and the cross-correlation time window. Consider Fig. 4.6 for the proposed
bit-stream correlator which comprises of several correlation stages, where each stage has a delay
block, a counter and a gate responsible for the cross-correlation. The delay line is constructed
by connecting serially several unit delays" (discrete time-lags). The two bit-streams are cross-
correlated using an XNOR function and an accumulator at the output of every unit delay". Each
accumulator increases its current value by 1 when xd (n) matches the corresponding time-lag Dn .
When the reference signal x(n) reaches the incremental delay which matches the time delay between
x(n) and xd (n), then the accumulator value corresponding to that incremental delay will be the
maximum out of every accumulator along the delay path. Thus, the computed delay as a result of
cross correlating the two signals is given by Eq. 4.7 where n is the number of discrete time-lag, C
is the accumulator value and i is the time of evaluation.

Dn,i = max(C1 , C2 ...Cn ) (4.7)
t=i
It is observed that in the proposed architecture the buer size of x(n) in [33] is reduced to
the time delay between the two signals. In this approach, buering of the two signals is perceived
in time. Instead of saving the incoming data, they are processed in real-time for the duration of a
Adde
temp xd(n) 1 2 3 ... n
Buffer
b)
xd(n)
Correlation Stage
x(n) Delay Delay Delay ... Delay
Counter n
Counter 1
Counter 2
Counter 3
Figure 4.6: Proposed architecture of the bit-stream cross correlator. The buering blocks have
been replaced with delay blocks.
specied time window. Periodic reset of the accumulators is necessary in order to obtain a new cross
correlation result. The disadvantage of this architecture is the loss of temporal accuracy. However,
there is a trade-o between temporal and delay-estimation accuracy. The longer the reset period,
the more data points the system accumulates. Thus, time-averaging is performed, resulting to a
more indicative estimation of the constantly varying delay between the two signals. The advantages
of these approach are the following. There is a signicant reduction of the number of transistors
needed. This will have a positive impact on power consumption and silicon area. Furthermore, the
computational complexity of the correlator is reduced to a minimum.
In order to decide the number of correlations stages to be added to the system, the following
assumptions were made. The longest delay would occur at the end of a fatiguing contraction.
Physiological values reported in literature for a fatiguing isometric contraction indicate that MFCV
m
very rarely is measured to be less than 3
s [17, 3539]. Referring back to Table 4.3, a sampling
(shifting) frequency of 6 kHz was selected in order to achieve a good accuracy (low MARD values)
while keeping the required delay buer length to a minimum. Thus, the delay buer was designed
with a length of 40 correlation-stages. The correlation window in not inuenced by the buer length.
Hence, the trade-o between buer length and correlation window is swapped with the trade-o
between maximum detection window and sampling frequency. However, increasing the detection
window in the proposed bit-stream cross-correlator is less costly that increasing the correlation
window in the architecture proposed in [33].
The dynamic range of the system can be varied according the sampling frequency and the
delay estimation accuracy according to the cross-correlation time window. In order to characterise
the system, it was modelled in SIMULINK and compared with MATLAB

® xcorr() function using
Mean Ab
2.5
4.3 MFCV Tracking Algorithm 1.5

0 5 10 15 20 25 90
30
Comparator Threshold over Commom Mode (mV)
5 9
W = 0.5
W = 0.66
4.5 8 W=1
Mean Absolute Relative Difference (%)

W = 1.5
W=2
4 7
3.5 6
3 5
2.5 4
2 3
1.5 2
0 5 10 15 20 25 30 1000 2000 3000 4000 5000 6000 7000 8000
Comparator Threshold over Commom Mode (mV) Sampling Frequency (Hz)
(a) (b)
9
Figure 4.7: Performance of the proposedWWbit-stream
= 0.5 cross-correlator with varying threshold, correla-
®
tion8 window and sampling frequency: (a) MARD
= 0.66
between MATLAB xcorr() function and modelled
system with a varying threshold, (b) MARD
W=1
W = 1.5 between MATLAB
® xcorr() function and modelled
system,
7 where W is the correlation window (s).
W = 2
real sEMG data from the biceps brachii muscle of a test subject during a static, fatiguing contraction.
5
The sEMG signal was amplied to a peak-to-peak voltage of 300 mV. Fig. 4.7a shows the MARD
®
4
between the modelled system and MATLAB simulations for a varying comparator threshold. It
3
can be seen that the relative error has a small sensitivity to the chosen threshold level. The modelled
2
1000 2000 3000 4000 5000 6000 7000 8000
system's output is Sampling
compared to
Frequency the results or the
(Hz) xcorr() function. By looking into Table 4.2, a
window of 1s was used in the xcorr() function as it exhibited good accuracy across a wide range
of sampling frequencies. Fig. 4.7b shows a parametric simulation with the sampling frequency and
correlation window as variables. Increasing the sampling frequency results to higher accuracy but
limits the dynamic range of the system. The results show that sEMG bit-stream cross-correlation
is an accurate representation of cross-correlation which allows it to used for tracking MFCV. A
longer correlation window would yield better delay estimation accuracy but would result to a slowly
tracking system. The proposed approach of this bit-stream cross-correlator resembles the basic
architecture of a time-to-digital converter. According to [40], the measurement interval ∆T in a
time-to-digital converter can be expressed as in Eq. 4.8, where Tc is the period of the clock. The
error in the measured time interval ∆T is r and can be equal to twice the clock period. Thus, a
delay chain length of 40 yields a relative error of 5%, which is considered acceptable for the intended
application.
INC INC
RST RST RST
counter x 39 counter
14
Out (1) Out (N)
Stages 3-39
Correlation Stage 1 Correlation Stage 2 Correlation Stage 40
{
xd(n)
x(n) D Q D Q D Q
CLK CLK CLK CLK
INC INC INC

RST RST RST RST RST RST RST
counter counter counter
14 14 14
Out (1) Out (2) Out (40)
Figure 4.8: Correlation stage with delay block, counter and correlator (XNOR).
∆T = N Tc + r
(4.8)
r ∈ [−Tc ; +Tc ]
Correlation Stage
As mentioned in the section above, the bit-stream correlator is made of several correlation stages
connected in series. Each stage comprises of a delay block, a counter and a gate responsible for
the cross-correlation. The cross-correlator proposed by Lande et al. utilised an inverter chain as
a buer. However, it was designed to operate at a much higher frequency than the one required
for this application. As such, an inverter chain would be too long, increasing silicon area and
power consumption. A simple D-type Flip Flop (DFF) acts as a delay block, and the delay time
is controlled by the sampling frequency of the system (CLK). A 14 bit ripple counter was found
to be enough to meet the requirements of the system. The use of XNOR gates as a bit correlator
improves the original AND gate design by taking all possible digital cases into consideration. The
circuit schematic is shown in Fig. 4.8. A second clock (RST) with a much lower frequency than
CLK is used to reset the counters and thus dene the cross-correlation time window. The dynamic
range of the system can be varied according the sampling frequency CLK and the delay estimation
accuracy according to the cross-correlation time window (RST).
Fig. 4.9 shows the signal propagation in the system. Fig. 4.9(a) shows an EMG signal segment
as detected by the biopotential front end ampliers. Fig. 4.9(b) shows the resulting bit-stream after
amplitude (mV)
2
a) 0
−2
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
2
amplitude (V)
b) 1
0
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
2
amplitude (V)
c) 1
0
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
2
amplitude (V)
d) 1
0
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
2
amplitude (V)
e) 1
0
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
200
f)
counter
100
0
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Time (s)
Figure 4.9: Correlation stage signals. (a) EMG signal, (b) xd (n) bit-stream, (c) x(n) bit-stream
delayed by 10 previous correlation stages, (d) x(n) bit-stream delayed by all previous stages, (e)
XNOR bit correlation result between (b) and (c), (e) counter result after adding all XNOR values in
real time.
single bit digitization. Fig. 4.9(c) and (d) show the propagation of the bit-stream trough the delay
blocks of the system. Finally, Fig. 4.9(e) and (f ) show the correlation results and how these are
interpreted by the counters in real time. At the end of the correlation time window, the maximum
value counter is selected. The delay this counter is linked to, represents the time lag between the
input signals. After that, the counters are reset to zero and correlation starts again.
Maximum Detector Stage
At the end of the correlation time window (RST), all the counters of the system are read. The
correlation stage (i.e. delay) of the counter with the maximum value best represents the time lag
between the two input signals. A maximum function compares all the counters in cycles. The
maximum function comprises of smaller blocks, each comparing two 14 bit numbers, and operates
using sequential logic. It starts by comparing all the results in pairs and then proceeds with
evaluating the results of the last comparison. The maximum function operates using sequential
out (1-40) from counters
& & & & & & & & & & & & enable
max(0) max(1) max(2) max(3) max(4-17) max(18) max(19)

1 f0 f1 f2 f3 f 18 f 19
max(20) max(21) max(22-28) max(29)
2 f 20 f 21 f 29
max(28)
max(30) max(31-33) max(34)

3 f 30 f 34
max(31) max(32) max(33)
max(35) max(36)
4 f 35 f 36
max(37) max(38)
5 f 0-38 f 37 f 38
thermometer coding out(1-40)
Figure 4.10: Maximum Function Flow.
logic and is shown in Fig. 4.10. Every maximum operation returns a binary ag, which passes down
to the next comparison and indicates which of the two compared numbers is the maximum. Thus, a
binary one means the rst of two numbers is bigger. This way, the counter position number (hence
delay number) and not the counter value is returned when the operation is nished. Consider Eq. 4.9
as an example for calculating the counter path. The maximum detector was generated using VHDL
coding, shown in Appendix B. The max function is enabled only prior to the counter evaluation
stage, thus saving power. All the counters and delay blocks stop processing, in order to allow the
max function to settle to one output (sequential logic). Then, the system output is ready to read.
out(1) = f38 ∧ f37 ∧ f35 ∧ f30 ∧ f20 ∧ f0
out(2) = f38 ∧ f37 ∧ f35 ∧ f30 ∧ f20 ∧ f 0
out(3) = f38 ∧ f37 ∧ f35 ∧ f30 ∧ f 20 ∧ f1

(4.9)
.
.
.
out(8) = f38 ∧ f37 ∧ f35 ∧ f 30 ∧ f 21 ∧ f 3

.
.
.
4.4 MFCV Tracking ASIC Architecture 94
enable CLK READY RST
del
CLK2 D Q D Q D Q D Q D Q D Q D Q
CLK Q CLK Q CLK Q CLK Q CLK Q CLK Q CLK Q
CLK1
reset
Figure 4.11: Timing control unit circuit diagram.
Timing Control Unit
The timing control unit governs all the functions of the system. As previously mentioned, when
the time to calculate the delay comes (CLK2) the counters and delays need to stop functioning, the
max function needs to be enabled and when the result is ready, the output should be loaded and
sent out using a standard Serial Peripheral Interface (SPI).
All the operations of the MFCV estimator are orchestrated by the timing control unit, shown
in Fig. 4.11. The inputs to this block are two clocks (sampling clock and correlation time window
clock). When the correlation time window clock is asserted, the delay (DFF shifting) of the signal
stops, hence the counters stop operating. The maximum function uses an enable signal as a trigger.
The READY signal is delayed to allow the maximum function to settle. When the operation is
nished, the RST resets all the counters.
4.4 MFCV Tracking ASIC Architecture
The developed MFCV ASIC (code name Optimus) shown in Fig. 4.13 consists of four major building
blocks: a) a sEMG dual channel Instrumentation Amplier (IA), b) two Sallen Key low-pass lters,
c) a bit-stream converter comprised by two analogue comparators and d) a digital bit-stream cross-
correlator. In addition, the MFCV ASIC includes a bias circuit generator, a digital timing control
circuit and a SPI to communicate with any receiving circuit.
The operation of the MFCV ASIC can be described as follows: the dual IA amplies the
detected sEMG signals which are then low-pass ltered to extract the signal attributes in the
required frequency band (10-500 Hz). Following that, the bit-stream converter digitises the sEMG
signals and feeds them to the bit-stream cross-correlator that computes the time delay between
them.
amplitude (V)
a) 2
0
0 2 4 6 8 10 12 14 16 18 20
4
amplitude (V)
b) 2
0
0 2 4 6 8 10 12 14 16 18 20
4
amplitude (V)
c) 2
0
0 2 4 6 8 10 12 14 16 18 20
4
amplitude (V)
d) 2
0
0 2 4 6 8 10 12 14 16 18 20
4
amplitude (V)
e) 2
0
0 2 4 6 8 10 12 14 16 18 20
4
amplitude (V)
f) 2
0
0 2 4 6 8 10 12 14 16 18 20
Time (ms)
Figure 4.12: Timing Control Unit signal ow. (a) sampling clock, (b) correlation time clock, (c)
delay clock, (d) enable maximum function, (e) read data for output and (f) reset counters.
Analog Out 1 Threshold 1 Test 1 CLK 1 CLK 2

MFCV Tracking
Synchronisation
ASIC
R2 Bias Generator Timing Control

C3
R2
R2 R2 R3
R4 R2
R2 C2
Sampling
El1 C1 Correlation Correlation

R2 Stage x40 Stage
R1 R1
El2 a) sEMG readout Stage b) Filtering Stage c) Bit-Stream

Counter Counter
Output Delay
X
Epidermis
Enable Max SPI

El3
High-Pass d) Bit-Stream Cross

Filter Max Selectror
Correlator
Bias Generator
Vref : Setting DC, close
proximity to electrodes
Analog Out 2 Threshold 2 Test 2
R1 = 10 MΩ, R2 = 100 KΩ, R3 = 1 MΩ, R4 = variable, C1 = 1.5 nF, C2,3 = 33 pF
Figure 4.13: System block architecture of the Muscle Fibre Conduction Velocity tracking ASIC.
4.4.1 sEMG Signal Processing
The circuit has been implemented in a commercially available 0.35 µm CMOS technology provided
by AMS (C35B4). Standard cell libraries were used for the implementation of the operational
ampliers and comparators in stages a, b and c (OP05B, COMP) as the focus of this chapter is the
bit-stream cross-correlation algorithm application for MFCV tracking and system implementation.
A common front-end architecture in biopotential measurements is a DC-coupled fully dier-
ential amplier followed by a dierence amplier, as in the classical three opamp instrumentation
amplier [4146]. The equivalent input noise only depends on the two opamps constituting the fully
dierential amplier. Instrumentation ampliers should be capable of rejecting up to 300 mV DC
Polarization Voltage (PV) from the biopotential electrodes [46, 47]. Thus, AC coupling is needed
to avoid saturating the output due to input DC osets. Existing architectures that can implement
rejection of such high DC voltages without using o-the-shelf components lead to lower performance
IAs [48, 49] or suer from trade-o 's resulting in reduced PV [5052]. On the other hand, the use
of conventional, simple, o-chip high-pass lters signicantly reduces the input impedance, where
especially the common mode input impedance is very important for achieving high Common Mode
Rejection Ratio (CMRR).
Thus, an external oating high-pass lter is used, as shown in Fig. 4.13. The intended appli-
cation of the ASIC as a wearable device allows the use of external components. The main advantage
of oating high-pass lter compared to conventional passive high-pass lters is the elimination of the
grounded resistor, implementing very large common mode input impedance [46, 47]. With R1=10
MΩ and C1=1.5 nF the resulting cuto is 10.6 Hz to best lter out motion artefacts [53]. However,
this lter structure requires a fourth electrode to bias the input lter structure and set the DC
voltage of the body around the electrodes. This reference DC level must satisfy the amplier's
input common mode range requirements and is set at a voltage Vref = Vsupply /2. The common
mode voltage of the detected sEMG signal feeding the ampliers is the averaged DC electrodes
(1-3) voltage at node X (Fig. 4.13). Since the input network is not grounded, when a common
mode input voltage is applied, no currents ow through the network.
The low-pass lter of Sallen Key topology is implemented with a cuto of 2.5 kHz, with
R2=100 kΩ, R3=1 MΩ and C2, C3=33 pF. The structure is duplicated to allow two channel
operation. The reference voltages of the two comparators are kept separate to allow oset mismatch
compensation. Furthermore, the feedback resistors responsible for the amplier gain (R4) were not
4.5 MFCV Tracking ASIC Experimental Results 97
Figure 4.14: Die micro-photograph of the MFCV tracking ASIC.
implemented on the IC, but left to be completed with external components to allow gain exibility.
4.5 MFCV Tracking ASIC Experimental Results
The MFCV tracking ASIC micro-photograph is shown in Fig. 4.14. The layout area for the digital
core is 1122 µm by 494 µm and for the analogue front end is 800 µm by 500 µm, with a total area of
1 mm2 . 1 The total static and dynamic power consumption of the digital core (clocked at 10 kHz)
is 1 µW and of the analogue front end is 2.071 mW from a 3.3 V supply. The power consumption
of the analogue front end includes the power of two IAs, two high-pass lters and two bit-stream
converters.
4.5.1 System Validation
The characterization of the IA is performed at the output of the sEMG readout stage. Fig. 4.15a
shows that the measured integrated input-referred noise of the IA (0.1-2.5 kHz) is 3.579 µV with
nV
a noise oor at 135 √ . The high corner frequency (300 Hz) is due to the choice of standard cell
Hz
1 Layout and vhdl designs for sections of the bit-stream cross-correlator were designed by Vlad Cretu under the
direct supervision of Ermis Koutsos.
Table 4.5: MFCV ASIC parameters, size and power breakdown (C35, 3.3V).
Core Measured Current Size
Analogue 616 µA 0.4 mm2
Digital 290 nA 0.554 mm
2
Biasing 11.5 µA -
Total Power 2.071 mW -
libraries for the ampliers. The noise measurements originate from a single chip due to limitations
on the number of dies. The MATLAB

® code used for the noise analysis is shown in Appendix C.
The sEMG has a range of 50 µV to 5 mV. Thus, the presented front end biopotential amplier
is more that adequate to amplify sEMG signals for the proposed application. Fig. 4.15b shows
the measured dierential and common-mode gain of the IA followed by the high-pass lter. The
external oating low pass lter was included during measurements. The resulting CMRR is 85.24
dB in the bandwidth of interest (83.1 dB at 50 Hz).
Following that, in order to verify the functionality of the system, the front end amplier was
disconnected from the digital core. Thus, the bit-stream cross correlator was tested on its own. Two
square signals of amplitude 3.3 Vp−p and frequency 100 Hz were placed on the inputs of Channel 1
and Channel 2. The aim was to resemble a prefect case scenario of amplied and digitised sEMG
signals. The reset clock (RST ) of the system was set to 1 Hz to maximise the correlation window
accuracy. The sampling clock (CLK ) was set to 10 kHz. Starting from a phase delay between
the two pulses of 0 s, it was gradually increased all the way to 3.9 ms with 1 µs increments. The
measured delay appeared on the output in the form of a step ladder. The mean time to switch from
one unit delay to another was 101.3 µs, close to the ideal time of 100 µs, resulting in a mean error
of less that 2 µs with a standard deviation of 7.4 µs.
A segment of the detected and amplied sEMG signals is shown in Fig. 4.16. The delay
between the two signals can be observed visually. However, the cross-correlator will perform this
task more accurately. It can be seen that Channel 1 and Channel 2, although have the same gain,
do not have the same peak-to-peak amplitude. If the middle electrode in bipolar (single dierential)
amplication was not shared, the amplitude mismatch would be worse. Since the cross-correlation
algorithm is not very sensitive to threshold levels it is believed that this issue did not pose any
problems. The bottom graph shows the output of the comparator from Channel 2 for reference.
The common mode of the two ampliers was not the same, due to amplied osets. An external C-R
−5 60
10
Differential
−6
10 40
Common Mode
−7
135nV
Power spectrum (V/sqrt(Hz))
10
20
−8
10
Gain (dB)
0 CMRR = 85.24 dB
−9
10
−20
−10
10
Simulation
−11
Measurement −40
10
−12
10 −60
−4 −2 0 2 4 6 −1 0 1 2 3 4 5
10 10 10 10 10 10 10 10 10 10 10 10 10
(a) (b)
Figure 4.15: (a) Input-referred noise measurement (red) and simulation (black) results of the IA
channel 1, (b) Common-mode rejection ratio measurement of the IA (chan. 1) using a 500mVpp input
signal.
lter restored this value to 1.65 V. Thus, the applied threshold can be the same for both channels
and must be larger that 1.65 + EM Gnoise .
A volunteer was asked to lift a 10 kg weight with their biceps while standing until exhaustion.
The sEMG signal was recorded from the inner left biceps brachii muscle. Furthermore, the test
subject was asked to perform the task two times repetitively. However, the subject was not given
enough time to completely rest between the two repetitions. Thus, some visible changes are expected
in the MFCV from these two repetitions. The trial results for the estimated delay are presented in
Fig. 4.17. It is clear that the second lift, the subject could not hold the weight for as long as in the
rst lift. Furthermore, the estimated delay starts from a higher value the second time, indicating that
the subject was already fatigued when again starting the test. The threshold for the comparators
was set to 1.655 V, such as it stands above the noise oor. To verify that, the comparator output
was being monitored. The goal was to set the threshold to the level that minimised any accidental
noise crossings.
In order to study the eect of threshold on MFCV estimation, the two amplied sEMG
signals from this test were saved at a sampling frequency of 2 kHz. Following that, using a waveform
generator (Tektronix AFG3000) the sEMG signals were played back to the system in a loop. This
way, it was made sure that the same response is expected every time. On every iteration, the
threshold voltage was increased. Consider Fig. 4.18 for the eect of V ref (threshold) on the
Time Delay
Amplitude (V)
1.65
1.6
200 400 600 800 1000 1200 1400

Amplitude (V)
1.65
1.6
200 400 600 800 1000 1200 1400
3
Amplitude (V)
0
200 400 600 800 1000 1200 1400
Time (samples, f = 2KHz)
Figure 4.16: Top: sMEG segment from Channel 1, middle: sMEG segment from Channel 2. Delay
between the two channels is easily detectable by eye, bottom: Comparator output from channel 2 for
reference.
Static Contractions with resting phase

40
Cross−Correlation Delay in units (unit = 1/f, f = 10KHz)
35
30
25
20
Lack of Rest
15
10
5 Wrong
exercise
0
0 20 40 60 80 100 120 140 160 180
Time (sec)
Figure 4.17: Cross-correlation delay estimation under static contraction from rst test subject.
sEMG was recorded from the inner left biceps brachii. Sampling clock is 10 kHz and window size
clock is 1 Hz.
Static Contraction EMG data playback with varying Vth from 1.654 V to 1.670 V
40
Cross−Correlation Delay in units (unit = 1/f, f = 6.25KHz)

35
30
25
20
15
10
0
0 100 200 300 400 500 600 700 800
Time (sec)
Figure 4.18: The recorded sEMG signals from Subject 2 are played back to the system using a
funciton generator. Thus, the eect of thresshold can be studied.
estimated delay. It can be seen that the optimum threshold lies within the region 1.657-1.667
V. The main limitation for the use of such a low sampling frequency was the on-board memory
of the waveform generators. It is believed that the limited resolution had a small impact on the
measurements.
4.5.2 Validation on Human Subjects
Twenty (20) volunteers (4 female, 16 male) were used to evaluate the performance of the MFCV
tracking ASIC. This work has been approved by the Imperial College Joint Research Compliance
Oce ethics committee (ICREC ref: 15IC2481).
4.5.2.1 Experimental Protocol
Subjects were seated or standing, depending on their preference, with the back and elbow xed
against a wall in order to minimise compensatory movements. Subjects were required to keep in
contact with the wall throughout the whole duration of the testing. Consider Fig. 4.19 for a graphical
representation of the experimental set-up. The subject specic calibration of EMG sensors involved
the performance of ve Maximum Voluntary Contractions (MVCs). Following that, the subjects
were asked to perform isometric contractions by pulling against a handle attached to an electronic
scale. The bottom end of the scale was attached to the ground. Subjects were required to sustain a
Output Force
sEMG ch1
sEMG ch2
Electrodes
Bit-Stream
Detailed force
output with
limmits
Force Monitor
monitor
8-ch ADC PC
Amplified sEMG USB link
ASIC clk, SPI,

uC
threshold
Figure 4.19: Experimental set-up protocol. The subjects keep the force output constant and within
limmits with the aid of a visual, real-time representation of the applied force.
constant force, pre-set at 70% of their MVC force, for as long as they deemed possible. After the force
reading dropped more than 10% of the pre-set level, the experiment would stop. A normal luggage
scale was used as a force monitor. The scale was modied by adding an instrumentation amplier
to the scale's resistor bridge network that monitors force. The voltage output of the modied scale
was sampled by the same ADC that sampled the two amplied sEMG outputs of the MFCV ASIC.
A visual representation of the bit-stream allowed for correct threshold positioning. Furthermore,
a detailed and real-time plot of the force output with visible limits helped the subjects monitor
their force output and keep it steady between the acceptable limits. The force boundaries were
adjusted for each subject according to their MVC levels. Four disposable surface EMG electrodes
were placed on the skin of the participant's arm to monitor activity of the biceps brachii muscle,
according to Fig. 4.20. Prior to electrode placement, the skin was prepared by cleaning it with
medical alcohol. A detailed architectural diagram and a photograph of the experimental setup are
also shown in Fig. 4.20.
A requirement of the system is the positioning of the two independent thresholds for the bit-
stream converters. The thresholds could be placed at the common-mode voltage, as in the analysis
System Architecture Left Bicept Brachii Experimental Setup
Innervation Zone
Faraday Cage
Myotendinous Zone
Battery
PCB uC Board
Clock IED
ASIC
SPI Data
1
Threshold 2 3
A Battery Pack
B
PCB with voltage regulator and floating filter
Muscle Fibre
USB Data
Dirrection
Vref
KL26z Freedom Board

Electrodes Faraday Cage
PC
Figure 4.20: Left: Architectural diagram of experimental setup. Middle: Electrode conguration for
bipolar single dierential sEMG amplication. Muscle zones are displayed. Right: Experimental setup
including microcontroller for data transmission, clock generation and threshold position. Faraday cage
used in noise measurements only.
of the modelled system. However, in order to avoid noise in the absence of sEMG to contribute to
MFCV estimation, the thresholds are placed just above the noise oor. The threshold levels should
be low enough to be closer to the base of the sEMG MUAP spike and at the same time be above
the peak noise amplitude. This could also be achieved with the use of a comparator with hysteresis.
4.5.2.2 Results from Human Trials
The accuracy of the ASIC is established by a direct comparison between the ASIC and MATLAB
® .
The amplied sEMG signal from the ASIC is recorded using a 16-bit ADC by ADInstruments. Fol-
lowing that, the MFCV is computed in MATLAB

® using the xcorr() function and then compared
to the ASIC estimate.
An example of the estimated MFCV from MATLAB

® and that measured form the ASIC
from one subject is presented in Fig. 4.21a. The subject starts the static contraction form a rested
m
state, with an MFCV of 6
s . As the subject gets tired, the MFCV drops to lower values, until the
point that the subject is completely fatigued and stops the experiment. The MFCV has reached a
m
new lower value of 4
s.
Table 4.6 presents the change in MFCV from a rested state to a complete fatigue state of the
muscle for all 20 subjects. The estimated values of MFCV decrease in all subjects during fatigue,
as seen in Table 4.6 (B), (C) by the relative change in MFCV from the ASIC and MATLAB
®
respectively. Although MFCV is not always monotonically decreasing, the end point is at a lower
Table 4.6: Subject trial results: a) subject number, b) change in MFCV from chip, c)
change in MFCV from MATLAB® , d) MARD and e) MVC force.
A) # 1 2 3 4 5 6 7 8 9 10
B) (m/s) 1.13 1.40 2.38 2.13 1.59 3.01 1.64 4.11 0.88 4.80
C) (m/s) 1.22 1.71 2.66 3.28 1.69 3.48 1.32 2.93 1.01 5.47
D) % 1.12 1.77 1.26 2.56 2.67 6.47 2.37 3.04 0.93 9.64
E) Kg 19.5 23.0 23.5 23.5 13.2 23.5 15.0 13.5 20.5 16.0
A) # 11 12 13 14 15 16 17 18 19 20
B) (m/s) 1.78 5.72 2.6 20.67 1.32 2.13 0.48 1.09 3.35 2.28
C) (m/s) 2.14 5.73 2.85 18.41 1.44 2.27 2.15 0.99 3.54 2.35
D) % 1.60 2.08 1.86 5.38 1.34 1.39 3.11 3.63 6.78 3.70
E) Kg 28.5 27.6 25.4 24.6 36.0 32.0 19.0 34.0 35.0 17.6
Table 4.7: Comparison between MFCV physiological values.

Literature [17] [35] [36] [37] [38] [39] This
work
Mean/Range 5.1 4.2 2.6 4.64 2.53 4.4 5.4
MFCV (m/s) 5.3 4.87
value than the starting point in all subjects. The relative error between MATLAB
® and the ASIC
is estimated across the complete MFCV trends and not only on the relative change of MFCV and
is shown in Table 4.6 (D). Time lag is converted to velocity using Eq. 4.10, where CLK1 is the
sampling frequency, OutputDelay# is the ASIC delay estimate between 1-40 and IED is the Inter
Electrode Distance (Fig. 4.20, IED = 2 cm). The estimated MFCV values match the physiological
values reported in literature for a fatiguing isometric contraction [17, 3539]. The reported MFCV
values are compared with the MFCV values seen in this study in Table 4.7.
The correlation window was set to 1 second (CLK2=1 Hz) to maximise the accuracy of
the bit-stream cross-correlator. The sampling/delay clock was set to 6.25 kHz to establish a wide
MFCV dynamic range. The same parameters were imported in MATLAB

® for analysis. The
ASIC is found to have a Mean Absolute Relative Error (MARD) of 3.2%. Subject specic MARD
results are presented in Table 4.6. The two MFCV trends (MFCV: MATLAB
® , MFCV: ASIC)
are compared point by point and the distribution of comparison data is displayed in a box plot in
Fig. 4.21b.
m IED · CLK1

CV = (4.10)
sec OutputDelay#
6.5
Matlab Simulation
35
MFCV ASIC
6
30
Absolute Relative Difference (%)

5.5
25
Rested State
MFCV (m/s)
5
20
4.5 15
4 10
5
3.5
Fatigued State
0
3
0 5 10 15 20 25 30 35 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Time (s) Subject Number (#)
(a) (b)
Figure 4.21: (a) MFCV trial results from subject 11. The rested state MFCV is 6 sec
m
and progresses
to 4 sec as the muscle fatigues, (b) Fatigue trial results comparison between MFCV tracking ASIC and
MATLAB® . The MARD is presented in percentile.
m
4.5.3 Discussion
The ASIC is capable of estimating MFCV with a mean relative error of 3.2% compared to
MATLAB
® analysis. A large proportion of the error can be attributed to the nature of the
bit-stream cross-correlator, as it resembles a time-to-digital converter. Hence, some error arises
by quantising in time with limited resolution, as the bit stream cross-correlator has a total of 40
quantisation values. Furthermore, misalignment between ASIC output data and MATLAB
® esti-
mations gives rise to unexpected error. Nevertheless, the system provides good estimation accuracy
which is suitable for the purposes of muscle fatigue monitoring. Increasing the number of delays
(above 40) will yield a wider dynamic range and allow higher sampling clock rate thus achieving
higher estimation accuracy. This however will have a direct impact on the ASIC size.
The systems performance specications are summarised in Table 4.8. The MFCV tracking
ASIC realises a real time cross-correlator with a very small footprint of 0.55 mm2 and an overall size
of less than 1 mm2 . Furthermore, it oers a large degree of freedom by allowing user exibility on the
correlation time window, sampling frequency and amplier gain. With a low power consumption
of 628 µA, the system is particularly suitable for wearable fatigue monitors. Implementation of
the sEMG cross-correlator using a low power microcontroller (Nordic Semiconductor nRF51822, 4
kHz ADC) yielded a signicantly larger dynamic power consumption of 4.8 mA. Implementing the
MFCV tracking algorithm in an ASIC enables the reduction of data transmission power costs, as
4.6 Wearable Muscle Fatigue Monitor 106
Table 4.8: Summary of MFCV tracking ASIC.

Technology 0.35 µm 2P4M
Circuit Size 0.954 mm2
Voltage Supply 3.3 V
Current Consumption Total: 628 µA
Bit-Stream Cross-Correlator: 290 nA
Biasing Circuits: 11.5µA
IA: 180µA
Low-Pass Filter: 60µA
Bit-Stream Converter: 68µA
CMRR 85.24 dB
Gain Adjustable (set to 111)
Cut-o frequencies 10.6 Hz, 2.5 kHz internal
Input Referred Noise Density nV
135 √
Hz
Correlation window Adjustable by CLK2
Estimation Accuracy Dependable on CLK1
MFCV dynamic range Dependable on IED & CLK1
Number of Delays 40
MARD Error 3.2 %
well as area, without the need for EMG compression algorithms [54, 55]. Moreover, wearable devices
benet from integrated, low power solutions, allowing portability and prolonged use. The resulting
small error of the proposed system supports that the system is capable of accurately estimating
MFCV during static contractions.
4.6 Wearable Muscle Fatigue Monitor
The MFCV tracking ASIC was integrated into a miniaturised PCB bearing a microcontroller in an
eort to design a complete monitoring platform for muscle fatigue. The board featured BLE com-
munication and was controlled by a custom designed Android application. Following that, a custom
3D printed case was developed to hold the electronics and interface them with the electrodes. This
eort manifested to the creation of the rst wearable device dedicated to muscle fatigue tracking .
2
The wearable node was designed to interface with the commercially available, low cost AMBU
Neuroline (72000-S/25) self adhesive, gel electrodes. The chosen electrodes are widely available and
since they are self adhesive, they would be responsible for attaching and holding the wearable device
in place. First, the four electrodes (3 for sensing, 1 for biasing) were arranged in a pattern that
minimises their footprint. A rough sketch of the electrode area and position on paper is shown in
2 The wearable device PCB and the control application were designed by Jack Heaey under the direct supervision
of Ermis Koutsos.
Optimu s
Placeholder
62.00 mm
nR F5182 2
DAC
40.00 mm EEP ROM
(a) (b) (c) (d)
Figure 4.22: (a) Minimum area for 4 electrode conguration. (b) PCB dimensions constrains based
on 4 electrode conguration. (c) Bottom side of the case with electrode snap sockets. (d) Completed
PCB with soldered components.
Fig. 4.22a. This area would dene the maximum dimensions for the wearable device, as seen in
Fig. 4.22b. In order for the electrodes to be interchangeable, an electrode attaching mechanism
with snap sockets was designed and 3D printed, shown in Fig. 4.22c. The distance between each
electrode is 2 cm. As the ASIC requires one electrode to be the furthest from the innervation
zone, the wearable contains a slide switch that swaps the internal connections of the two furthest
electrodes. Consequently, the wearable can be used in any orientation on the human body, allowing
it to be worn in a more comfortable position. The developed two-sided PCB would hold the following
components; the MFCV tracking ASIC, power managements circuits (regulator, battery charger),
USB UART communication controller, on-board Electrically Erasable Programmable Read-Only
Memory (EEPROM), a microcontroller with Bluetooth Low Energy (BLE) capabilities, a Digital
to Analogue Converter (DAC), signal buers and diagnostics ribbon connectors. The miniatured
PCB with the components soldered on top can be seen in Fig. 4.22d.
The wearable contains a two-layer PCB with a gold nish, as the gold wires are used to
bond the ASIC to the PCB. The Muscle Fatigue Monitor wearable utilises a Nordic Semiconductor
nRF51822 2.4 GHz Bluetooth Low Energy System-on-Chip (SoC), which oers 256 KB of ash
memory and 16 KB of RAM in a 6 by 6 mm QFN-48 package. To sample the sEMG signal, the
10-bit ADC of the nRF51822 is used with a 1.2 V reference. To set the comparator threshold voltage
and the common mode voltage, the nRF51822 uses its SPI capabilities to write to the Microchip
MCP4822 dual DAC. The wearable is powered by a 3.7 V Lithium-Polymer rechargeable battery,
with a capacity of 300 mAh. With the battery weighing just 7 g, the wearable is light-weight and
does not hinder movement. The wearable is recharged via a USB Mini Type B connector. As all
Figure 4.23: Exploded view of the case with the PCB inside.Snap sockets are screwed in place in
the hole oppenings. The device is attached on the user by the electrode adhesive.
of the components on the Muscle Fatigue Monitor wearable PCB require a 3.3 V supply, the 3.7
V supply from the battery is steppeddown to 3.3 V using a Diodes Incorporated AP7333-33SAG-
7 voltage regulator. If streaming the sEMG signal to the Android application continuously, the
wearable has a battery life of 27 h. The nRF51822 also communicates with a Microchip 25LC1024
EEPROM via SPI. The EEPROM has a memory capacity of 128 KB in a Small Outline Integrated
Circuit (SOIC) package. If the wearable is not connected to Android and collecting delay values
from the ASIC and storing them in the EEPROM, the wearable has a battery life of 64 h. If the
correlation window has a frequency of 1 Hz, one delay value will be obtained from the ASIC per
second. Consequently, the 128 KB EEPROM would be able to store 36 hours worth of data. The
completed PCB inside its 3D printed case is shown in Fig. 4.23.
The Android application provides the user interface to the wearable. When the application
acquires data from the wearable, it stores the data in a dierent directory for each registered user.
After connecting to the wearable, the Android application allows the user to set the comparator
threshold voltage and the common mode voltage. The user can also change the sampling frequency
and correlation time window. The calibration tab is shown in Fig. 4.24a. After setting the threshold
voltage or during an auto-calibration, the output of the comparator at the end of the delay path is
streamed to the Android application. The Android application can stream both the sampled sEMG
and the MFCV values from the ASIC. Alternatively, it can stream only the MFCV values if the user
wishes to save power. The sEMG and delay values are plotted on graphs, as shown in Fig. 4.24b.
After a live stream is halted by the user, the received sEMG and/or MFCV values are saved to
a CSV le. By pressing the button at the bottom right of Fig. 4.24b, the delay values that have
been written to the EEPROM since the last time this button was pressed are sent to the Android
4.7 Summary 109
(a) (b)
Figure 4.24: (a) Calibrate parameters tab of the Android application. The used can specify the
common mode voltage, threshold voltage, sampling frequency and correlation window, (b) Live cap-
ture tab of the Android application. The user can visualise sEMG and muscle fatigue data, as well
as EEPROM saved data and choose to export the recorded values.
(a) (b)
Figure 4.25: (a) Wearable Muslce Fatigue Tracking System, showning the Android application and
the wearable device, (b) Complete Muslce Fatigue Tracking System, showning the Android application
and the wearable device attached on a user.
application and saved to a CSV le. Instead of using the Android application to capture sEMG
and fatigue data, the wearable can stream data directly to MATLAB
® via USB. The wearable is
light and unobtrusive (Fig. 4.25a) and it can be seen attached on a user in Fig. 4.25b. Table 4.9
summarises the specications of the developed muscle fatigue tracking wearable device.
4.7 Summary
This chapter presented a rst of its kind MFCV tracking ASIC and its detailed implementation and
operation. The MFCV ASIC utilises a novel bit stream approach that greatly simplies the sEMG
signal without any loss of information, thus reducing computational complexity and minimising
4.7 Summary 110
Table 4.9: Summary of Muscle Fatigue Tracking Wearable Device.

Power consumption during EMG/fatigue stream 41.48 mW
Battery life for continuous EMG/fatigue streaming 27 h
Power consumption with BLE continuously inactive 17.43 mW
Battery life with BLE inactive 64 h
10-bit ADC LSB 1.172 mV
EEPROM fatigue data capacity 130816 values
power consumption. The developed ASIC is capable of tracking MFCV from surface EMG signals
using the method of cross-correlation. The system consists of a sEMG instrumentation amplier, a
ltering stage, a bit-stream converter and a bit-stream cross-correlator. A test group of 20 people
was used to demonstrate the ability of the proposed bit-stream correlator to accurately estimate
MFCV in real time during static fatiguing contractions. The ASIC has a MARD error of 3.2 %
compared to MATLAB
® analysis for the same dataset. The system draws 628 µA from a 3.3 V
power supply and is implemented in a commercially available 0.35 µm CMOS technology. Hence
this is the rst working and validated System-on-Chip for muscle fatigue monitoring.
The ASIC was embedded in a Muscle Fatigue Monitoring wearable device capable of stream-
ing live sEMG and muscle fatigue data wirelessly to a user controlled smart-phone application using
low power Bluetooth technology. The wearable node is lightweight and attached to the skin through
the self adhesive, pre-gelled, commercial surface electrodes. When the node is not transmitting the
two sEMG channels, the operational time is not limited by the battery life but by the onboard
memory.
REFERENCES 111
References
no. 1, pp. 1836, 2000.
[2] T. Masuda, H. Miyano, and T. Sadoyama, The measurement of muscle ber conduction ve-
locity using a gradient threshold zero-crossing method, Biomedical Engineering, IEEE Trans-
actions on, no. 10, pp. 673678, 1982.
2001.
[4] K. Masuda, T. Masuda, T. Sadoyama, M. Inaki, and S. Katsuta, Changes in surface emg
parameters during static and dynamic fatiguing contractions, Journal of Electromyography

[5] D. Farina and R. Merletti, Methods for estimating muscle bre conduction velocity from
surface electromyographic signals, Medical and biological Engineering and Computing, vol. 42,
no. 4, pp. 432445, 2004.
[6] E. Koutsos and P. Georgiou, An analogue instantaneous median frequency tracker for emg
fatigue monitoring, in Circuits and Systems (ISCAS), 2014 IEEE International Symposium
on, pp. 13881391, IEEE, 2014.
[7] M. Stana¢evi¢, S. Li, and G. Cauwenberghs, Micropower mixed-signal vlsi independent com-
ponent analysis for gradient ow acoustic source separation, IEEE Transactions on Circuits
and Systems I: Regular Papers, vol. 63, no. 7, pp. 972981, 2016.
[8] P. Julián, A. G. Andreou, and D. H. Goldberg, A low-power correlation-derivative cmos vlsi
circuit for bearing estimation, IEEE Transactions on Very Large Scale Integration (VLSI)
Systems, vol. 14, no. 2, pp. 207212, 2006.
[9] P. Julián, A. G. Andreou, L. Riddle, S. Shamma, D. H. Goldberg, and G. Cauwenberghs, A
comparative study of sound localization algorithms for energy aware sensor network nodes,
IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 51, no. 4, pp. 640648,
2004.
REFERENCES 112
[10] M. M. Lowery, C. L. Vaughan, P. J. Nolan, and M. J. O'Malley, Spectral compression of the
electromyographic signal due to decreasing muscle ber conduction velocity, Rehabilitation

[11] G. Dimitrov, Z. Lateva, and N. Dimitrova, Eects of changes in asymmetry, duration and
propagation velocity of the intracellular potential on the power spectrum of extracellular po-
tentials produced by an excitable ber., Electromyography and clinical neurophysiology, vol. 28,
no. 2-3, p. 93, 1988.
[12] N. Dimitrova and G. Dimitrov, Interpretation of emg changes with fatigue: facts, pitfalls, and
fallacies, Journal of Electromyography and Kinesiology, vol. 13, no. 1, pp. 1336, 2003.
[13] L. Arendt-Nielsen and K. Mills, The relationship between mean power frequency of the emg
spectrum and muscle bre conduction velocity, Electroencephalography and clinical Neurophys-
iology, vol. 60, no. 2, pp. 130134, 1985.
[14] R. Merletti, M. Knaitz, and C. J. De Luca, Myoelectric manifestations of fatigue in voluntary
and electrically elicited contractions, J Appl Physiol, vol. 69, no. 5, pp. 181020, 1990.
[15] W. H. Linssen, D. F. Stegeman, E. M. Joosten, S. L. Notermans, M. A. van't Hof, and R. A.
Binkhorst, Variability and interrelationships of surface emg parameters during local muscle
fatigue, Muscle & nerve, vol. 16, no. 8, pp. 849856, 1993.
[16] F. B. Stulen and C. J. De Luca, Frequency parameters of the myoelectric signal as a measure
of muscle conduction velocity, Biomedical Engineering, IEEE Transactions on, no. 7, pp. 515
523, 1981.
[17] M. Zwarts, T. Van Weerden, and H. Haenen, Relationship between average muscle bre con-
duction velocity and emg power spectra during isometric contraction, recovery and applied
ischemia, European journal of applied physiology and occupational physiology, vol. 56, no. 2,
pp. 212216, 1987.
[18] G. McVicar and P. Parker, Spectrum dip estimator of nerve conduction velocity, Biomedical
[19] E. Spinelli, C. Felice, M. Mayosky, J. Politti, and M. Valentinuzzi, Propagation velocity mea-
surement: autocorrelation technique applied to the electromyogram, Medical and Biological

REFERENCES 113
[20] D. Farina, M. Pozzo, E. Merlo, A. Bottin, and R. Merletti, Assessment of average muscle
ber conduction velocity from surface emg signals during fatiguing dynamic contractions,
Biomedical Engineering, IEEE Transactions on, vol. 51, no. 8, pp. 13831393, 2004.
[21] I. Hunter, R. Kearney, and L. Jones, Estimation of the conduction velocity of muscle ac-
tion potentials using phase and impulse response function techniques, Medical and Biological
[22] D. Farina and R. Merletti, Comparison of algorithms for estimation of emg variables during
voluntary isometric contractions, Journal of Electromyography and Kinesiology, vol. 10, no. 5,
pp. 337349, 2000.
[23] D. Farina and R. Merletti, A novel approach for estimating muscle ber conduction veloc-
ity by spatial and temporal ltering of surface emg signals, Biomedical Engineering, IEEE
[24] D. Farina and D. Falla, Estimation of muscle ber conduction velocity from two-dimensional
surface emg recordings in dynamic tasks, Biomedical Signal Processing and Control, vol. 3,
no. 2, pp. 138144, 2008.
[25] P. Parker and R. Scott, Statistics of the myoelectric signal from monopolar and bipolar elec-
trodes, Medical and biological engineering, vol. 11, no. 5, pp. 591596, 1973.
[26] M. Schwartz, ed., EMG Methods for Evaluating Muscle and Nerve Function, ch. 2, 9, 22.
InTech, 2012.
[27] G. Kamen and D. Gabriel, Essentials of electromyography. Human Kinetics, 2010.
[28] M. Barbero, R. Merletti, and A. Rainoldi, Atlas of Muscle Innervation Zones. Springer, 2011.
[29] R. Merletti and L. L. Conte, Advances in processing of surface myoelectric signals: Part 1,
Medical and Biological Engineering and Computing, vol. 33, no. 3, pp. 362372, 1995.
[30] H. Broman, G. Bilotto, and C. J. De Luca, A note on the noninvasive estimation of muscle
ber conduction velocity, IEEE transactions on biomedical engineering, vol. 5, no. BME-32,
pp. 341344, 1985.
[31] R. Merletti and L. R. L. Conte, Surface emg signal processing during isometric contractions,

REFERENCES 114
[32] anatomy diagram, Arm muscles, 2015.
[33] T. S. Lande, T. G. Constandinou, A. Burdett, and C. Toumazou, Running cross-correlation
using bitstream processing, Electronics letters, vol. 43, no. 22, pp. 11811183, 2007.
[34] T.-D. Chiueh, P.-Y. Tsai, and I.-W. Lai, Baseband Receiver Design for Wireless MIMO-OFDM
Communications. John Wiley & Sons, 2012.
[35] T. Sadoyama, T. Masuda, and H. Miyano, Relationships between muscle bre conduction
velocity and frequency parameters of surface emg during sustained contraction, European
Journal of Applied Physiology and Occupational Physiology, vol. 51, no. 2, pp. 247256, 1983.
[36] S. Andreassen and L. Arendt-Nielsen, Muscle bre conduction velocity in motor units of the
human anterior tibial muscle: a new size principle parameter., The Journal of Physiology,
vol. 391, no. 1, pp. 561571, 1987.
[37] T. Sadoyama, T. Masuda, H. Miyata, and S. Katsuta, Fibre conduction velocity and bre
composition in human vastus lateralis, European journal of applied physiology and occupational
physiology, vol. 57, no. 6, pp. 767771, 1988.
[38] X. Ye, T. Beck, and N. Wages, Relationship between innervation zone width and mean muscle
ber conduction velocity during a sustained isometric contraction, Journal of musculoskeletal

& neuronal interactions, vol. 15, no. 1, pp. 95102, 2015.
[39] M. Naeije, Estimation of the action potential conduction velocity in human skeletal muscle
using the surface emg cross-correlation technique, Electromyogr Clin Neurophysiol, vol. 23,
pp. 7380, 1983.
[40] S. Henzler, Time-to-digital converter basics, in Time-to-Digital Converters, pp. 518,
Springer, 2010.
[41] R. Pallas-Areny, Interference-rejection characteristics of biopotential ampliers: a comparative
analysis, Biomedical Engineering, IEEE Transactions on, vol. 35, no. 11, pp. 953959, 1988.
[42] J. H. Nagel, Biopotential ampliers, The Biomedical Engeneering Handbook, pp. 11851195,
1995.
[43] M. J. Burke and D. T. Gleeson, A micropower dry-electrode ecg preamplier, Biomedical

REFERENCES 115
[44] E. M. Spinelli, N. H. Martinez, and M. A. Mayosky, A single supply biopotential amplier,
Medical engineering & physics, vol. 23, no. 3, pp. 235238, 2001.
[45] E. M. Spinelli, N. Martínez, M. A. Mayosky, and R. Pallàs-Areny, A novel fully dierential
biopotential amplier with dc suppression, Biomedical Engineering, IEEE Transactions on,

vol. 51, no. 8, pp. 14441448, 2004.
[46] E. M. Spinelli, R. Pallàs-Areny, and M. A. Mayosky, Ac-coupled front-end for biopotential
measurements, Biomedical Engineering, IEEE Transactions on, vol. 50, no. 3, pp. 391395,
2003.
[47] R. F. Yazicioglu, S. Kim, T. Torfs, H. Kim, and C. Van Hoof, A 30 w analog signal processor
asic for portable biopotential signal monitoring, Solid-State Circuits, IEEE Journal of, vol. 46,
no. 1, pp. 209223, 2011.
[48] N. Verma, A. Shoeb, J. Bohorquez, J. Dawson, J. Guttag, and A. P. Chandrakasan, A micro-
power eeg acquisition soc with integrated feature extraction processor for a chronic seizure
detection system, Solid-State Circuits, IEEE Journal of, vol. 45, no. 4, pp. 804816, 2010.
applications, Solid-State Circuits, IEEE Journal of, vol. 38, no. 6, pp. 958965, 2003.
[50] R. Yazicioglu, P. Merken, R. Puers, and C. Van Hoof, A 60 uw 60 nv/ hz readout front-end
for portable biopotential acquisition systems, Solid-State Circuits, IEEE Journal of, vol. 42,
no. 5, pp. 11001110, 2007.
[51] R. F. Yazicioglu, P. Merken, R. Puers, and C. Van Hoof, A 200 w eight-channel eeg acquisi-
tion asic for ambulatory eeg systems, Solid-State Circuits, IEEE Journal of, vol. 43, no. 12,
pp. 30253038, 2008.
[52] T. Denison, K. Consoer, W. Santa, A.-T. Avestruz, J. Cooley, and A. Kelly, A 2 uw 100
nv/rthz chopper-stabilized instrumentation amplier for chronic measurement of neural eld
potentials, Solid-State Circuits, IEEE Journal of, vol. 42, no. 12, pp. 29342945, 2007.
[53] C. J. De Luca, L. Donald Gilmore, M. Kuznetsov, and S. H. Roy, Filtering the surface emg
signal: Movement artifact and line noise contamination, Journal of biomechanics, vol. 43,
no. 8, pp. 15731579, 2010.

REFERENCES 116
[54] A. M. Dixon, E. G. Allstot, D. Gangopadhyay, and D. J. Allstot, Compressed sensing system
considerations for ecg and emg wireless biosensors, Biomedical Circuits and Systems, IEEE
[55] J. Zhang, Y. Suo, S. Mitra, S. P. Chin, S. Hsiao, R. F. Yazicioglu, T. D. Tran, and R. Etienne-
Cummings, An ecient and compact compressed sensing microsystem for implantable neural
recordings, Biomedical Circuits and Systems, IEEE Transactions on, vol. 8, no. 4, pp. 485496,
2014.
117
Chapter 5
A MFM ASIC exploring the iMDF and

Zero Crossings relationship
5.1 Introduction
T
HERE is still a lot to learn about the physiological changes during fatigue and how they
are conveyed though the surface EMG (sEMG) signal. As shown in Chapter 2, during
fatigue the conduction velocity of the Motor Unit Action Potentials (MUAPs) changes. It appears
as if the MUAPs slow down", hence the sEMG spikes become broader. In addition, the ring
rate decreases and fast twitch bres give up while slow twitch bres take over. However, the exact
mechanisms behind these changes are not fully understood yet. It has been reported in the literature
that the Zero Crossing Rate (ZCR), mean spike duration and spike counting are closely related to
instantaneous Median Frequency (iMDF) and correlate well with muscle fatigue [17]. Moreover
the sEMG-force relationship as well as the Inter Spike Interval (ISI) [8, 9] can attribute to changes
in the MU ring rate.
All these changes are time based, yet they collectively manifest themselves as a frequency
shift to lower values and the iMDF is the best metric to track that spectral compression. Tracking
that compression is achieved by looking into the frequency content of the signal either using Short
Time Fourier Transform (STFT) or with the use of systems demonstrated in Chapter 3. In Chap-
ter 4, the use of bit-stream processing has proven benecial in terms of computational complexity
and robustness, and showed potential in simplifying the sEMG signal without loss of information.
Therefore, this approach can potentially combine the advantages of bit-stream processing with the
unique insight of iMDF in fatigue analysis.
This chapter exploits time domain parameters to approximate the iMDF metric. As stated
5.2 Median Frequency and sEMG Spike Duration 118
in the literature (Chapter 2), the sEMG signal is composed by a stream of superimposed MUAPs.
Hence, it is possible to establish a link between iMDF and the time distribution of the sEMG
spikes. This chapter presents a novel realisation of an Application Specic Integrated Circuit (ASIC)
dedicated to conditioning the sEMG signal and extracting fatigue related information through a
dierent metric and mixed signal processing techniques.
5.2 Median Frequency and sEMG Spike Duration
So far, two closed loop systems were presented which track the median frequency of the sEMG
power spectrum by processing the recorded signal in the frequency domain (Chapter 3). Processing
the sEMG signal in the frequency domain increases the systems robustness to Signal-to-Noise Ratio
(SNR) and the estimated iMDF values reect the true median of the power spectrum. However,
processing in the frequency domain can have an impact on system complexity, design tolerances
and overall power consumption. These can be leveraged by the appropriate use of time domain
methods to approximate the spectral changes the sEMG signal undergoes under fatigue.
It is well established that there is a dependence of the EMG spectrum on the propagation
velocity of the MUAPs [1, 10]. The EMG power spectrum can be written in the form shown in
Eq. 5.1 bellow:

1 fd
Pf = G (5.1)
v2 v
where v is the propagation velocity of the MUAPs, f the frequency, Pf the recorded spectrum, d the

fd
surface electrode distance and G v includes the spectrum of the MUAPs and other geometrical
factors. Thus, the observed median frequency shift is proportional to the MUAPs velocity [1].
The sEMG amplitude appears random in nature as the random inter-pulse intervals and
amplitudes result from the superposition of many independent pulse trains of individual Motor Units
(MUs). Hence, the sEMG signal is aected by the ring rate of the MUs (0-20 Hz) [11]. The ring
rate mainly aects the power spectrum for frequencies up to 40-80 Hz. Above that point, other
factors including conduction velocity, muscle type/size, electrode distance and muscle chemistry
dene the spectrum shape and position [8, 9]. During fatigue, the ring rate of most motor units
decreases [12]. Fast twitch muscle bres dominate high force production but get easily fatigued and
drop-out. This leaves fatigue resistant, slow twitch muscle bres with smaller amplitudes, deeper
in the muscle and slower conduction velocities [5]. Furthermore, MUAP durations increase during
peripheral fatigue due to decreased muscle bre conduction velocities, while MUAP amplitudes
either increase or decrease [13]. The low-frequency range of the myoelectric power spectrum depends
on the inter-pulse interval and is inuenced by fatigue, as it increases with fatigue [1]. As a result,
MU ring rate and MUAP spike durations accurately reect the changes in the sEMG properties
and the compression of the sEMG power spectrum under fatigue.
For this reason, several metrics including the number of spikes,zero crossings, turns and peaks
per unit time have been used as indicators of muscle fatigue [5, 14]. Specically, Inbar et al [1]
showed that a simple computational measure, the ZCR, can be used as a reliable monitor of surface
EMG spectral changes. The ZCR or zero-crossing frequency fz of the sEMG signal is dened as
half the number of zero crossings per second [1, 7]. The relationship between ZCR and the signal
power spectrum was dened by Rice [15] as shown in Eq. 5.2 below:
"R f
s /2
#
0 f 2 S(f )df
Zf = 2 R fs /2 (5.2)
0 S(f )df
where S(f ) is the Power Spectral Density (PSD) of the signal and fs is the sampling frequency.
The ZCR showed properties close to instantaneous Mean Frequency (iMNF) of the sEMG power
spectrum [7] and Inbar et al [1] provided a relationship between ZCR and iMNF, shown in Eq. 5.3,
where σ2 is the variance of the power spectrum and Fa the mean frequency. As expected, time
window length determined the variance and sensitivity of the ZCR to spectral variations.
Zf = 2(σ 2 + Fa2 )1/2 (5.3)
As seen from Eq. 5.3, the ZCR can reect changes in mean frequency. Similarly, the distribu-
tion of the time distance between crossings can oer a unique insight in the velocity of the MUAPs,
the ring rate and hence the progressive compression of the power spectrum during fatigue. Since
the sEMG signal is a superposition of several MUAPs and closely related to their velocity, then
the spike duration will be closely linked to the properties of the sEMG signal under fatigue and its
power spectrum. Examining the time duration properties of the sEMG crossings can oer a more
detailed look into sEMG properties. The aim is to get a representative and accurate reading and
insight of the spectral behaviour of the sEMG signal and the median frequency better reects the
spectral changes of the sEMG signal during fatigue [12, 1618].

0.4
0.3
0.2
0.1
sEMG Amplitude (V)

Z1 Z2 Z3 Z4 Z5 Z6
T2 T4
0
T1 T3 T5
-0.1
-0.2
-0.3
0 50 100
Time (samples)
Figure 5.1: sEMG segment indicating the zero crossings and the time interval between these
crossigns.
The Mean Spike Length (MSP) has been used as an estimator of fatigue in the past [19].
However, MSP is calculated from the number of crossings in a given window and is based on the
assumption that all crossing durations are equally spaced with a linear distribution among them.
Looking into the actual distribution of the time distance between the individual crossings [5] and
estimating their median value will yield a metric of fatigue comparable to the true median frequency
of the myoelectric power spectrum, as well as provide a detailed insight to the sEMG physical
changes during fatigue. Fig. 5.1 shows the sEMG zero crossings and the time distance between
these crossings used for sEMG fatigue analysis. This new metric will be dened as Spike Duration
based Median Frequency approximation (SDMF, in Hz) and will be dened as the reciprocal of
1
twice the median length of the crossings, SDM F = 2τ̃ .
The SDMF from three biceps brachii sEMG recordings of three dierent subjects was com-
puted using MATLAB

® and compared to the iMDF. The SDMF and iMDF plots can be seen in
Fig. 5.2 and the corresponding R, R2 and RM S values in Table 5.1. It is easily observed that
the changes in the power spectrum expressed by iMDF are reected with great detail in SDMF.
Although the absolute frequencies do not match, there is a strong relationship between the two
frequency trends. This observation is further supported by the computed R and R2 values between
iMDF and SDMF, where the correlation between them is above 90%.
The recording of the sEMG signal is very susceptible to the movement of the electrodes,
the recording device or the electrode leads. Residual artefacts such as motion artefacts, spikes and
180 75 110
MDF
160 SDMF
70 100
Smooth MDF
Smooth SDMF
140
65 90
120
60 80
Frequency (Hz)
100
55 70
80
50 60
60
45 50
40
20 40 40
0 35 30
0 20 40 60 80 100 120 140 0 20 40 60 80 100 0 10 20 30 40 50 60 70
Time (s) Time (s) Time (s)
(a) (b) (c)

Figure 5.2: SDMF and iMDF plots of biceps brachii sEMG recordings of three dierent subjects
(a), (b), (c) respectively).
Table 5.1: Correlation coecient values between the original and smoothed iMDF and
SDMF.
Subject 1 Subject 2 Subject 3
Direct Smooth Direct Smooth Direct Smooth
R 0.95 0.99 0.97 0.99 0.94 0.99
R2 0.90 0.98 0.93 0.99 0.88 0.99
RMS 5.77 1.38 2.42 0.53 3.82 1
other sources of low freq noise have to be identied and removed prior to processing the recording.
A test signal was constructed by inserting a 10 Hz sine wave segment at regular intervals in a
unaltered (clean") and ltered sEMG signal. The resulting test signal is shown in Fig. 5.3a. The
test signal resembles an extreme case of an sEMG recoding with the presence of motion artefacts.
It was constructed for the purpose of evaluating SDMF susceptibility to artefacts. Estimation of
the iMDF of the altered sEMG using normal frequency analysis tools (such as the computation of
the periodogram) without prior removal of the artefacts would result to a biased estimate that does
not reect the spectral content of the original sEMG signal. This can be observed in Fig. 5.3b.
The presence of low frequency artefacts can signicantly alter measurements of mean values and
concentrations, such as ZCR, mean spike length and peak and spike counting. The median of a set
of data is dened as the number halfway into the set after the data have been arranged in order
from least to greatest. Thus, the median is less sensitive to outlier values and can provide a more
reliable and accurate approximation of the power spectrum even in extreme recordings than other
metrics computing mean concentrations. This can be observed from Fig. 5.4 and Table 5.2, where
the same altered sEMG signal was used to estimate the SDMF and compared to the iMDF of the
original, unaltered, sEMG signal. It is clear that the SDMF can tolerate the presence of outlier
0.15
0.15 8080
sMEG
sMEG
7070 sEMG
sEMG
& artefacts
& artefacts
0.1
0.1
6060
0.05
0.05
sEMG Amplitude (V)
sEMG Amplitude (V)
5050
iMDF (Hz)
iMDF (Hz)
00 4040
3030
-0.05
-0.05
2020
-0.1
-0.1
1010
-0.15
-0.15 00
00 2000
2000 4000
4000 6000
6000 8000
8000 10000
1000012000
1200014000
1400016000
16000 00 2020 4040 6060 8080 100
100 120
120 140
140 160
160 180
180
Time
Time
(samples)
(samples) Time
Time
(samples)
(samples)
(a) (b)
Figure 5.3: (a) sEMG segment showing the addition of low frequency (10 Hz) motion artefacts, (b)
iMDF computed for the unaltered and the altered sEMG signal.
Table 5.2: Correlation coecient values between the original and smoothed iMDF and
SDMF with the presence of simulated motion artefacts.
Subject 1 Subject 2 Subject 3
Direct Smooth Direct Smooth Direct Smooth
R 0.89 0.98 0.90 0.99 0.76 0.97
R2 0.79 0.97 0.83 0.98 0.59 0.95
RMS 8.58 2.88 3.98 1.17 7.32 2.46
values and provide a representable and accurate reading, as a strong correlation with iMDF was
observed with values higher than 75%. The advantages of the SDMF metric are presented bellow:
Bit-stream based, mixed signal (robust to process variations) processing,
Closely reects the slowing down of MUAPs,
Not aected by high frequency noise,
Increased robustness to motion artefacts,
Robust to outlier values,
Reduced complexity over iMDF methods,
Potentially lower power than iMDF methods.

180 75 110
MDF
160 SDMF
70 100
Smooth MDF
Smooth SDMF
140
65 90
120
60 80
Frequency (Hz)
100
55 70
80
50 60
60
45 50
40
20 40 40
0 35 30
0 20 40 60 80 100 120 140 0 20 40 60 80 100 0 10 20 30 40 50 60 70
Time (s) Time (s) Time (s)
(a) (b) (c)

Figure 5.4: SDMF and iMDF plots of biceps brachii sEMG recordings of three dierent subjects
(a), b), (c) respectively). The SDMF was computed using the artefact infected sEMG signal and the
iMDF using the clean sEMG signal.
5.2.1 Algorithm for Spike Duration Based iMDF Approximation
In order to estimate the distribution and the median value of the time duration between zero
crossings, as seen in Fig. 5.1, the sEMG signal is converted to a series of pulses (bit-stream) with
the aid of a Pulse Width Modulator (PWM). In the presented application, the PWM is a simple
comparator, with the threshold set to the common mode of the amplied signal. The resulting
positive (Fig. 5.1, red periods) and negative (Fig. 5.1, green periods) pulses (toggles) are processed
separately. With the aid of a Pulse Length Detector (PLD), each toggle's length can be estimated.
The PLD calculates the toggle length concentration that is equal to several predened time constants
τ. The aim is to nd the required τ such as the number of toggles with length T >τ is equal to
the number of toggles with length T < τ. This is accomplished by comparing the total numbers
of toggles with the number of toggles with periods T >= τ . Consider Fiq. 5.5 for an architectural
block diagram illustration of the proposed algorithm. By progressively counting toggles with longer
length, starting from a small τ and increasing it, the distribution of toggle lengths is approximated.
For any probability distribution, the medial is by denition any real number k that P (X ≤ k) =
Rk 1
P (X ≥ k) = −∞ f (x)dx = 2 . Hence, the relationship between the counters and τ̃ is given by
τ̃ = min( Counter
2
1
− Counter2 ).
The PLD consists of a series of delay and overlap stages. Each stage progressively delays
the input bit-stream and compares the delayed version with the original undelayed one. Consider
Fiq. 5.6, where (a) is the undelayed positive toggle bit-stream, (b) the undelayed negative toggle
bit-stream, (c) the already delayed positive toggle bit-stream and (d) the comparison between the
two versions of the same bit-stream. If the two signals overlap, then the toggle length is greater or
possitive PLDup
τ1 τ2 τ3 ... τn
sEMG PWM
negative PLDdown
τ1 τ2 τ3 ... τn
Counter 2
_
Counter 1 +1/2 ~τ Output τ
Figure 5.5: Architectural block diagram illustration of the proposed SDMF tracking algorithml.
Stages
equal to the 3-39 delay at that stage and the comparison output increments a counter. Hence,
added
{
the counter values from all the stages represent the concentration of toggles with length larger than
Delay & Delay & Delay &
overlap overlap overlap
stage 1 the equivalent
stage 1 stage delay.
stage 1 If the two pulses do not overlap, it means that the toggle length is
longer that the current delay stage and it will increment one of the subsequent counters. Thus,
the expected distribution of counted toggles by each counter (from all stages) is a monotonically
Delay & Delay & Delay &
overlap overlap overlap
decreasing line. Furthermore, the starting value for each counter is expected to decrease as the
stage 1 stage 2 stage 40
) stage is incremented.
adder adder adder

The proposed SDMF circuit implementation is presented in Fiq. 5.7. The sEMG signal
Ca1 Ca2 Ca40

enters the PWM block, which produces the toggle pulses for spike period as shown in Fiq. 5.6a,b.
Output ~τ = max | ( Ct/2 - Ca ) |

The delay and overlap stage is presented in more detail in Fiq. 5.8. An 8 bit ripple counter was
1
empirically selected to meet the requirements of the system. The introduced delay is equal to
CLK .
In case there is a wrong overlap pulse (Fig. 5.6d), it is corrected using an Overlap Correction
Delay & overlap stage
circuit, comprising of two D-type ip ops (DFFs) and a xnor gate. A starting delay is introduced
to adjust the dynamic range of the system and eliminate very short toggles due to high frequency
D Q
noise. The aim of the system is to nd the required delay τ such as the number of toggles with
x(n) CLK
length above that delay is equal to the number of toggles bellow that delay. The counters are
evaluated regularly at a frequency CLK2 . Hence, the total number of toggles and the ones that
D Q
are longer than τ are compared and the as in Eq. 5.4. The min function is designed in VHDL
in a similar manner to the max function usedxd(n) D Q CLK

in the Muscle Fibre Conduction Velocity (MFCV)
CLK
estimator in Chapter 3, section 4.3.2) and is shown in Appendix
CLK B. The two dierences between
the two functions are the minimum operation change and the addition of a subtraction stage to the
INC
new function. RST RST RST
counter
8
Out
PWM sEMG signal

Toggle length
Positive
a) toggle
Negative
toggle
b)
Toggle length Delayed

c) >=
Delay
positive
toggle
sEMG PWM
τ1 τ
Wrong Overlap
d)
Overlap
Overlap
AND Output
to Counter C
Time
Counter 1
Figure 5.6: (a) is the undelayed positive toggle bit-stream, (b) the undelayed negative toggle bit-
stream, (c) the already delayed positive toggle bit-stream and (d) the comparison between the two
versions of the same bit-stream. If the two signals overlap, then the toggle length is greater or equal
to the added delay at that stage and the comparison output increments a counter.
Stages 3-39
{
xup(n) Delay & Delay & Delay &

PWM
Start overlap overlap overlap
delay stage 1 stage 2 stage 40
xdup(n)
xdown(n) Delay & Delay & Delay &

Start overlap overlap overlap
delay stage 1 stage 2 stage 40
xddown(n)
adder adder adder adder
Ct Ca1 Ca2 Ca40
Output ~τ = min | ( Ct/2 - Ca ) |
Figure 5.7: SDMF tracking system circuit implementation.

dder adder
Ca2 Ca40
)| 5.2 Median Frequency and sEMG Spike Duration 126
Delay & overlap stage
D Q Overlap
Correction
x(n) CLK
D Q
xd(n) D Q CLK
CLK CLK
INC
RST RST RST
counter
8
Out
Figure 5.8: Circuit implementation of the delay & overlap stage, including an overlap correction circuit.

Ct
Dn,i = min − (Ca1 , Ca2 ...Can ) (5.4)
2 t=CLK2
The SDMF is computed using Eq. 5.5, where P LDτ# is the estimated delay at which the
number toggles with length above that are equal to the number the ones with length bellow that.
Stages 3-39
e2 The Start Delay for Correlation
P LD Stage
T # is set to 40
an empirical value of 5 (1 unit delay = 1 CLK cycle).
{
A total of 40 stages were implemented to allow for a wide dynamic range of computation. As
mentioned in the analysis above, the expected distribution of counted toggles by each counter
D Q
should be a monotonically decreasing line. Consider Fig. 5.9a, where the counter value from all
CLK
stages is shown at three distinct points in time. All three plots show a decreasing trend with time.
It was observed that some Dof the

Q smaller length toggles, if closely spaced, are not removed by the
Overlap Correction
D Q circuit.
CLK This has the eect of increasing the counted toggles in the longer
delays (stage # ≥ 30). However, this does not impact the estimated values signicantly, as it only
CLK
occurs at the start of the contraction where the toggle lengths are shorter and is only observed
INC INC
RST RST in long delay stages. RSTeach RST
The starting value for counter is expected to decrease as the stage is
counter counter
8 8
incremented. This can be observed in Fig. 5.9b, where the counter values for three distinct stages
Out (2) Out (40)
are plotted for the complete duration of the sEMG signal of a fatiguing contraction. The SDMF
for the same signal as in Fig. 5.2c is computed using the median() function in MATLAB
® and
D two
the proposed algorithm. The Q estimates for SDMF are shown in Fig. 5.10. It is easily observed
CLK
D Q
5.3 SDMF Tracking ASIC Architecture 127
240
240 300
300
220
220 pulse
pulse # at
# at 10s
10s pulse
pulse # at
# at scale
scale 1 1
pulse
pulse # at
# at 40s
40s 250
250 pulse
pulse # at
# at scale
scale 2020
200
200 pulse
pulse # at
# at 60s
60s pulse
pulse # at
# at scale
scale 3535
180
180
Number of pulses (#)

200
200
160
160
140
140 150
150
120
120
100
100
100
100
8080
5050
6060
4040 0 0
0 0 5 5 1010 1515 2020 2525 3030 3535 4040 0 0 1010 2020 3030 4040 5050 6060 7070
Delay
Delay number
number (#)(#) Time
Time (s)(s)
(a) (b)
Figure 5.9: (a) the counter value from all stages at three distinct points in time. All three plots
show a decreasing trend with time. (b) the counter values for three distinct stages for the complete
duration of the sEMG signal of a fatiguing contraction.
that the proposed algorithm successfully estimates the median crossing length. The Mean Absolute
Relative Dierence (MARD) between the two plots is less than 6.5% and can be attributed to the
non-monotonic counting behaviour of some stages.
CLK
SDM F (Hz) = (5.5)
2(P LDT # + Start_Delay)
5.3 SDMF Tracking ASIC Architecture
The developed SDMF ASIC (code name Batman) shown in Fig. 5.11 consists of three major building
blocks: a) a sEMG Instrumentation Amplier (IA) with a Programmable Gain Amplier (PGA) and
a second order Low Pass Filter (LPF), b) a PWM bit-stream converter comprised by an analogue
comparator and d) a circuit implementation of the proposed SDMF tracking algorithm. The signal
gets amplied and band-pass ltered by the IA, PGA and LPF. Following that, it is converted
to a series of pulses (toggles) by the comparator and subsequently gets processed by the SDMF
algorithm. In addition, the SDMF ASIC includes a bias generator circuit, a digital timing control
circuit, a 10-bit Analogue to Digital Converter (ADC) and a Serial Peripheral Interface (SPI) to
communicate with any receiving circuit.

160
MDF
SDMF chip
140 SMDF matlab
120
100
Frequency (Hz)
80
60
40
20
0 10 20 30 40 50 60 70
Time (s)
Figure 5.10: Estimated SDMF comparison between the median() function in MATLAB® and the
proposed algorithm. The error between
b)
them is given by MARD = 6.5%.
240 300
220 pulse # at 10s pulse # at scale 1

pulse # at 40s 250 pulse # at scale 20
200 pulse # at 60s pulse # at scale 35
180
200
160
140 150
Analog Out Bit-Stream Out

120
100
100 a) sEMG readout Stage b) Bit-Stream
El1
Delay & Delay &
overlap overlap
80 TI TC PGA
50 stage stage
Delay
6-bit
SPI
60 El2 x40
/
C1 x1 Delay & Delay &

Epidermis
DC Servo 1 overlap overlap

40 0
0 5 10 15 20 25 30 35 40 0 10 20 30 40 50 stage 60 70 stage
Delay
El3 number (#) Bias Generator Time (s) Counter Counter
Vref : Setting DC, Timing Control
close proximity Min
to electrodes / Enable Min
6-bit Selectror
10-bit ADC 16-bit SPI
SDMF Tracking
ASIC c) SDMF Estimator
CLK 1 CLK 2 CLK ADC EOC 4-signal IN/OUT
Figure 5.11: System block architecture of the SDMF tracking ASIC.

Wet Electrode Dry Electrode

+ +
Metal VsEMG VsEMG
- -
Gel <1k
Skin Contact <100k 10nF >1M 10nF
Body Rbody Rbody
Figure 5.12: Skin-electrode impedance model for dry and wet electrodes.
5.3.1 sEMG Front End Signal Processing
The circuit has been implemented in a commercially available 0.35 µm CMOS technology provided
by AMS (C35B4). Standard cell libraries were used for the implementation of the ADC and the
comparator (ADC10A, COMP).
5.3.1.1 Choice of Electrodes
In Chapter 4, the developed MFCV tracking ASIC and the resulting wearable node were designed
to operate using self adhesive gel electrodes (AMBU Neuroline (72000-S/25)). The reasons for this
choice were to provide a lower and better matched impedance interface between the skin and the
electrodes and also provide a medium to attach and hold the wearable node on the user. Consider
Fig. 5.12 for the dry and wet electrode impedance model. By matching the skin-electrode impedance
between the sensing electrodes, mains and common mode noise can be reduced. Furthermore, as seen
in Eq. 5.6 where electrode impedance Zel1,2 is constant, by increasing the front-end amplier's input
impedance Zin common mode noise and other sources of noise to the electrodes can be minimised.
Following the analysis in [20], the skin-electrode to output signal voltage gain is given by Eq. 5.7
and the input refereed noise at the source is given by Eq. 5.8, where:
vs (ω) ⇒ signal source on skin interface,
ii,n (ω) ⇒ net current noise at amplier input,

Yc (ω) ⇒ gc + ωCc skin-electrode coupling admittance,
Yi (ω) ⇒ amplier input admittance,
vi,n ⇒ input referred amplier voltage noise,
Av ⇒ amplier voltage gain.

Zel2 − Zel1
vA − vB = vcm . (5.6)
Zin
Yc (ω)
G(ω) = Av . (5.7)
Yc (ω) + Yi (ω)
Yc (ω) + Yi (ω) ii,n

vs,n = vi,n + . (5.8)
Yc (ω) Yc (ω)
Finally, assuming a noiseless innite input impedance amplier, the source input referred
noise power density reduces to Eq. 5.9.
It is observed that by increasing the amplier input impedance, the electrode impedance
mismatch eects and input signal attenuation is minimised. Furthermore, as expected, a dry elec-
trode would be noisier than a wet electrode [20, 21]. However, wet electrodes can be problematic
with long term monitoring, since the gel slowly dries out and the Ag/AgCl contact does not touch
the skin properly [22]. In addition, dry electrodes are more practical, easy to use and long lasting.
However, due to the lack of a conductive medium (gel) they are susceptible to motion artefacts.
2 4kT gc
vs,rms = . (5.9)
|Yc (ω)|2
The wearable node developed in Chapter 4 performed as expected and the overall imple-
mentation was successfully. However, the device size was relatively large for a convenient and light
wearable. As seen form the analysis in in section 4.6, the device size is governed by the size and
location of the electrodes. Hence, in an eort to minimise device area, increasing monitoring time
and make the node easier to use, it was decided that the next generation of muscle fatigue moni-
toring ASIC would function using dry electrodes. Following the skin-electrode impedance analysis
above, in order to keep the amplier input impedance high the oating high pass lter could not
be used. Hence, the new amplier should be capable of rejecting the DC Polarization Voltage (PV)
c)
a) b) 20
IM8 = 1.6u A
10 2
10 Vbp1 Vbp2
5n A M7 M8 VCM M7 M8 Vb
Ibias = 1u A M3
R1
Vbp3 VCM
280 M5 M6
Vip Vin
IM3 = I+IR1
2 5
IM4 = I-IR1
10
M1 M2 10 R2 Ao Gm M1
C2
10 Cext
M3 M4 10 Ao
Von 5 Vop M3 M4 Vbn3
20
2 Vbn2
40 5
20 Von
M5 M6 Vop
M1 M2
Vbn1
c) d) 20
2
10
10 Vbp1 Vbp2 VCM M7 M8
Ibias = 1u A M3 M4
VCM/Ao- Vbp3
M5 M6
20 10
20 VCM 10
Gm M1 M2 Ao+
Cext Ao- 10 R3 R4
Ao/Ao+ 10
IM5 = 3.2u A
Vbn3 M3 M4
M5 Vbn2
5
Vop 20 Von
M1 M2
Figure 5.13: (a) transconductance, (b) transimpedance (single ended), (c) DC servo and (d) tran-
simpedance (double ended) circuit implementations.
from the biopotential electrodes. Although it is reported in literature that PV voltage can reach
300 mV [23, 24], measurements using the MFCV ASIC, by lowering its externally controlled gain,
and steel electrodes (1 cm2 ) indicated that such a high PV is unrealistic. If the electrodes are rmly
attached to the skin, then PV could be less than 50 mV. The solution would be to use an IA with
a DC servo loop to cancel out the electrode oset, in the expense, however, of power consumption.
5.3.1.2 Analogue Front End
Another aspect of the MFCV ASIC that could be improved was the power consumption of the
instrumentation amplier. The standard three opamp IA, although a popular choice for biopotential
measurements [23, 2529], triples the system's power consumption and doubles the amplier noise
contributions. Moreover, the Common Mode Rejection Ratio (CMRR) of the three-opamp IA
depends on the matching of the resistors. This leaves two other choices for a front end architecture
with a high imput impedance. Either a Harrison amplier architecture [30] or a current feedback
amplier architecture. Although the Harrison amplier oers great performance, the long time
constants associated with the big input capacitors and huge feedback resistors and the possible
leakage eects that can cause the output DC level to drift (described in Chapter 3, section 3.5)
make it not the optimum choice for the intended application. As the sEMG signal is susceptible
to large spikes/motion artefacts, the Harrison amplier would have to reset very often and possibly
take a long time to return to the nominal DC operating point. Since the presented algorithms
work with zero crossings, it would be better to avoid this topology. The current feedback topology
is also widely used in biopotential measuring systems [24, 3136] and oers a solution to all the
considerations described so far. The gain of the current feedback IA is precisely dened by the ratio
of two resistors and the CMRR does not rely on the matching of the resistors. The developed IA
design was adapted from [24].
As seen in Fig. 5.11, the current feedback IA consists of a transconductance (TC) input
stage, a transimpedance (TI) output stage and a DC servo loop for rejecting the electrodes' as well
as the amplier's DC oset, allowing the use of a higher gain. For simplicity, the IA is followed by
a PGA with the same architecture. The overall noise performance is dominated by the rst stage
(IA) while the gain is precisely controlled by the second stage. The TC, TI (single ended), DC servo
and TI (double ended) circuit implementations are shown in Fig. 5.13a,b,c and d respectively. At
the TC stage, the dierential input signal from the electrodes is copied to the terminals of R1 and
the same current deects the currents of transistors M3 and M4 . The input dynamic range of the
amplier is given by Eq. 5.10. R1 was set to 12.5 kΩ and the quiescent IM 3,M 4 current to 800 nA
resulting to an input dynamic range of 10 mV, in an eort to balance power consumption, dynamic
range and noise performance. According to [24], transistor sizing allowed for Rout,M 5,6 to be much
1 1 1
larger than
gm1 , gm3 and R1 .
DR = ±(IM 3,M 4 × R1 ). (5.10)
The TI stage converts the TC output current from transistors M5,6 into voltage using the
replicated transistors M1,2 . Transistors M7,8 of the TI stage force the same biasing current as M7,8
of the TC stage. Thus, half the current dierence of M1,2 of the TI stage ows through the load
R2/3+4
resistor R2 or R3 + R4 . The voltage gain of the two combined stages is derived as G= R1 and
with R3 + R4 = 200M Ω the expected gain is 44.1 dB. For the PGA, R1 is set to R1 = 770kΩ for
a bigger DR and the gain is controlled with a 5-bit adjustable R2 resistor, oering a gain range
between 1 and 32. Hence, the overall gain varies from 160 to 5130 (44.1-74.2 dB).
Ibias = 6.5n A
Ibias = 6.5n A
Gm2
Gm1 C2
C1
Vin
Figure 5.14: Second order, low pass lter, gm-C architecture.
Noise is mostly dened by the TI stage, and it is optimised by forcing transistors M3−8
to operate in strong inversion. Finally, the DC servo loop is responsible for ltering out the DC
electrode oset. A gm − C low pass lter in negative feedback forms an integrator which settles to
the output DC operating point of the IA. Due to the large time constants needed, the capacitor used
is external with a value of 10 µF. The resulting DC level is used to adjust the bias current of the
TI stage and hence, control the common mode voltage of the IA. The high pass corner frequency
of the DC servo loop is 6.3 Hz and it is capable of rejecting an electrode oset up to 48.9 mV.
The resulting high-pass lter behaviour is advantageous for ltering out motion artefacts and other
sources of noise to the amplier, since the sEMG power is concentrated above 10 Hz.
5.3.1.3 sEMG Filtering
A second order, low-pass gm − C lter from [37] was implemented to limit the bandwidth of the
recorder signal. The lter architecture is shown in Fig. 5.14, where the transfer function and corner
frequency are given by Eq. 5.11 and Eq. 5.12. The operation transconductance amplier (OTA)
used employed bump linearisation transistors to further reduce the transconductance and hence the
lter corner frequency. By operating the dierential pair in weak inversion, the gm is accurately
Ids
controlled by the biasing current as gm = nVt . With gm1 = gm2 , Ibias = 6.5 nA, C1 = 14.3 pF and
C2 = 7 pF the resulting corner frequency was 788 Hz.
gm1gm2
G(s) = . (5.11)
s2 C 1 C2 + sC1 gm2 + gm1gm2
gm
ωo = √ . (5.12)
C1 C2
5.4 SDMF Tracking ASIC Experimental Results 134
Figure 5.15: Die micro-photograph of the SDMF tracking ASIC.
5.4 SDMF Tracking ASIC Experimental Results
The SDMF tracking ASIC micro-photograph is shown in Fig. 5.15. The IA and LPF performance
characteristics are a function of the biasing current used. A master current of 3.2 µA was created
and then divided and distributed across the circuits. A copy of this current was sent to a pad in
order to better characterise the performance of the ASIC. The averaged master current from four
chips was 3.63 µA which implies a 13.4% increase in the biasing current. The average measured
power consumption of the IA, including the TI, TC and DC servo loop, was 18.1 µA. However,
most of the power, 8.2 µA, is consumed by the DC servo loop. The total static and dynamic
power consumption of the digital core (measured at 10 kHz) is 4.32 µW and of the analogue front
end (measured) is 125.1 µW from a 3.3 V supply. Table 5.3 summarises the power consumption
breakdown of the SDMF ASIC. The SDMF algorithm comprises of 50% more DFFs than the MFCV
algorithm. Moreover, the min function also performs subtraction, hence the overall number of gaits
is increased compared to the max function used in the MFCV algorithm. This increase in used
gates explains the increase in power of the SDMF algorithm compared to the MFCV algorithm.
Table 5.3: SDMF ASIC power breakdown (C35, 3.3V).

Core Measured Current
IA 18.1 µA
PGA 18.5 µA
SDMF 1.31 µA
Estimated
TI 4.95 µA
TC 4.95 µA
DC servo 8.2 µA
ADC 45 µA
Bit-Stream Converter 68 µA
Total power 498 µW
5.4.1 System Validation
Firstly, the analogue front end was characterised with noise, gain and frequency measurements. Fol-
lowing that, the SDMF algorithm was independently veried. Subsequently, appropriate comparator
thresholds were chosen and all three stages (front end, bit-stream converter, SDMF algorithm) were
connected together to test the complete system.
The lowest measured integrated input-referred noise of the complete front-end (0.1-2 kHz)
was 1.45 µV (highest of 4 chips at 1.54 µV), with a noise oor at 106 √nV and a corner frequency
(Hz)
of 2.5 Hz. The IA noise performance is shown in Fig. 5.16a. The measured dierential and common
mode gain of the IA and LPF are shown in Fig. 5.16b. The corner frequencies are 6.9 Hz and 880.5
Hz (6.3, 788.2 simulated) with a CMRR of 90.66 dB at 50 Hz. The measured gain was 48.6 dB,
higher than the simulated value of 45.8 dB. The rst order High Pass Filter (HPF) behaviour of
the DC servo loop and the second order behaviour of the LPF are easily observed. The increase in
the low pass cuto can be attributed to the increase of the master bias current, since the chosen
gm − C lter architecture is very sensitive to biasing. The total harmonic distortion (THD) of the
IA was measured with the help of a SRS SR760 FFT spectrum analyser and a Tektronix AFG3102
arbitrary waveform generator. The Tektronix 100 Hz, 1 Vpp sine wave, after averaging 30 FFT
measurements had THD = 0.539%. The IA, driven by a 100 Hz, 1 mVpp sine wave and an output
of 218 mVpp had THD=0.91%.
The DC servo has a large time constant in order to lter out any DC electrode oset and
stop the amplier from saturating. As a result, when the amplier is turned on, it requires 10
seconds before it reaches it's stable operating point. After the DC servo settles down, it carries on
Simulation
Measurement
60
10-6
10-7 106nV 40
Differential
Power spectrum (V/sqrt(Hz))
10-8 20 Common Mode
Gain (dB)
10-9 CMRR = 90.66 dB
0
10-10
-20
10-11
-40
10-12
10-13 -60
10-4 10-2 100 102 104 106 100 101 102 103 104
(a) (b)
Figure 5.16: (a) Input-referred noise measurement results of the analogue front-end signal path
(IA, PGA, LPF), (b) Common-mode rejection ratio measurement of the IA using a 500mVpp input
signal.
ltering the output of the amplier as expected. Due to the increase of the basing current, the DC
servo dynamic range increased to approximately 60 mV. The start-up transient response of the IA
is shown in Fig. 5.17.
The SDMF algorithm was independently veried before proceeding with complete system
measurements. After isolating the SDMF processing unit, a 100 Hz square wave was applied to it's
input, resembling the output of the PWM. The SDMF was driven with a 4 kHz sampling clock and
a 1 Hz time window. The simulated SDMF settled to an output value of 100 Hz, or an output delay
value of 40 (× 2, including starting delay), which was matched by the measurement results.
The experimental set-up, including the measuring board, microcontroller and dry electrode
conguration, is shown in Fig. 5.18. The electrodes were made of steel disks with 1cm diameter and
Interelectrode Distance (IED) of 2 cm and were attached to metal studs inside an epoxy mould.
In order to minimise external noise interference picked up by the long leads, the electrode cables
were shielded. The shield potential was set to the bias electrode potential of 1.65 V (VCM ). The
positioning requirements for the SDMF ASIC are relatively relaxed compared to the MFCV ASIC.
Thus, there is no need to place the electrodes according to the muscle's innervation zone, allowing
the user to chose the desired location more freely. Nevertheless, the electrodes should be placed
away form the myotendinous zones of the muscle as well as the boarders of the muscle in order to
avoid cross-muscle interference. Interestingly, when the electrodes were placed on the top part of
1.82 Saturated output
1.8
1.78
Output common mode (V)
1.76
1.74
Cext charging
1.72
1.7
1.68
1.66 Chip turn-on 1.65 V input VCM Stable operation
1.64
0 2 4 6 8 10 12 14
Time (s)
Figure 5.17: Transient responce of the IA start-up. The ouput reached a stadle operating point
after 10 seconds due to the large time constant involved with the DC servo loop.
the inner left biceps brachii muscle, a faint Electrocardiography (ECG) signal was picked up, as
seen in Fig. 5.19. The signal was almost burred in noise and hence would not pose any problems
during the sEMG measurement and processing stages.
The issue of setting the correct threshold for the comparator was addressed for the rst time
in Chapter 4, section 4.5.1. In the MFCV ASIC, careful threshold selection placed it above the
noise oor. The threshold should be low enough to be closer to the base of the sEMG MUAP
spike and at the same time be above the peak noise amplitude. The comparators used in both
ASICs exhibit some small (<5 mV) hysteresis to combat with the presence of high frequency noise
while switching. Only one transition from high-to-low or low-to-high is desired. Without hysteresis
every time the positive terminal crosses the negative, due to the presence of noise, there would be
several transitions of the output. As a result, the wrong bit-stream would be produced and alter the
MFCV/SDMF results. Moreover, due to parasitics capacitances in the layout of the ASIC (or even
PCB designs) some hysteresis is required to deal with kickback noise every time the comparator
switches. In Chapter 4, it was concluded that a comparator with increased hysteresis limits could be
a better approach than manually setting the threshold level. The hysteresis threshold levels should
be placed above and below the maximum noise amplitude and centred around the output common
Coax-cable, shield at Vcm
Battery supply
Steel disks
d = 1 cm
Cext KL26Z to PC
ASIC
VB Vin+ Vin-
Electrodes Measuring Unit

Figure 5.18: PCB measurement board for the SDMF ASIC. The dry electrodes are made of steel
disks, 1cm in diameter. The ASIC communicates withthe microcontroller via SPI and the microcon-
troller to the PC via UART.
2.5
Heartbeats 2
2
1.5
1.5
1
1
0.5
Amplitude (mV)
0.5
0 0
-0.5 -0.5
-1
-1
-1.5
-1.5
-2
-2
-2.5
0 1 2 3 4 5 6 7 3.8 4 4.2 4.4 4.6 4.8
Time (s) Time (s)
Figure 5.19: Transient reconding from the upper part of the inner left biceps brachii muscle. ECG
heartbeats were detected and shown in detail in the left gure.
2.1
VLH = 1.658 V
2 VCM = 1.650 V
VHL = 1.642 V
1.9
1.8
Amplitude (v)
1.7
1.6
1.5
VCM _
1.4 VOUT
+
1.3
RI RF
VIN
1.2
1 1.1 1.2 1.3
Time (s)
Figure 5.20: Shown are an sEMG recording and the position of the selected hysteresis thresholds.
Hysteresis is introduced with the aid of positive feedback, ass seen in the bottom left conrer.
mode voltage. Since this IA uses a DC servo loop to lter out electrode and amplier DC osets,
the output VCM is well dened and constant across dierent chips. Furthermore, the comparator's
positive and output terminal were connected to pads to allow for post-fabrication changes or even
bypassing of the comparator. By connecting two external resistors and forming a positive feedback,
hysteresis was added to the comparator. Consider Fig. 5.20 for the circuit implementation and
graphical representation of the hysteresis limits. By selecting RI = 25 kΩ and RF = 5.1 MΩ the
resulting hysteresis was 16 mV and centred around VCM. This way, when there is no contraction
and there is not any sEMG present, there would be no change at the input of the SDMF algorithm.
Hence, when there is sEMG present, the algorithm would not compute wrong values by taking into
account segments with noise.
5.4.2 Physiological Measurements
Simulations of the SDMF algorithm were performed using a 4 kHz sampling clock and a operating
window of 1 s. However, it was found experimentally that tracking accuracy and system responsive-
ness were increased by using a 3 kHz sampling clock and a 330 ms operating window. A total of ve
volunteers were used to evaluate the performance of the SDMF tracking ASIC. The experimental
2.22.2 0.250.25
2.12.1
2 2 0.2 0.2
1.91.9
1.81.8 0.150.15
Amplitude (V)
Amplitude (V)
Amplitude (V)
Amplitude (V)
1.71.7
1.61.6 0.1 0.1
1.51.5
smoothed
smoothed sEMG
sEMG Increasing
Increasing withwith fatigue
fatigue
1.41.4 0.050.05
1.31.3
1.21.2 0 0
0 0 10 10 20 20 30 30 40 40 50 50 60 60 70 70 0 0 10 10 20 20 30 30 40 40 50 50 60 60 70 70
Time
Time (s) (s) Time
Time (s) (s)
(a) (b)
Figure 5.21: (a) sEMG recording of a contraction starting from a rested state until exhaustion, (b)
smoothed sEMG amplitude using a 300 ms moving average window.
protocol was a simplied version of the protocol used in Chapter 4, section 4.5.2.1). A smaller
subject group of n=5 and a simplied protocol were chosen due to time constrains involving the
approval of a new ethics application.
The subjects were asked to stand or sit with their back and elbow xed against a wall in
order to minimise compensatory movements. Subjects were required to keep in contact with the wall
throughout the whole duration of the testing. Following that, the subjects were asked to perform
an isometric contraction by lifting a 10 kg load. The muscles started from a completely rested state
until exhaustion. Fig. 5.21a shows a sEMG recording from the left bicept brachii muscle during
a fatiguing contraction. The signal amplitude is approximately 500 mVpp . Fig. 5.21b shows the
smoothed response of the absolute of the sEMG signal. A moving average lter with a window
length of 300 ms was used. It can be observed that as the muscle fatigues, the amplitude of the
sEMG signal gradually increases. During this contraction, the smoothed amplitude almost doubles
in amplitude. There is a rapid increase in the sEMG amplitude at the start of the contraction as
force is applied gradually. Following that, the sEMG amplitude increases steadily as force remains
constant and the muscle fatigues. As explained in Chapter 2 and observed in this measurement,
the amplitude of the sEMG signal could be used as a reliable indicator of fatigue.
The accuracy of the SDMF ASIC was established by comparing the ASIC results to
MATLAB
® . The amplied sEMG signal from the ASIC is recorded using a 16-bit ADC by ADIn-
struments. The iMDF is computed in MATLAB

® using the medf req() function and then compared
to the ASIC SDMF estimate. Fig. 5.22b-f shows the SDMF estimate (blue) compared to the equiva-
lent iMDF values (red). By adding hysteresis to the comparator when there is no sEMG present the
SDMF algorithm does not compute wrong estimates by taking into account noise crossings. Hence,
as seen in Fig. 5.22a in the highlighted green regions, it is easier to discriminate the start of the
contraction using SDMF than observing iMDF. When computing the iMDF of an sEMG signal, the
part of the signal containing only noise should be identied and removed. It is clear that the SDMF
ASIC is capable of tracking the spectral changes and compression of the sEMG power spectrum.
As a result, it can be used as an estimator of muscle fatigue.
5.4.3 Discussion
Although the SDMF is used to get an insight on the spectral changes of the sEMG signal, it cannot
be compared iMDF directly. Hence, the two metrics are smoothed using a moving average lter
with a width of 20 samples. The corresponding R, R

2 and RMS values of the comparison between
the smoothed SDMF and iMDF results are shown in Table 5.4. The average correlation between
them is 0.945, verifying the strong relationship between the two metrics. If the DC oset between
SDMF and iMDF is removed, then the MARD between them can be calculated, although they are
two dierent metrics. The average MARD is 3.4%.
Table 5.4: Correlation coecient values between the smoothed iMDF and measured
SDMF.
Subject 1 Subject 2 Subject 3 Subject 4 Subject 5
Smooth Smooth Smooth Smooth Smooth
R 0.97 0.98 0.89 0.92 0.97
R2 0.94 0.96 0.79 0.85 0.94
RMS 1.56 1.68 3.32 3.11 2.92
MARD SDMF-8 Hz
2.62%
SDMF-14 Hz
2.05%
SDMF-5 Hz
3.67%
SDMF-8 Hz
4.07%
SDMF-0 Hz
4.62%
Having established the functionality of the analogue front end and the SDMF algorithm, the
eect of the gain setting on the SDMF performance was evaluated. The sEMG signal from two
subjects was saved at a sampling frequency of 50 kHz and played back to the system using the
12-bit DAC on the FRDM-K64F microcontroller board followed by a 40 dB signal attenuator. The
input signal had an amplitude of 1 mV and the PGA gain was increased for every measurement
from 1 to 10. Fig. 5.23 shows the playback SDMF results for two subjects. It was observed that
after a certain point, the chosen gain had little to no eect on the SDMF output. On average, after
120 100 100
120 1
110 90 90
100 100 1
80 80
90
80 70 70
SDMF (Hz)
SDMF (Hz)
SDMF (Hz)
SDMF (Hz)
SDMF (Hz)
80
60 60 60
70
40 50 50
60
50 20 40 40
40 30 30
60 0 50 20 40
100 60 80
150 100 120
200 140 160
250 180 0 10 0 2010 30
20 3040 40 50 50 60 60 70
Time (s) Time (s) Time (s) Time (s)
(a) (b)
120 100 100 90
120
110 90 90
80
100 100
80 80
70
90
80 70 70
SDMF (Hz)
SDMF (Hz)
SDMF (Hz)
SDMF (Hz)
SDMF (Hz)
80 60
60 60 60
70
50
40 50 50
60
40
20 50 40 40
40 30 30 30
20 400 60 80
50 100 120
100 140 160
150 180 200 250 0 10 0 20 10 30 20 40 30 50 60
40 70
50 60 0
(c) (d)
90 90 90 90
80 80 80 80
70 70 70 70
SDMF (Hz)
SDMF (Hz)
SDMF (Hz)
SDMF (Hz)
60 60 60 60
50 50 50 50
40 40 40 40
30 30 30 30
70 60 70 0 5 0 10 5 15 1020 1525 2030 2535 3040 3545 40 45 0 5 0 10 5 15 1020 1525 2030 2535 3040 3545 40 45
(e) (f )
Figure 5.22: (a) Comparison of the SDMF (red) and iMDF (blue) trends before and after the muscle
contraction (green regions), (b-f) SDMF (red), iMDF (blue) and their smoothed (black) trends for
Subjects 1-5.
9090 100100
G =G7= 7
8585 G =G6= 6
G =G5= 5 G =G
8=8
G =G4= 4 90 90 G =G
5=5
8080 G =G3= 3 G =G
3=3
G =G2= 2
G =G1= 1
7575
80 80
7070
SDMF (Hz)
SDMF (Hz)
6565 70 70
6060
60 60
5555
PGA
PGA gain
gain
5050
50 50
4545
4040 40 40
0 0 20 20 40 40 60 60 80 80 100100 120120 0 0 50 50 100100 150150 200 200 250 250
Time
Time (s) (s) Time
Time (s) (s)
(a) (b)
Figure 5.23: (a) Smoothed SDMF responce with sEMG playback from Subject 1 and a varying
PGA gain from 1 to 7, (b) SDMF plot and smoothed responce with sEMG playback from Subject 2
and a varying PGA gain from 3 to 8.
the sEMG signal was amplied to an amplitude above 0.4 Vpp the SDMF algorithm could produce
an accurate measurement. In the measurements presented in Fig. 5.22, the PGA gain was set to 5
and the sEMG amplitude was approximately 0.5 Vpp .
There are several sources of error for the SDMF algorithm its comparison with iMDF. The
algorithm resembles the operation of the MFCV estimator. Hence, the accuracy is limited by the
number of available delay stages. Moreover, due to the operation of the algorithm, it is possible
for smaller length pulses to be counted in the longer delay stages, thus skewing o the SDMF
reading. The processing windows should be carefully aligned such as every signal epoch contains
the same information to compute either SDMF or iMDF. Any misalignment of the signals could
yield dierent absolute values per segment, but keep the overfall shape and trend of the spectral
compression. Finally, a better way to integrate the pulse length distribution produced by the
SDMF algorithm would be to progressively shift through all the delay stage counters starting from
the bottom until the desired value is reached. Compared to the current method, which looks into
all the counter values at the same time, the proposed method would always settle to the rst
intersection of the correct value and not be susceptible to errors due to the occurrence of equal
values. Also, it would not be biased by the shorter pulses counter in the longer delay stages, as it
would probably have stopped shifting through them before reaching the longer delays. The system
performance specications are summarised in Table 5.5. The developed ASIC realised a real-time
medial pulse length tracker with a dedicated sEMG IA and SPI programmable PGA.
Table 5.5: Summary of SDMF tracking ASIC.

Technology 0.35 µm 2P4M
Voltage Supply 3.3 V
Current Consumption Total: 151µA
SDMF: 1.31 µA
Amplier: 36.6 µA
Bit-Stream Converter: 68 µA
ADC: 45 µA
Total power 498 µW
CMRR 90.66 dB
Gain Adjustable (48.6-78.7 dB)
Bandpass frequency range 6.9 Hz, 880.5 kHz
Input dierential voltage range 10 mV
DC oset cancellation 60 mV
Input Referred Noise Density nV
106 √
Hz
Computation window Adjustable by CLK2
Estimation Accuracy Dependable on CLK1
Number of Delays 40
SDMF, iMDF correlation 0.945
It has been demonstrated that SDMF can be a reliable estimator of muscle fatigue and that is
correlates well with iMDF. At the same time, the SDMF algorithm is simpler to implement than any
of the two iMDF systems presented in Chapter 3. Moreover, the implementation is lower power and
robust to motion artefacts and process and mismatch variation due to its digital architecture. The
SDMF algorithm is more reliable, accurate, lower power and for these reasons it will be preferred
over the iMDF tracking systems.
5.4.4 The second version of the Wearable Muscle Fatigue Monitor
The SDMF tracking ASIC was integrated into a Wearable Muscle Fatigue Monitoring (MFM) node,
similar to the wireless node presented in Chapter 4, section 4.6

1. The second MFM node is shown in
Fig. 5.24a,b. The node features the same Nordic Semiconductor nRF51822 2.4 GHz Bluetooth Low
Energy System-on-Chip (SoC). The rst MFM device featured a large 300 mAh battery, oering 27
h of continuous BLE operation. In order to decrease the overall size and weight of the new device,
the battery capacity was reduced from 300 mAh to 165 mAh, oering an approximated battery
life of 15 h of continuous BLE operation. These changes allowed the node to be less obstructive
and hence achieve a highly ergonomic design. A size comparison between the two MFM versions is
1 The wearable device PCB was designed by Silvia Gallo and the and the control application was updated by
Usama Arif under the direct supervision of Ermis Koutsos.
(a) (b) (c)
(d) (e)
Figure 5.24: (a) Revised Wireless MFM node, bottom view, (b) Revised Wireless MFM node, top
view, (c) Size comparison between old and new MFM versions, (d) Size comparison between old and
new MFM versions with their cases on, (e) Size comparison between wet electrode sockets and custom
made dry electrodes.
shown in Fig. 5.24c,d and e.
A major advancement of the SDMF ASIC front end amplier was its ability to operate with
dry electrodes. Therefore, the revised wireless node was updated to feature dry, custom made
electrodes. The electrodes for the new node were made from stainless steel bars, 2 mm in diameter
and 15 mm in length for the sensing electrodes and 25 mm in length for the biasing electrode. This
has allowed for a 65% decrease in weight (from 37.6 g to 13.1 g) and 61% decrease is volume (from
40×62×16 mm to 34×41×11 mm). The rst MFM device featured long cables to connect the PCB
to the electrode studs attached on the case. Although these cables allowed for an easy access to the
electronics once the case was opened, they increased the overall height requirements and contributed
to the accumulation of additive noise. Thus, metal springs were employed to ensure the metal bars
(electrodes) were in contact with the PCB, shown in Fig. 5.24c. This allowed for easier handling of
the device, reduced overall height and elimination additional sources of noise.
(a) (b)
Figure 5.25: (a) Application window where the user can select number and location of wireless
MFM nodes, (b) Application window where the user can save muscle names and locations used.
The new node was designed to be attached to the user with the use of adhesive plasters
wrapping around the device, instead of adhesive electrodes. Thus, the monitor can stay rmly
attached to the user for longer periods of time. Moreover, due to the increase in the adhesive
area and decrease is size and weight, the new node is better secured on the skin, allowing for less
movement. The increased stability can combat the tendency of dry electrodes to generate motion
artefacts due to the absence of a conductive gel.
The Android application provides the user interface to the wearable. When the application
acquires data from the wearable, it stores the data in a dierent directory for each registered user.
After connecting to the wearable, the Android application allows the user to set the sampling
intervals control the ASIC's functions. The application was updated to allow the user to connect
up to four nodes into one mobile device and save their location on the body, as seen in Fig. 5.25a,b.
Lastly, a colour coded scheme (green, yellow, red) as an indication of the approximated level of
fatigue (Fig. 5.25a). This addition allows the users to quickly asses their fatigue level without the
need to continuously monitor the detailed SDMF results.

5.5 Summary 147
5.5 Summary
This chapter presented a system capable of tracking the sEMG spectral changes by processing the
signal in the time domain. The sEMG power spectrum is closely linked to the MUAPs' conduction
velocity and ISI. It has been shown that the ZCR correlates with the mean frequency of the power
spectrum, however it is a simplistic representation of the mean number of spikes. Creating a
distribution of the time interval of the crossings and estimating its median is a more promising,
accurate and robust metric for tracking the power spectrum characteristics. In this chapter it is
demonstrated that it can be a reliable estimator of muscle fatigue.
The developed system oers a simpler, low power and robust to motion artefacts method for
estimating the median spike length of an sEMG signal. It is further shown that this estimate strongly
correlates with the median frequency of the sEMG signal (correlation > 0.9). The sEMG signal is
processed as a bit-stream, showing that it is possible to accurately track the spectral compression
of the sEMG during fatigue with one bit representation. Furthermore, the developed algorithm
can be easily adapted to estimate the distribution of spike duration, ISI and spike period. This
information would not be possible to obtain without the use of more advanced signal processing
techniques (STFT) since it cannot be cannot be retrieved simply with ZCR monitoring or peak
counting. Hence, the presented SDMF metric can oer signicant insight on muscle fatigue and
sEMG statistics and help to better understand muscle fatigue.
A digital algorithm for nding the median length between crossings was coupled with a low
noise amplier, a PGA and a LPF and utilises a standard SPI communication protocol. The sEMG
front end amplier was optimised to operate with wet and dry electrodes, increasing the system's
exibility. Moreover, the presented ASIC was integrated in a small and lightweight wearable device
which can serve as a platform to expand the knowledge about the formation of sMEG from MUAPs
during fatigue. As such, it serves as an expansion of the capabilities of the iMDF and MFCV tracking
systems, which provide information about muscle fatigue and recovery, but not the underlying
mechanisms.
REFERENCES 148
References
[1] G. Inbar, O. Paiss, J. Allin, and H. Kranz, Monitoring surface emg spectral changes by the zero
crossing rate, Medical and Biological Engineering and Computing, vol. 24, no. 1, pp. 1018,
1986.
[2] R. Merletti, L. Lo Conte, and C. Orizio, Indices of muscle fatigue, Journal of Electromyography
no. 1, pp. 1836, 2000.
[4] M. Cifrek, V. Medved, S. Tonkovi¢, and S. Ostoji¢, Surface emg based muscle fatigue evalua-
tion in biomechanics., Clinical biomechanics (Bristol, Avon), vol. 24, no. 4, p. 327, 2009.
2001.
[6] O. Dayan, I. Spulber, A. Eftekhar, P. Georgiou, J. Bergmann, and A. McGregor, Applying
emg spike and peak counting for a real-time muscle fatigue monitoring system, in Biomedical
Circuits and Systems Conference (BioCAS), 2012 IEEE, pp. 4144, IEEE, 2012.
[7] G. Hägg, Electromyographic fatigue analysis based on the number of zero crossings, Pügers
Archiv European Journal of Physiology, vol. 391, no. 1, pp. 7880, 1981.
[8] C. J. De Luca, Physiology and mathematics of myoelectric signals, IEEE Transactions on

Biomedical Engineering, no. 6, pp. 313325, 1979.
[9] P. Lago and N. Jones, Eect of motor-unit ring time statistics on emg spectra, Medical and
Biological Engineering and computing, vol. 15, no. 6, pp. 648655, 1977.
[10] L. H. Lindstrom and R. I. Magnusson, Interpretation of myoelectric power spectra: a model
and its applications, Proceedings of the IEEE, vol. 65, no. 5, pp. 653662, 1977.
[11] M. B. Reaz, M. Hussain, and F. Mohd-Yasin, Techniques of emg signal analysis: detection,
processing, classication and applications, Biological procedures online, vol. 8, no. 1, pp. 1135,
2006.
REFERENCES 149
[12] C. J. De Luca, The use of surface electromyography in biomechanics, Journal of applied

biomechanics, vol. 13, pp. 135163, 1997.
[13] J. Finsterer and B. Mamoli, Turn/amplitude parameter changes during sustained eort,
Electroencephalography and Clinical Neurophysiology/Electromyography and Motor Control,

vol. 101, no. 5, pp. 438445, 1996.
no. 1, pp. 1836, 1999.
[15] S. O. Rice, Mathematical analysis of random noise, Bell System Technical Journal, vol. 23,
no. 3, pp. 282332, 1944.
[16] F. B. Stulen and C. J. De Luca, Frequency parameters of the myoelectric signal as a measure of
muscle conduction velocity, IEEE Transactions on Biomedical Engineering, no. 7, pp. 515523,
1981.
[17] T. Schweitzer, J. Fitzgerald, J. Bowden, and P. Lynne-Davies, Spectral analysis of human
inspiratory diaphragmatic electromyograms, Journal of Applied Physiology, vol. 46, no. 1,
pp. 152165, 1979.
[18] R. Merletti, M. A. Sabbahi, and C. J. De Luca, Median frequency of the myoelectric signal,
1984.
[19] D. A. Gabriel, Reliability of semg spike parameters during concentric contractions, Elec-
tromyography and clinical Neurophysiology, vol. 40, no. 7, pp. 423430, 2000.
[20] Y. M. Chi, T.-P. Jung, and G. Cauwenberghs, Dry-contact and noncontact biopotential elec-
trodes: methodological review, IEEE reviews in biomedical engineering, vol. 3, pp. 106119,
2010.
[21] P. Laferriere, E. D. Lemaire, and A. D. Chan, Surface electromyographic signals using dry
electrodes, IEEE Transactions on Instrumentation and Measurement, vol. 60, no. 10, pp. 3259
3268, 2011.
[22] D. Hewson, J. Duchêne, and J.-Y. Hogrel, Changes in impedance at the electrode-skin interface
of surface emg electrodes during long-term emg recordings, in Engineering in Medicine and
REFERENCES 150
Biology Society, 2001. Proceedings of the 23rd Annual International Conference of the IEEE,
vol. 4, pp. 33453348, IEEE, 2001.
[23] E. M. Spinelli, R. Pallàs-Areny, and M. A. Mayosky, Ac-coupled front-end for biopotential
measurements, IEEE Transactions on Biomedical Engineering, vol. 50, no. 3, pp. 391395,
2003.
[24] R. F. Yazicioglu, S. Kim, T. Torfs, H. Kim, and C. Van Hoof, A 30 w analog signal processor
asic for portable biopotential signal monitoring, IEEE Journal of Solid-State Circuits, vol. 46,
no. 1, pp. 209223, 2011.
[25] R. Pallas-Areny and J. G. Webster, Ac instrumentation amplier for bioimpedance measure-
ments, IEEE Transactions on Biomedical Engineering, vol. 40, no. 8, p. 831, 1993.
[26] M. J. Burke and D. T. Gleeson, A micropower dry-electrode ecg preamplier, IEEE Trans-
actions on Biomedical Engineering, vol. 47, no. 2, pp. 155162, 2000.
[27] E. M. Spinelli, N. Martínez, M. A. Mayosky, and R. Pallàs-Areny, A novel fully dieren-
tial biopotential amplier with dc suppression, IEEE transactions on biomedical engineering,

vol. 51, no. 8, pp. 14441448, 2004.
[28] C.-J. Yen, W.-Y. Chung, and M. C. Chi, Micro-power low-oset instrumentation amplier
ic design for biomedical system applications, IEEE Transactions on Circuits and Systems I:
Regular Papers, vol. 51, no. 4, pp. 691699, 2004.
[29] J. Huijsing, Operational ampliers: theory and design. Springer Science & Business Media,
2011.
applications, IEEE Journal of solid-state circuits, vol. 38, no. 6, pp. 958965, 2003.
[31] H. Krabbe, A high-performance monolithic instrumentation amplier, in Solid-State Circuits

Conference. Digest of Technical Papers. 1971 IEEE International, vol. 14, pp. 186187, IEEE,
1971.
[32] F. Eatock, A monolithic instrumentation amplier with low input current, in Solid-State
Circuits Conference. Digest of Technical Papers. 1973 IEEE International, vol. 16, pp. 148
149, IEEE, 1973.

REFERENCES 151
[33] A. P. Brokaw and M. P. Timko, An improved monolithic instrumentation amplier, IEEE
Journal of Solid-State Circuits, vol. 10, no. 6, pp. 417423, 1975.
[34] M. S. Steyaert, W. M. Sansen, and C. Zhongyuan, A micropower low-noise monolithic instru-
mentation amplier for medical purposes, IEEE Journal of Solid-State Circuits, vol. 22, no. 6,
pp. 11631168, 1987.
[35] R. Martins, S. Selberherr, and F. A. Vaz, A cmos ic for portable eeg acquisition systems,
IEEE Transactions on Instrumentation and measurement, vol. 47, no. 5, pp. 11911196, 1998.
[36] R. F. Yazicioglu, P. Merken, and C. Van Hoof, Integrated low-power 24-channel eeg front-end,
Electronics Letters, vol. 41, no. 8, pp. 457458, 2005.
[37] R. L. Geiger and E. Sanchez-Sinencio, Active lter design using operational transconductance
ampliers: a tutorial, Circuits and Devices Magazine, IEEE, vol. 1, no. 2, pp. 2032, 1985.
152
Chapter 6
A sEMG Analysis System-on-Chip for

Muscle Fatigue Monitoring
6.1 Introduction
S
O far, from the three systems presented, two have been tested and validated with human
subjects. Each one looks into the dierent properties of the surface EMG (sEMG) signal
during fatigue. Furthermore, each chip iteration oered a more complete version of a muscle fatigue
monitoring system. This chapter presents the nal attempt to create a complete, reliable and
autonomous system capable of tracking muscle fatigue. Improvements were made to the previous
systems to make the new Application Specic Integrated Circuit (ASIC) more robust, requiring less
external components on the interfacing board and allow it to work in real-life scenarios.
The Muscle Fibre Conduction Velocity (MFCV) can oer valuable information about the
muscle recovery and lactic acid concentration. Furthermore, Spike Duration based Median Fre-
quency approximation (SDMF) closely relates to instantaneous Median Frequency (iMDF), thus
oering a unique look into the spectrum compression. Spike counting is a simple metric that relates
to the Motor Unit (MU) ring rate and is believed to relate to muscle engagement and force output
[13]. Lastly, the Mean Absolute Value (MAV) of the sEMG amplitude is an indication of muscle
engagement, applied force, muscle fatigue/recovery and as mentioned in Chapter 2, it plays a vital
role in the establishment of a fatigue index using Joint Analysis of Amplitude and Spectrum (JASA).
Creating a system capable of providing all that information allows for a collective monitoring of the
muscle that oers a unique insight in muscle function, fatigue and recovery.
As seen in the previous chapters, there was a natural progression in the each of the presented
systems' capabilities. Chapter 3 showcased an circuit implemented in CMOS capable of tracking

6.2 sEMG SoC Requirements 153
iMDF. However, the iMDF ASIC did not have an supporting circuits to form a fatigue monitoring
tool. Chapter 4 presented a MFCV tracking system with a xed gain amplier and a simple Serial
Peripheral Interface (SPI) communication interface. The SDMF ASIC in Chapter 5 included an
programmable Gain Amplier (PGA), a better band-pass lter and was designed to work with
dry electrodes, oering greater exibility and increased ergonomics for the user. This chapter
presents the combination of the best aspects of the human validated systems along with certain
improvements to increase the device usability. All previous systems were designed to operate under
static conditions. These type of contractions are very useful to help characterise the muscle, better
understand sEMG and in turn, the underlying reasons for muscle fatigue and loss of force generation.
Hence, the developed systems are ideal tools intended for laboratory use. Contractions is real life
however, can be static but are mostly dynamic (e.g. walking, exercising, lifting). A system able to
perform well under static and dynamic contractions would vastly increase the areas of application,
expanding the device's use to real life, everyday, personal monitoring. This chapter presents a
complete System-on-Chip (SoC) for the analysis of sEMG signals and muscle fatigue monitoring.
6.2 sEMG SoC Requirements
The presented MFCV tracking ASIC in Chapter 3 proved that an ASIC can be used to monitor
muscle fatigue and take part in the development of wearable monitoring systems. However, with
a xed amplier gain, any dierence between the two sEMG channels could not be accounted for.
Also, due to the presence of DC osets, it was not possible to achieve high gains. Although there was
a timing circuit present in order to control the algorithm operation and data output of the chip, all
the clocks were supplied individually, leaving the task to the microcontroller. Lastly, there was no
common mode generator or ground buers inboard, requiring external components. Furthermore, it
was designed to operate with gel electrodes. This type of electrode oers good surface interface with
low resistance, low noise and adhere well to the skin. However, it was observed that wet electrodes
have a limited working life, as the gel and glue dry out. Since there was not an ADC on-board the
ASIC, this task was left to to an external microcontroller. The SDMF ASIC featured a variable
gain amplier and was designed to operate with dry electrodes. As described in Chapter 5, the
SDMF ASIC outperforms the iMDF tracking systems presented in Chapter 3 in terms of power,
robustness to motion artefacts and process and mismatch variations, reliability and accuracy. Hence,
the SDMF algorithm is preferred over the iMDF tracking algorithms. Again ,this ASIC did not
6.2 sEMG SoC Requirements 154
feature a clock signal generator, thus requiring several input clock signals. The on-board ADC was
a great addition, relaxing the requirements of the microcontroller and increasing the robustness of
the system, as the signal was digitised on chip and was less susceptible to external sources of noise,
either from the instrumentation board or the microcontroller.
As a result, it was sensible to combine the best aspects of both ASICs and try to increase
their capabilities. Hence, the MFM ASIC features both the SDMF and MFCV algorithms. It has an
ADC to digitise and stream the amplied sEMG from both channels. Both sEMG channels feature
a separate, adjustable gain, controlled digitally. There is also a ground generation and adjustment
circuit. A timing generator is added that generates all the necessary clock signals and from one
master clock of 1 MHz, while oering the exibility to the user to adjust each signal desperately.
The master clock can be provided by the microcontroller or even by an external oscillator (due to
the common frequency used). A Finite State Machine (FSM) is added to control all the functions of
the ASIC as well as to control the data exchange through the SPI interface. As a result, the number
of external components needed to operate this chip is minimal. The required external components
consist of capacitors only; Cext for the DC servo loop of the ampliers and decoupling capacitors
for the supplies and ground.
As mentioned above, an important step in the creation of a complete fatigue monitoring
SoC is enabling the device to monitor dynamic contractions. In dynamic contractions, muscle
engagement varies with time. In a periodic movement such as walking, one muscle gets activated
and contacts to push us forward and then relaxes as another muscle from a dierent part of the body
takes over. As a result, the observed sEMG signal from that muscle consists of segments where there
is sEMG present and segments where there is not. Proper analysis requires the identication of the
active sEMG segments in order to exclude parts of the recording with no valuable information and
avoid processing the static noise present in the recording. In order to identify the starting points
and stopping points of the sEMG segments in a dynamic contraction recording, an onset detector
is used. The developed Muscle Fatigue Monitoring (MFM) ASIC will compute the MFCV, SDMF,
mean spike count and MAV sEMG amplitude in real time enabling the use of the ASIC in more
applications and more demanding fatigue monitoring systems.

6.3 Adaptation for Dynamic Contractions 155
VTHR
VIN VON
5 x RC
VCM VCM
LPF
Figure 6.1: Onset detector circuit on prototyping board.
6.3 Adaptation for Dynamic Contractions
Both the MFCV and SDMF algorithms use delay stages to process the incoming sEMG bit-stream
in a certain time window. The delay and processing interval are dened by a clock signal. Hence,
it is possible to pause the algorithms when there is no sEMG present so that they do not process
the noise. In order to test this concept, an external onset circuit was constructed on a prototyping
board and paired to the existing SDMF ASIC.
A basic onset detector is fairly simple to construct. The sEMG signal is rectied so as to
get its absolute value. Then, the rectied sEMG signal is low pass ltered to remove the amplitude
uctuations and produce a smoother signal [48]. The ltered signal is compared to a known
threshold value, typically a multiple of the noise standard deviation. When sEMG is present, the
ltered signal will start to rise above the common mode and cross the threshold. By adding a
comparator to the end of this circuit and adjusting the threshold, a pulse indicating the start and
end of an active sEMG segment during a dynamic contraction is generated. It has been shown
[9, 10] that a non-linear signal pre-processor can further increase the EMG onset detection accuracy
in low Signal-to-Noise Ratios (SNR) and can easily be included in further iterations of the current
design, if deemed necessary.
The described onset detection circuit was constructed on a prototyping board and is shown
in Fig. 6.1. A simple inverting rectifying circuit takes the absolute of the incoming sEMG record-
ing. The rectied signal is buered and amplied using an inverting amplier. Following that, a
fth order R-C Low Pass Filter (LPF) tracks the average amplitude. The corner frequency was
empirically set to 25 Hz in order to produce a smooth output DC value equivalent to the MAV of
the sEMG signal. Finally, a comparator produces the required onset pulse, indicating the active
regions in the recording. This circuit operates in real time, along with the SDMF ASIC. The input
to the onset detector circuit is the amplied sEMG from the ASIC.
6.3 Adaptation for Dynamic Contractions 156
The onset pulse is now fed to the microcontroller responsible for generating the two SDMF
clocks, one dening the delay and one the processing time window. A dynamic contraction can
be seen in Fig. 6.2a, where the subject was asked to repetitively lift a load while standing, using
only the arm muscles. The sEMG recording is from the inner left biceps brachii. A closer look to
the individual sEMG active segments can be seen in Fig. 6.2b. As the subject lifted the load, the
sEMG increased in amplitude. When the load was fully lifted, then the subject started lowering it.
During that period, the muscles were less active, hence the sEMG amplitude was smaller. Finally,
as the subjects waited before repeating the process, the sEMG amplitude was zero and only noise
was recorded.
If the ASIC were to process the signal from Fig. 6.2a, the resulting SDMF output would be
as in Fig. 6.2c. The long inactive sEMG periods are considered as pulses and mislead the estimation
algorithm. Given that the onset positions of the sEMG signal are known, one solution would be to
discard any SDMF results where no sEMG is present. This approach is based on the assumption
that the processing window is perfectly aligned with the inactive sEMG period. However, this is
rarely is the case. Hence, as the contraction starts while the algorithm still processes the inactive
period, a wrong estimate is produced. The resulting eect could be described as an oscillation
around the correct SDMF value, as seen in Fig. 6.2c.
Another more promising solution would be to pause the SDMF algorithm during these in-
active periods. In order to examine the eect of the pausing mechanism on the SDMF ASIC a
dynamic sEMG signal was constructed from previous recordings and played back to the system.
This way, it was easier to monitor and adjust the circuit until the desired eect was achieved, with-
out wearing out a human subject. The reconstructed signal and the resulting onset pulse are shown
in Fig. 6.2d. The microcontroller code was modied to take into account the onset detector output.
When the onset was high, the clocks would function normally, thus enabling the SDMF algorithm.
When the onset was low, then the low frequency clock responsible for the processing window, would
stop. Hence, the algorithm would pause generating new results. The computed SDMF for the re-
constructed dynamic sEMG signal without the onset detector is shown in Fig. 6.2e and the SDMF
for the same signal with the onset detector turned on is shown in Fig. 6.2f. The corresponding R,
R
2 and RMS values of the comparison between the smoothed SDMF and iMDF results are shown
in Table 6.1. The correlation between them is 0.97, and the oset free error is less than 5%. It
is clear that the addition of the onset circuit greatly and pausing mechanism greatly improves the
6.4 Circuit Implementation 157
accuracy of the SDMF algorithm during dynamic contractions and expands the ASIC's capabilities.
Table 6.1: Examples of dynamic contractions and the eect of the onset detector and
pausing algorithm on SDMF accuracy.
Subject 1
Smooth
R 0.97
R2 0.94
RMS 1.75
MARD SDMF-10 Hz
4.6%
6.4 Circuit Implementation
The amplier from the SDMF ASIC has been tested and validated in the lab with real measurements.
Hence, the same front-end will be used for the MFM ASIC. Since the MFCV algorithm requires
two channels, the amplier is duplicated. Separate control of the gain between the two channels
will increase the accuracy and performance of the MFCV ASIC. The bias generator is updated with
the addition of a virtual ground generator. It comprises of a proportional to absolute temperate
circuit (provided by AMS standard cell library) and a buering opamp with an adjustable feedback
to control the virtual ground (common mode voltage -VCM-). During testing of the MFCV and
SDMF ASICs it was observed that the bias electrode coupling to the human body had a big eect
on the stability of VCM. In order to isolate the ground for the biasing electrode, as well as the rest of
the circuits, on chip ground buers are added. Since the overall aim is to create a truly miniaturised
fatigue monitoring system, suitable for a wearable device, the onset detector is implemented in the
ASIC, along with the rest of the sEMG processing circuits. The circuit implementation is shown
in Fig. 6.3. The rectifying opamp is replaced with a comparator and a chopper in order to increase
the accuracy in small amplitude and speed. Since a very low corner frequency is required for the
LPF, a switch-capacitor (SC) lter is chosen, to avoid the use of either very large components or
very small transconductance (gm ) OTAs which are hard to achieve and require calibration. The
LPF corner frequency is accurately controlled with a clock signal. The frequency response of the
third order SC LPF is shown in Fig. 6.4a (fc = 11 Hz, CLK = 4 kHz).
A transient response of the simulated onset detector is shown in Fig. 6.4b. A 100 Hz sine wave
is used to simulate the sEMG signal. After rectication and ltering, an analogue LPF was added
to smooth out the residual spikes from the SC lter. Without the analogue LPF, the comparator
3 3
2.5 2.5
2 2
Amplitude (V)
SDMF (Hz)
1.5 1.5
1 1
0.5 0.5
Start
0 0
0 20 40 60 80 100 120 140 110 111 112 113 114 115 116 117 118 119
Time (s) Time (samples)
(a) (b)
120
0.05
0.04
100
0.03
80 0.02
0.01
SDMF (Hz)
Amplitude (V)
60 0
-0.01
40
-0.02
-0.03
20
-0.04
Start Start
0 -0.05
0 50 100 150 0 20 40 60 80 100 120 140
Time (s) Time (s)
(c) (d)
120 100
Smoothed
SDMF chip
110
iMDF matlab
90
100
80
90
Frequency (Hz)
SDMF (Hz)
80 70
70
60
60
50
50
Start
40 40
0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140
Time (s) Time (s)
(e) (f )
Figure 6.2: (a) sEMg recording of a dynamic contraction, (b) close-up section of dynamic contrac-
tion, (c) SDMF output for the dynamic contraction recording, (d) synthesised dynamic contraction
(blue) and the onset pulse (red), (e) SDMF output for the synthesised dynamic contraction and (f)
SDMF output for the synthesised dynamic contraction with the onset detector on. From (a) it is
clear that the SDMF algorithm is better at detecting the start/end point of the contraction.
VIN 400f 4p
VCM 1x SC-RC
VTHR
VON
3x SC-RC
VF
Figure 6.3: Onset detector circuit implementation for the MFM ASIC.
0
3.5
VON
-50 3
VRECT
2.5
-100
Amplitude (V)
Gain (dB)
2
VFILT VTHR
1.5 VCM
-150
1
-200
0.5 VIN
CLK
-250 0
10-1 100 101 102 103 104 105 0 0.01 0.02 0.03 0.04 0.05
Frequency (Hz) Time (s)
(a) (b)
Figure 6.4: (a) Frequency responce of the fth order SC LPF with CLK = 4 kHz, (b) Transient
responce of the onset detector. The sEMG signal is approximated with a 100 Hz sine wave.
would switch uncontrollably. Finally, the threshold is controlled with a 10-bit digital to analogue
converter (DAC) provided by AMS standard cell libraries. The threshold is externally controlled
through the SPI from a microcontroller. The threshold should can positioned at a level above the
noise oor, as a multiple of its standard deviation.
Although there exist several circuit implementations of real-time, adaptive, noise-tracking
threshold generators [1115], it was decided to adjust the threshold manually. Thus, more exibility
is oered when customising the onset detector operation to a specic dynamic contraction. As a
result, the onset detector can be set to ignore noise as well as small amplitude sEMG contractions
and help distinguish between a weak one and a strong one.
6.4.1 Digital Control
All the operations of the MFM ASIC are controlled by a FSM. The FSM receives instructions from
the microcontroller through the SPI and distributes them across the ASIC's circuits. Similarly, it
sends all the processed data back in the same way. The FSM has four states to govern the ASIC
functions; idle, read, send and test. The state selection is controlled by the rst two bits of the
input SPI instruction and the state diagram can be seen in Fig. 6.5. The ASIC goes to the idle
state as soon as its powered on and reset automatically using a Power-on-Reset (POR) circuit. In
that state the PGA gains are set to unity, the VCM to 1.65 V and the ADC is set to power down
mode. Also, the DAC outputs are set to zero along with all the output registers. Finally, the timer
control sets all the clocks to a predened starting value. The VHDL implementation of the FSM
and the timer controller are shown in Appendix B.
Following the initialisation that occurs in the idle state, the user can either program the ASIC
with a new set of commands or start receiving data. This is selected again with the appropriate SPI
command. Moving on to the read state, the ASIC is expecting the incoming commands and does
not send anything back to the microcontroller. In this state the user can select the desired PGA gain
for both sEMG channels, set the onset detector DACs to a specic value, turn the ADC ON or OFF,
adjust the VCM, set the SDMF and MFCV clocks to the desired frequency and select how both
algorithms respond to the onset pulse. All this information requires a set of ve SPI transactions.
Nevertheless, if the user does not wish to program all the registers, one SPI transaction could be
enough. A detailed input SPI command breakdown is shown in Table 6.2. As mentioned earlier,
the rst two bits are allocated for state selection. The SPI[13:11] bits are used to correctly allocate
11
01
read 00 test
11
01 00
idle
00
01 10
10 00 10
11
send
Figure 6.5: FSM state diagram. The SPI[15:14] bits control the state transitions.
Table 6.2: Input and output SPI control sequences.

SPI BITS: 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
IN 1 state sel in1 timer[10:0]
IN 2 state sel in2 timer[11] Gain Amp 1 Gain Amp 2
IN 3 state sel in3 X DAC1
IN 4 state sel in4 X DAC2
IN 5 state sel in5 ADC on SDMF on MFCV on GND X
OUT 1 IMDF ADC: EMG ch1

OUT 2 MFCV ADC: EMG ch2
OUT 3 Spike count[8:3] ADC: EMG RMS ch1
OUT 4 0 0 0 Spike count[2:0] ADC: EMG RMS ch2
OUT 5 SPI_IN[15:0]
the incoming command to the appropriate registers.
After the ASIC has been successfully programmed its time to start sending data back. Al-
though the ASIC is producing the data, it is congured as the SPI slave. This means that the
microcontroller governs the initiation of the SPI transaction. Hence, the ASIC has a dedicated
output to inform the microcontroller that it is ready to send data. In the case that the ADC is ON,
the interrupt is controlled by the ADC's End of Conversion output (EOC). When EOC goes high,
the ADC has a valid output and the microcontroller reads the results. When the ADC is OFF, the
interrupt is controlled by the individual SDMF and MFCV timing controllers. At the start of each
processing window, both algorithms have a valid result for the previous window. However, both
algorithms do not share the same processing window. Hence, the FSM selects the algorithm with
the shortest window and sends forwards its ready-to-send (RTS) interrupt to the microcontroller.
Contrary to the read state where the user can only send one SPI command to selectively
6.5 Mucle Fatigue Monitoring ASIC Overview 162
Analog Out
10-bit DAC
Bias Generator Analog Buffers
El1
Vref : Setting DC, Timing Onset
close proximity X
to electrodes
Onset Detector 10-bit ADC
El2
SDMF
Epidermis
IA, PGA MFCV Onset

Estimator
El3
C
FSM Gain Control Cyclic
MUX Timing
FSM
C Control
El4
SDMF & MFCV
IA, PGA
SDMF Onset
El5 Estimator
Onset Detector
Onset
16-bit SPI
X
Timing
Ground Generator
Power
& Ground Buffers
10-bit DAC On MFM ASIC
Reset
EOC CLK IN/OUT

C
Figure 6.6: System block architecture of the Muscle Fatigue Monitoring ASIC.
program the ASIC, in the send state all the results are sent sequentially. To achieve that, four
SPI transactions are required, even when the ADC is OFF. The FSM counts the number of SPI
transactions and sends the data in the correct order. The detailed output SPI command breakdown
is shown in Table 6.2. Finally, the last state is the test state, where the input SPI bits are forwarded
to the output SPI registers. Hence, the SPI and FSM operation can be veried prior to using the
MFM ASIC.
6.5 Mucle Fatigue Monitoring ASIC Overview
The developed MFM ASIC (code name Atlas) is shown in Fig. 6.6. The ADC sampling rate is set
to twice the LPF corner frequency, at 1.5k Hz per channel. When the ASIC EOC goes high, the
microcontroller initiates a series of SPI transactions and receives all the data. The chip is designed
to be completely autonomous, with the only external components required being the charging
capacitors for the front-end and the decoupling capacitors for the ground and power supplies. As a
result, a single microcontroller can control several MFM ASICs concurrently. Hence, this is the rst
ASIC completely dedicated to sEMG analysis and muscle fatigue monitoring. It serves as a tool to
visualise and process two sEMG channels in close proximity in order to monitor and understand
the physiological changes the muscles undergo during fatigue.

6.6 Summary 163
Figure 6.7: Die micro-photograph of the MFM ASIC.
A micro-photograph of the MFM ASIC is shown in Fig. 6.7. The chip has two power supply
rails, one dedicated for the analogue and one for the digital circuits. A single pad-ring is used
to oer increased Electrostatic Discharge (ESD) protection. Rail-splitters oer complete isolation
between the two supplies. A separate ground pad has been used to isolate the analogue and digital
circuits, by biasing the ground area between them. Furthermore, the analogue and digital circuits
are arranged in separate areas in an eort to minimise cross-talk and digital noise feeding into the
sensitive analogue circuits. Finally, the digital circuits have been metal-lled to meet the metal and
polysilicon density requirements as this chip was part of a wafer run and each design had to be self
sustained.
6.6 Summary
This chapter presented a sEMG System-on-Chip (SoC) for the monitoring of muscle fatigue. It
combines the MFCV and SDMF algorithms, providing a unique insight in muscle fatigue. The
computation of amplitude parameters such as the MAV of the sEMG amplitude and the spike
concentration oers a complete look into the amplitude and spectrum parameters of the sEMG.
As a result, a fatigue index under the JASA protocol can be established. It features two sEMG
ampliers and an ADC to stream the two channels and MAV values to the user. Moreover, a
sEMG onset detector is included, oering dynamic contraction compatibility and unlocking the
6.6 Summary 164
full potential of the developed technology. It is designed to be standalone and programmed though
standard SPI protocol. Therefore, several MFM SoCs can be combined to work in parallel, centrally
controlled by a microcontroller, thus reducing data transmission costs and bandwidth requirements.
REFERENCES 165
References
[1] R. Conwit, D. Stashuk, B. Tracy, M. McHugh, W. Brown, and E. Metter, The relationship of
motor unit size, ring rate and force, Clinical neurophysiology, vol. 110, no. 7, pp. 12701275,
1999.
[2] H. Milner-Brown, R. B. Stein, and R. Yemm, Changes in ring rate of human motor units
during linearly changing voluntary contractions, The Journal of physiology, vol. 230, no. 2,
p. 371, 1973.
[3] B. Bigland-Ritchie, Emg/force relations and fatigue of human voluntary contractions., Exer-
cise and sport sciences reviews, vol. 9, no. 1, pp. 75118, 1981.
[4] G. Allison, Trunk muscle onset detection technique for emg signals with ecg artefact, Journal
of Electromyography and Kinesiology, vol. 13, no. 3, pp. 209216, 2003.
[5] X. Li and A. Aruin, Muscle activity onset time detection using teager-kaiser energy operator,
in 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference, pp. 75497552,
IEEE, 2005.
[6] M. Magda, Emg onset detectiondevelopment and comparison of algorithms, 2015.
[7] J. Drapaªa, K. Brzostowski, A. Szpala, and A. Rutkowska-Kucharska, Two stage emg onset
detection method, Archives of Control Sciences, vol. 22, no. 4, pp. 427440, 2012.
[8] R. P. Di Fabio, Reliability of computerized surface electromyography for determining the onset
of muscle activity, Physical Therapy, vol. 67, no. 1, pp. 4348, 1987.
[9] S. Solnik, P. Rider, K. Steinweg, P. DeVita, and T. Hortobágyi, Teagerkaiser energy operator
signal conditioning improves emg onset detection, European journal of applied physiology,
vol. 110, no. 3, pp. 489498, 2010.
[10] X. Li, P. Zhou, and A. S. Aruin, Teagerkaiser energy operation of surface emg improves muscle
activity onset detection, Annals of biomedical engineering, vol. 35, no. 9, pp. 15321538, 2007.
[11] H.-L. Chan, M.-A. Lin, T. Wu, S.-T. Lee, Y.-T. Tsai, and P.-K. Chao, Detection of neuronal
spikes using an adaptive threshold based on the maxmin spread sorting method, Journal of
Neuroscience Methods, vol. 172, no. 1, pp. 112121, 2008.
REFERENCES 166
[12] L. Traver, C. Tarin, P. Marti, and N. Cardona, Adaptive-threshold neural spike detection by
noise-envelope tracking, Electronics Letters, vol. 43, no. 24, p. 1, 2007.
[13] J. Holleman, A. Mishra, C. Diorio, and B. Otis, A micro-power neural spike detector and fea-
ture extractor in. 13µm cmos, in 2008 IEEE Custom Integrated Circuits Conference, pp. 333
336, IEEE, 2008.
[14] R. R. Harrison, A low-power integrated circuit for adaptive detection of action potentials in
noisy signals, in Engineering in Medicine and Biology Society, 2003. Proceedings of the 25th
Annual International Conference of the IEEE, vol. 4, pp. 33253328, IEEE, 2003.
[15] D. Kim, M. Stanacevic, R. Kamoua, and Z. Mainen, A low-power low-data-rate neural record-
ing system with adaptive spike detection, in 2008 51st Midwest Symposium on Circuits and
Systems, pp. 822825, IEEE, 2008.
167
Chapter 7
Conclusion
T
HIS thesis explores and develops for the rst time a CMOS-based technology dedicated to
real-time EMG analysis that targets muscle fatigue estimation using bespoke signal pro-
cessing. Various challenges associated with implementing a fatigue monitor in a standard CMOS
technology have also been identied and methods of solution have been proposed. Furthermore,
this work demonstrates the use of such miniaturised systems in the development of wearable muscle
fatigue monitoring devices. As a result, biofeedback facilitated by the use of the developed platforms
can become a reality.
Chapter 2 introduces the muscle physiology, in particular the mechanisms of muscle contrac-
tion and the surface EMG (sEMG) formation. The key sEMG attributes that relay muscle fatigue
information are identied and appropriate methods to detect them are investigated. Finally, a
review of existing approaches for fatigue monitoring paves the way for the approach of this thesis.
7.1 Contributions
Chapter 3 comprises the rst attempt for a low power CMOS system that estimates the instan-
taneous Median Frequency (iMDF) of the sEMG signal, which is considered the golden standard
for muscle fatigue assessment. Two systems are demonstrated; the rst system bypasses the need
for complex digital architectures by employing a fully analogue approach. A second system was
designed to advance the rst, using a mixed signal architecture in order to increase robustness to
process variations and decrease computational complexity and power consumption.
Chapter 4 initially reviews available methods for muscle bre conduction velocity estima-
tion, and concludes that cross-correlation is the most appropriate for implementation in CMOS
technology. This conclusion provides the springboard to develop a novel bit stream approach that
7.2 Recommendations for Future Work 168
greatly simplies the sEMG signal without any loss of information, thus reducing computational
complexity and minimising power consumption. The developed technology was evaluated in a clin-
ical study with a set of healthy individuals, which conrmed the accuracy and eciency of the
developed technology. Finally, the microchip was embedded in a wireless (Bluetooth) node operat-
ing with commercially available wet electrodes, introducing the use of custom Application Specic
Integrated Circuits (ASICs) in wearable electronics for unsupervised muscle fatigue monitoring.
Chapter 5 introduces a new metric similar to iMDF combining the advantages of the bit-
stream approach developed in Chapter 4, showing that it is possible to accurately track the spectral
compression of the sEMG during fatigue with one bit representation. This was achieved by exploiting
a link between zero crossing and median frequency. The developed technology was tested with
human subjects, demonstrating that the validity of the proposed fatigue metric. A second wireless
node was designed to accommodate the new microchip, further reducing the size and weight of the
previously developed node. This was partially achieved by switching to the use of custom-designed,
dry electrodes, oering a more ergonomic design.
Chapter 6 integrates the work presented in Chapters 4 and 5 to deliver the rst System-
on-Chip (SoC) oering complete insight into the underlying mechanisms and physiological changes
during muscle fatigue through sEMG analysis. This is achieved by exploiting both amplitude and
spectrum parameters of the sEMG signal. The new platform was designed to operate under both
static and dynamic contractions to maximise impact in real-life applications. The proposed approach
is scalable, as several muscle fatigue monitoring SoCs can operate in parallel and periodically relay
key information about the muscle, thus reducing data transmission costs and bandwidth require-
ments. Therefore, the technology can be used for multiple concurrent muscle fatigue feedback.
7.2 Recommendations for Future Work
Future developments based on material described in this thesis are proposed in the following areas:
7.2.1 System Optimisation and Enhancement
Having established methods and algorithmic approaches for reliable fatigue assessment, the devel-
oped circuits can be redesigned with the sole purpose of minimising power consumption. In addition,
measurement and characterisation of the SoC presented in Chapter 6 will oer a completed view
on the systems presented in this thesis. This work presented a CMOS implementation of the most
7.2 Recommendations for Future Work 169
appropriate sEMG analysis tools for fatigue evaluation. Nevertheless, establishing a fatigue index,
a metric indicating the fatigue level calibrated to each individual, remains a complicated task yet to
be realised in CMOS. Incorporating this feature in the developed systems would constitute a more
comprehensive tool suitable for real-life situations.
7.2.2 Next Generation Wearable Electronics
There is an increasing trend to develop wearable electronics, however, these lack integrated pro-
cessing and consist of sensing components such as electrodes. In addition, to date biosensors are
regarded as stand-alone measuring devices. However, distributed and coordinated body sensing can
substantially increase the amount of available information and in turn, biofeedback eectiveness.
Thus, integrating multiple smart (autonomous processing), centrally controlled EMG monitoring
nodes in a wearable yet unobtrusive platform (such as a smart clothing), has the potential to revo-
lutionise patient monitoring process and expand lab based testing into everyday life situations. The
technology proposed in this thesis can signicantly advance wearable sensors, as it oers integrated,
unobstructive, multi-node sensing.
7.2.3 Multimodal Biosensors
While monitoring electrophysiological parameters is useful for monitoring general well-being and
specic muscle performance, acquiring a more comprehensive picture of underlying physiology can
be achieved via expanding the biosensing modalities of the wearable device. More specically,
electrochemical measurements can provide important additional insight into the overall health and
peripheral fatigue state through physiologically relevant electrolytes such as potassium and sodium.
A distributed multimodal sensing platform that monitors electrolytes in sweat in coalition with local
EMG muscle performance, would constitute the next generation of wearable biosensors and pave
the way for researchers to embed smart biofeedback mechanisms in our modern way of living. So
far, body chemistries are measured via lab or point-of-care based testing of blood samples, extracted
via needles or ngerprick sensors. These highly inconvenient methods cannot provide a continuous,
real-time feedback. The development of a multi-node, multimodal wearable platform would, for
the rst time, provide a quantiable set of metrics to physiotherapists, sports professionals and
patients, such that they can accurately manage and optimise treatment through feedback and a
powerful insight in human body performance. The technology presented in this thesis has the
potential to be integrated with chemical sensors, creating a high-information content device.

List of publications 170
List of publications
Journal papers
E. Koutsos, V. Cretu,P. Georgiou, A Muscle Fibre Conduction Velocity Tracking ASIC for Local
Fatigue Monitoring, IEEE Transactions on Biomedical Circuits and Systems, vol. 10, no. 6, pp.
1119-1128, 2016
Conference papers
D. Sun, E. Koutsos, P. Georgiou, A neuromorphic based median frequency tracker for muscle
fatigue monitoring, 2016 IEEE Biomedical Circuits and Systems Conference (BioCAS), 2016, (in
proceedings)
M. Douthwaite, N. Moser, E. Koutsos, D. Yates, P. Mitcheson, P. Georgiou, Averaging in an
ISFET array for wearable thermoelectrically powered perspiration analysis, 2016 IEEE Biomedical
Circuits and Systems Conference (BioCAS), 2016, (in proceedings)
D. Sun, E. Koutsos, P. Georgiou, Comparison of sEMG bit-stream modulators for cross-correlation
based muscle fatigue estimation, 2016 IEEE International Symposium on Circuits and Systems
(ISCAS), pp. 838841, 2016
J. Heaey, E. Koutsos, P. Georgiou, Live demonstration: Wearable device for remote EMG and
muscle fatigue monitoring, 2015 IEEE Biomedical Circuits and Systems Conference (BioCAS), pp.
1-5, 2015
E. Koutsos, P. Georgiou, An analogue instantaneous median frequency tracker for EMG fatigue
monitoring, 2014 IEEE International Symposium on Circuits and Systems (ISCAS), pp. 1388-
1391, 2014
171
Appendix A
A.1 VHDL code for neuromorphic iMDF tracker
In order to improve the robustness of the system and to further simplify the design, a digital adder
is used instead of an analogue integrator. The integrated power dierence is used to control the
cut-o frequency of the high-pass lter, thus allowing the lter to select dierent portion of the
spectrum and estimate the median frequency of the input signal. The VHDL implementation of the
controller is shown bellow

1.
1 module Intg (CLKIN, RST, data_in_HPF, data_in_LPF, LFCLK, tap_out, data_reg,
count_reset, rst_period);
2
3 //CLKIN is the period of integration, LFCLK is the rate of tap changing.
4 //Proposed CLKIN = 10kHz, LFCLK = 10Hz.
5
6 input CLKIN;
7 input LFCLK;
8 input RST;
9 input [7:0] rst_period;
10 input [7:0] data_in_HPF;
11 input [7:0] data_in_LPF;
12 output [6:0] tap_out;
13 output [8:0] data_reg;
14 output count_reset;
15
1 The presented implementation was designed by Daiwen Sun under the direct supervision of Ermis Koutsos.
A.1 VHDL code for neuromorphic iMDF tracker 172
16 reg [7:0] rst_count;
17 reg signed [8:0] signed_data_in_HPF;
18 reg signed [8:0] signed_data_in_LPF;
19 reg signed [8:0] data_reg;
20 reg signed [8:0] data_conv;
21 reg [6:0] tap_out;
22 reg count_reset;
23
24 initial begin
25 signed_data_in_LPF = 9'sd0;
26 signed_data_in_HPF = 9'sd0;
27 tap_out = 7'd25;
28 data_reg = 9'sd0;
29 rst_count = 6'b0;
30 end
31
32 always @(posedge CLKIN)
33 begin
34 count_reset = 1'b0;
35 signed_data_in_HPF <= data_in_HPF;
36 signed_data_in_LPF <= data_in_LPF;
37 data_conv <= signed_data_in_HPF − signed_data_in_LPF;
38 //Reset integrator and tap_out if RST is asserted
39 if (RST == 1'b0)
40 begin
41
42 //Lock integrator when saturated
43 if (data_reg > 9'sd127) begin
44 data_reg <= 9'sd127;
45 end
46 else if (data_reg < −9'sd127) begin

47 data_reg <= −9'sd127;
48 end
49 else begin
50 data_reg <= data_reg + data_conv;
51 end
52 end
53
54 if (RST == 1'b1)
55 begin
56 data_reg <= 9'd0;
57 tap_out <= 7'd64;
58 end
59
60 //give reset signal when rst_period passed
61 rst_count <= rst_count + 1'b1;
62 if (rst_count == rst_period)begin
63 count_reset = 1'b1;
64 rst_count <= 6'b0;
65 end
66 end
67
68 //Do tap operations following rising edge of LFCLK
69 always @(posedge LFCLK)
70 begin
71 if (data_reg < 9'sd0 && tap_out > 7'd1)
72 begin
73 tap_out <= tap_out − 7'd1;
74 end
75 if (data_reg >= 9'sd0 && tap_out < 7'd125)
76 begin
77 tap_out <= tap_out + 7'd1;

78 end
79 end
80 endmodule
175
Appendix B
B.1 VHDL code for MFCV ASIC (OPTIMUS)
Listed bellow is the VHDL code for max selection function.
1 LIBRARY ieee;
2 USE IEEE.std_logic_1164.ALL;
3 USE IEEE.numeric_std.ALL;
4 USE work.ALL;
5 ENTITY max_funct IS
6 PORT(
7 --input signals
8 --2 vectors of 14 bit
9 input1, input2 : IN std_logic_vector(13 downto 0);
10 --enable : IN std_logic;
11 --maximum of the 2 input signals
12 output_max : OUT std_logic_vector(13 downto 0);
13 flag : OUT std_logic
14 );
15 END max_funct;
16 ARCHITECTURE comb OF max_funct IS
17 BEGIN
18 bigger: PROCESS(input1, input2)
19 BEGIN
20 if input1 > input2 then
21 output_max <= input1;
22 flag <= ’1’;
23 else
24 output_max <= input2;
25 flag <= ’0’;
26 end if;
27 END PROCESS bigger;
B.1 VHDL code for MFCV ASIC (OPTIMUS) 176
28 END comb;
1 LIBRARY ieee;
4 USE work.ALL;
5 ENTITY max_full IS
6 PORT(
7 --input signals
8 --2 vectors of 14 bit
9 input1, input2, input3, input4, input5, input6, input7, input8, input9, input10 : IN
std_logic_vector(13 downto 0);
10 input11, input12, input13, input14, input15, input16, input17, input18, input19,
input20 : IN std_logic_vector(13 downto 0);
13 --enable : IN std_logic;
14 --maximum of the 2 input signals
15 output : OUT std_logic_vector(39 downto 0)
16 );
17 END max_full;
18 ARCHITECTURE comb OF max_full IS
19 signal f : std_logic_vector(39 downto 0);
20 type output_full_mem is array (0 to 39) of std_logic_vector(13 downto 0);
21 signal output_max : output_full_mem;
22 BEGIN
23 output(39) <= f(38) and f(37) and f(35) and f(30) and f(20) and f(0);
24 output(38) <= f(38) and f(37) and f(35) and f(30) and f(20) and not(f(0));
25 output(37) <= f(38) and f(37) and f(35) and f(30) and not(f(20)) and f(1);
26 output(36) <= f(38) and f(37) and f(35) and f(30) and not(f(20)) and not(f(1));
27 output(35) <= f(38) and f(37) and f(35) and not(f(30)) and f(21) and f(2);
28 output(34) <= f(38) and f(37) and f(35) and not(f(30)) and f(21) and not(f(2));
29 output(33) <= f(38) and f(37) and f(35) and not(f(30)) and not(f(21)) and f(3);
30 output(32) <= f(38) and f(37) and f(35) and not(f(30)) and not(f(21)) and not(f(3));
31 output(31) <= f(38) and f(37) and not(f(35)) and f(31) and f(22) and f(4);
32 output(30) <= f(38) and f(37) and not(f(35)) and f(31) and f(22) and not(f(4));
33 output(29) <= f(38) and f(37) and not(f(35)) and f(31) and not(f(22)) and f(5);
34 output(28) <= f(38) and f(37) and not(f(35)) and f(31) and not(f(22)) and not(f(5));
35 output(27) <= f(38) and f(37) and not(f(35)) and not(f(31)) and f(23) and f(6);
36 output(26) <= f(38) and f(37) and not(f(35)) and not(f(31)) and f(23) and not(f(6));
37 output(25) <= f(38) and f(37) and not(f(35)) and not(f(31)) and not(f(23)) and f(7);
38 output(24) <= f(38) and f(37) and not(f(35)) and not(f(31)) and not(f(23)) and not(f
(7));
39 output(23) <= f(38) and not(f(37)) and f(36) and f(32) and f(24) and f(8);
40 output(22) <= f(38) and not(f(37)) and f(36) and f(32) and f(24) and not(f(8));
41 output(21) <= f(38) and not(f(37)) and f(36) and f(32) and not(f(24)) and f(9);
42 output(20) <= f(38) and not(f(37)) and f(36) and f(32) and not(f(24)) and not(f(9));
43 output(19) <= f(38) and not(f(37)) and f(36) and not(f(32)) and f(25) and f(10);
44 output(18) <= f(38) and not(f(37)) and f(36) and not(f(32)) and f(25) and not(f(10));
45 output(17) <= f(38) and not(f(37)) and f(36) and not(f(32)) and not(f(25)) and f(11);
46 output(16) <= f(38) and not(f(37)) and f(36) and not(f(32)) and not(f(25)) and not(f
(11));
47 output(15) <= f(38) and not(f(37)) and not(f(36)) and f(33) and f(26) and f(12);
48 output(14) <= f(38) and not(f(37)) and not(f(36)) and f(33) and f(26) and not(f(12));
49 output(13) <= f(38) and not(f(37)) and not(f(36)) and f(33) and not(f(26)) and f(13);
50 output(12) <= f(38) and not(f(37)) and not(f(36)) and f(33) and not(f(26)) and not(f
(13));
51 output(11) <= f(38) and not(f(37)) and not(f(36)) and not(f(33)) and f(27) and f(14);
52 output(10) <= f(38) and not(f(37)) and not(f(36)) and not(f(33)) and f(27) and not(f
(14));
53 output(9) <= f(38) and not(f(37)) and not(f(36)) and not(f(33)) and not(f(27)) and f
(15);
54 output(8) <= f(38) and not(f(37)) and not(f(36)) and not(f(33)) and not(f(27)) and
not(f(15));
55 output(7) <= not(f(38)) and f(34) and f(28) and f(16);
56 output(6) <= not(f(38)) and f(34) and f(28) and not(f(16));
57 output(5) <= not(f(38)) and f(34) and not(f(28)) and f(17);
58 output(4) <= not(f(38)) and f(34) and not(f(28)) and not(f(17));
59 output(3) <= not(f(38)) and not(f(34)) and f(29) and f(18);
60 output(2) <= not(f(38)) and not(f(34)) and f(29) and not(f(18));
61 output(1) <= not(f(38)) and not(f(34)) and not(f(29)) and f(19);
62 output(0) <= not(f(38)) and not(f(34)) and not(f(29)) and not(f(19));
63 A1: entity max_funct port map( input1 => input1, input2 => input2, output_max =>
output_max(0), flag => f(0) );
83 B1: entity max_funct port map( input1 => output_max(0), input2 => output_max(1),
output_max => output_max(20), flag => f(20) );
B.2 VHDL code for SDMF ASIC (BATMAN) 179
93 C1: entity max_funct port map( input1 => output_max(20), input2 => output_max(21),
98 D1: entity max_funct port map( input1 => output_max(30), input2 => output_max(31),
99 D2: entity max_funct port map( input1 => output_max(32), input2 => output_max(33),
100 E1: entity max_funct port map( input1 => output_max(35), input2 => output_max(36),
101 F1: entity max_funct port map( input1 => output_max(37), input2 => output_max(34),
102 END comb;
B.2 VHDL code for SDMF ASIC (BATMAN)
Listed bellow is the VHDL code for min dierence selection function.
1 LIBRARY ieee;
4 USE work.ALL;
5 ENTITY diff_funct IS
6 PORT(
7 input, ca : IN std_logic_vector(8 downto 0);
8 diff : OUT std_logic_vector(8 downto 0)
9 );
10 END diff_funct;
11 ARCHITECTURE comb OF diff_funct IS
12 BEGIN
13 bigger: PROCESS(input,ca)
14 BEGIN
15 diff <= std_logic_vector(
16 resize(
17 unsigned(
18 abs(
19 signed(resize(unsigned(input), 10))
20 -
21 (signed(resize(unsigned(ca), 10)) / 2)
22 )
23 ),9)
24 );
25 END PROCESS bigger;
26 END comb;
1 LIBRARY ieee;
4 USE work.ALL;
5 ENTITY min_funct IS
6 PORT(
7 input1, input2 : IN std_logic_vector(8 downto 0);
8 output_min : OUT std_logic_vector(8 downto 0);
9 flag : OUT std_logic
10
11 );
12 END min_funct;
13 ARCHITECTURE comb OF min_funct IS
14 BEGIN
15 bigger: PROCESS(input1, input2)
16 BEGIN
17 if input1 > input2 then
18 output_min <= input2;
19 flag <= ’0’;
20 else
21 output_min <= input1;
22 flag <= ’1’;
23 end if;
24 END process bigger;
25 END comb;
1 LIBRARY ieee;
4 USE work.ALL;
5 ENTITY min_full IS
6 PORT(
7 input1, input2, input3, input4, input5, input6, input7, input8, input9, input10 : IN
std_logic_vector(8 downto 0);
11 ca : IN std_logic_vector(8 downto 0);
12 output_top : OUT std_logic_vector(39 downto 0)
13 );
14 END min_full;
15 ARCHITECTURE comb OF min_full IS
16 signal f : std_logic_vector(39 downto 0);
17 type output_full_mem is array (0 to 39) of std_logic_vector(8 downto 0);
18 signal output_min : output_full_mem;
19 signal diff : output_full_mem;
20 BEGIN
21 output_top(39) <= not(f(38)) and not(f(34)) and not(f(29)) and not(f(19));
22 output_top(38) <= not(f(38)) and not(f(34)) and not(f(29)) and f(19);
23 output_top(37) <= not(f(38)) and not(f(34)) and f(29) and not(f(18));
24 output_top(36) <= not(f(38)) and not(f(34)) and f(29) and f(18);
25 output_top(35) <= not(f(38)) and f(34) and not(f(28)) and not(f(17));
26 output_top(34) <= not(f(38)) and f(34) and not(f(28)) and f(17);
27 output_top(33) <= not(f(38)) and f(34) and f(28) and not(f(16));
28 output_top(32) <= not(f(38)) and f(34) and f(28) and f(16);
29 output_top(31) <= f(38) and not(f(37)) and not(f(36)) and not(f(33)) and not(f(27))
and not(f(15));
30 output_top(30) <= f(38) and not(f(37)) and not(f(36)) and not(f(33)) and not(f(27))
and f(15);
31 output_top(29) <= f(38) and not(f(37)) and not(f(36)) and not(f(33)) and f(27) and
not(f(14));
32 output_top(28) <= f(38) and not(f(37)) and not(f(36)) and not(f(33)) and f(27) and f
(14);
33 output_top(27) <= f(38) and not(f(37)) and not(f(36)) and f(33) and not(f(26)) and
not(f(13));
34 output_top(26) <= f(38) and not(f(37)) and not(f(36)) and f(33) and not(f(26)) and f
(13);
35 output_top(25) <= f(38) and not(f(37)) and not(f(36)) and f(33) and f(26) and not(f
(12));
36 output_top(24) <= f(38) and not(f(37)) and not(f(36)) and f(33) and f(26) and f(12);
37 output_top(23) <= f(38) and not(f(37)) and f(36) and not(f(32)) and not(f(25)) and
not(f(11));
38 output_top(22) <= f(38) and not(f(37)) and f(36) and not(f(32)) and not(f(25)) and f
(11);
39 output_top(21) <= f(38) and not(f(37)) and f(36) and not(f(32)) and f(25) and not(f
(10));
40 output_top(20) <= f(38) and not(f(37)) and f(36) and not(f(32)) and f(25) and f(10);
41 output_top(19) <= f(38) and not(f(37)) and f(36) and f(32) and not(f(24)) and not(f
(9));
42 output_top(18) <= f(38) and not(f(37)) and f(36) and f(32) and not(f(24)) and f(9);
43 output_top(17) <= f(38) and not(f(37)) and f(36) and f(32) and f(24) and not(f(8));
44 output_top(16) <= f(38) and not(f(37)) and f(36) and f(32) and f(24) and f(8);
45 output_top(15) <= f(38) and f(37) and not(f(35)) and not(f(31)) and not(f(23)) and
not(f(7));
46 output_top(14) <= f(38) and f(37) and not(f(35)) and not(f(31)) and not(f(23)) and f
(7);
47 output_top(13) <= f(38) and f(37) and not(f(35)) and not(f(31)) and f(23) and not(f
(6));
48 output_top(12) <= f(38) and f(37) and not(f(35)) and not(f(31)) and f(23) and f(6);
49 output_top(11) <= f(38) and f(37) and not(f(35)) and f(31) and not(f(22)) and not(f
(5));
50 output_top(10) <= f(38) and f(37) and not(f(35)) and f(31) and not(f(22)) and f(5);
51 output_top(9) <= f(38) and f(37) and not(f(35)) and f(31) and f(22) and not(f(4));
52 output_top(8) <= f(38) and f(37) and not(f(35)) and f(31) and f(22) and f(4);
53 output_top(7) <= f(38) and f(37) and f(35) and not(f(30)) and not(f(21)) and not(f(3)
);
54 output_top(6) <= f(38) and f(37) and f(35) and not(f(30)) and not(f(21)) and f(3);
55 output_top(5) <= f(38) and f(37) and f(35) and not(f(30)) and f(21) and not(f(2));
56 output_top(4) <= f(38) and f(37) and f(35) and not(f(30)) and f(21) and f(2);
57 output_top(3) <= f(38) and f(37) and f(35) and f(30) and not(f(20)) and not(f(1));
58 output_top(2) <= f(38) and f(37) and f(35) and f(30) and not(f(20)) and f(1);
59 output_top(1) <= f(38) and f(37) and f(35) and f(30) and f(20) and not(f(0));
60 output_top(0) <= f(38) and f(37) and f(35) and f(30) and f(20) and f(0);
61 X1: entity diff_funct port map( input => input1, ca => ca, diff => diff(0) );
101 A1: entity min_funct port map( input1 => diff(0), input2 => diff(1), output_min =>
output_min(0), flag => f(0) );
121 B1: entity min_funct port map( input1 => output_min(0), input2 => output_min(1),
output_min => output_min(20), flag => f(20) );
B.3 VHDL code for MFM ASIC (ATLAS) 185
131 C1: entity min_funct port map( input1 => output_min(20), input2 => output_min(21),
136 D1: entity min_funct port map( input1 => output_min(30), input2 => output_min(31),
137 D2: entity min_funct port map( input1 => output_min(32), input2 => output_min(33),
138 E1: entity min_funct port map( input1 => output_min(35), input2 => output_min(36),
139 F1: entity min_funct port map( input1 => output_min(37), input2 => output_min(34),
140 END comb;
B.3 VHDL code for MFM ASIC (ATLAS)
Listed bellow is the VHDL code for the Finite State Machine.
1 library IEEE;
2 USE IEEE.STD_LOGIC_1164.ALL;
3 USE IEEE.NUMERIC_STD.ALL;
4 --after read only, wait for EOC to change state to send only, so that the reset_adc
is correct. also, send 2 send only commands to get the 1st result, and then 1 per
EOC
5 -- Address is a 16-bit input coming from the SPI
6 ------------------SSXXXXXXXXXXXXXX------------------
7 -- SS: Select Operation
8 ---- 00: idle
9 ---- 01: read_only
10 ---- 10: send_only
11 ---- 11: test
12 entity safe_state is
13 port(
14 clk: in std_logic;
15 EOC: in std_logic;
16 READY: in std_logic;
17 CS: in std_logic;
18 data_from_spi: in std_logic_vector(15 downto 0);
19 reset: in std_logic;
20 ADC: in std_logic_vector(9 downto 0);
21 MFCV: in std_logic_vector(5 downto 0);
22 IMDF: in std_logic_vector(5 downto 0);
23 N: in std_logic_vector(8 downto 0);
24 G1: out std_logic_vector(4 downto 0);
25 G2: out std_logic_vector(4 downto 0);
26 PD: out std_logic;
27 AGND: out std_logic_vector(2 downto 0);
28 SEL_IMDF: out std_logic_vector(1 downto 0);
29 SEL_MFCV: out std_logic_vector(3 downto 0);
30 CLOCK_SEL: out std_logic_vector(11 downto 0);
31 DAC1: out std_logic_vector(9 downto 0);
32 DAC2: out std_logic_vector(9 downto 0);
33 INTERRUPT: out std_logic;
34 RESET_ADC: out std_logic;
35 data_to_spi: out std_logic_vector(15 downto 0)
36 );
37 end entity;
38 architecture rtl of safe_state is
39 type state_type is (idle, read_only, send_only, test);
40 -- Register to hold the current state
41 signal state,nstate : state_type;
42 signal SEND : std_logic;
43 signal PD_in : std_logic;
44 signal EOC_in,TRIGGER : std_logic;

45 signal CS_in, CS_in2, CS_inin, CS_ininin : std_logic;
46 signal data_from_spi_in, data_from_spi_in2: std_logic_vector(13 downto 0);
47 signal TEST_COUNTER: std_logic_vector(1 downto 0);
48 begin
49 -- Logic to advance to the next state
50 process (clk, reset)
51 begin
52 if reset = ’0’ then
53 state <= idle;
54 nstate <= idle;
55 elsif (clk’EVENT and clk=’1’) then
56 state<=nstate;
57 case state is
58 when idle =>
59 RESET_ADC<= ’0’;
60 if ((data_from_spi(15) = ’0’) and (data_from_spi(14) = ’1’)) then
61 nstate <= read_only;
62 elsif ((data_from_spi(15) = ’1’) and (data_from_spi(14) = ’0’)) then
63 nstate <= send_only;
65 nstate <= test;
66 end if;
67 when read_only=>
69 nstate <= idle;
72 RESET_ADC<=’1’;
74 nstate <= test;
75 end if;
76 when send_only=>
78 nstate <= idle;
82 nstate <= test;
83 end if;
84 when test=>
86 nstate <= idle;

91 end if;
92 when others=> NULL;
93 end case;
94 end if;
95 end process;
96 -- Logic to determine output
97 process (clk)
98 begin
99 if (clk’EVENT and clk=’1’) then
100 EOC_in<= EOC;
101 CS_in<= CS;
102 CS_inin<= CS_in;
103 CS_ininin<= CS_inin;
104 data_from_spi_in2<= data_from_spi(13 downto 0);
105 data_from_spi_in<= data_from_spi_in2;
106 case state is
107 when idle =>
108 G1<= "11110";
109 G2<= "11110";
110 PD_in<= ’1’;
111 AGND<= (OTHERS=>’0’);
112 SEL_IMDF<= (OTHERS=>’0’);
113 SEL_MFCV<= (OTHERS=>’0’);
114 CLOCK_SEL<= (OTHERS=>’0’);
115 DAC1<= (OTHERS=>’0’);
116 DAC2<= (OTHERS=>’0’);
117 when read_only =>
118 if(data_from_spi_in(13 downto 11) = "000") THEN
119 CLOCK_SEL(10 downto 0) <= data_from_spi_in(10 downto 0);
120 elsif(data_from_spi_in(13 downto 11) = "001") THEN
121 G1 <= data_from_spi_in(9 downto 5);
122 G2 <= data_from_spi_in(4 downto 0);
123 CLOCK_SEL(11) <= data_from_spi_in(10);
125 DAC1<= data_from_spi_in(9 downto 0);
127 DAC2<= data_from_spi_in(9 downto 0);

129 PD_in<= data_from_spi_in(10);
130 SEL_IMDF<= data_from_spi_in(9 downto 8);
131 SEL_MFCV<= data_from_spi_in(7 downto 4);
132 AGND<= data_from_spi_in(3 downto 1);
133 end if;
135 end case;
136 end if;
137 end process;
138 process(clk, PD_in, EOC, READY ,EOC_in, CS_in2)
139 VARIABLE COUNTER1 : STD_LOGIC_VECTOR(1 DOWNTO 0);
140 begin
141 if PD_in = ’0’ then
142 SEND <= EOC;
143 else
144 SEND <= READY;
145 end if;
146 if (clk’EVENT and clk=’1’) then
147 case state is
148 when idle =>
149 data_to_spi<= (OTHERS=>’0’);
150 COUNTER1:= (OTHERS=>’0’);
151 when send_only =>
152 --if rising_edge(TRIGGER) then
153 if (CS_inin = ’1’ and CS_ininin = ’0’) then
154 data_to_spi(9 downto 0) <= ADC;
155 IF(COUNTER1 = "00") THEN
156 data_to_spi(15 downto 10) <= IMDF;
157 COUNTER1 := std_logic_vector(unsigned(COUNTER1) + "01");
158 ELSIF (COUNTER1 = "01") THEN
159 data_to_spi(15 downto 10) <= MFCV;
162 data_to_spi(15 downto 10) <= N(8 downto 3);
165 data_to_spi(12 downto 10) <= N(2 downto 0);
166 data_to_spi(15 downto 13) <= "000";
167 COUNTER1 := (OTHERS=>’0’);
168 end if;
169 end if;
170 when test =>

171 data_to_spi <= data_from_spi;
173 end case;
174 end if;
175 TEST_COUNTER<= COUNTER1;
176 end process;
177 process (CS)
178 begin
179 CS_in2<= not(CS);
180 end process;
181 PD <= PD_in;
182 INTERRUPT<=SEND;
183 end rtl;
Listed bellow is the VHDL code for the timing control unit.
1 library IEEE;
2 USE IEEE.STD_LOGIC_1164.ALL;
3 USE IEEE.NUMERIC_STD.ALL;
4 ENTITY ATLAS_timming is
5 PORT (
6 MASTER_CLOCK : IN STD_LOGIC;
7 RESET : IN STD_LOGIC;
8 SEL : IN std_logic_vector(11 downto 0);
9 CLK_1 : OUT STD_LOGIC;
13 CLK_5 : OUT STD_LOGIC);
14 END ATLAS_timming;
15 ARCHITECTURE Behavioral OF ATLAS_timming IS
16 -- (Fm/F)/2 = 1M/2/f = 500K/f
17 signal temporal1: STD_LOGIC;
22 signal COUNTERHOLD1 : STD_LOGIC_VECTOR(3 DOWNTO 0);--ADC
23 signal COUNTERHOLD2 : STD_LOGIC_VECTOR(19 DOWNTO 0);--LF MFCV
24 signal COUNTERHOLD3 : STD_LOGIC_VECTOR(7 DOWNTO 0);--HF MFCV
25 signal COUNTERHOLD4 : STD_LOGIC_VECTOR(19 DOWNTO 0);--LF IMDF
26 signal COUNTERHOLD5 : STD_LOGIC_VECTOR(7 DOWNTO 0);--HF IMDF
27 CONSTANT COUNTERMAX1 : STD_LOGIC_VECTOR(3 DOWNTO 0):= "1000";--%66KHz clock

generation ADC
generation HF
generation
generation
generation
generation
generation
generation
35 CONSTANT COUNTERMAX9 : STD_LOGIC_VECTOR(19 DOWNTO 0):= "00011000011010100000";--%5Hz
clock generation LF
36 CONSTANT COUNTERMAX10 : STD_LOGIC_VECTOR(19 DOWNTO 0):= "00011110100001001000";--%4
Hz clock generation
Hz clock generation
Hz clock generation
Hz clock generation
40 CONSTANT COUNTERMAX14 : STD_LOGIC_VECTOR(19 DOWNTO 0):= "11110100001001000000";--
%0.5Hz clock generation
41 BEGIN
42 PROCESS(RESET, MASTER_CLOCK)
43 VARIABLE COUNTER1 : STD_LOGIC_VECTOR(3 DOWNTO 0);--ADC
44 VARIABLE COUNTER2 : STD_LOGIC_VECTOR(19 DOWNTO 0);--LF MFCV
45 VARIABLE COUNTER3 : STD_LOGIC_VECTOR(7 DOWNTO 0);--HF MFCV
46 VARIABLE COUNTER4 : STD_LOGIC_VECTOR(19 DOWNTO 0);--LF IMDF
47 VARIABLE COUNTER5 : STD_LOGIC_VECTOR(7 DOWNTO 0);--HF IMDF
48 BEGIN
49 IF (RESET = ’0’) THEN
50 temporal1 <= ’0’;
60 ELSIF (MASTER_CLOCK’EVENT and MASTER_CLOCK=’1’) THEN --1MHz MASTER clock
61 IF(COUNTER1 = "0000") THEN
62 temporal1 <= NOT(temporal1);
63 COUNTER1 := COUNTERHOLD1;
64 ELSE
65 COUNTER1 := std_logic_vector(unsigned(COUNTER1) - "0001");
66 END IF;
67 IF(COUNTER2 = "00000000000000000000") THEN
70 ELSE
72 END IF;
73 IF(COUNTER3 = "00000000") THEN
76 ELSE
78 END IF;
79 IF(COUNTER4 = "00000000000000000000") THEN
82 ELSE
84 END IF;
85 IF(COUNTER5 = "00000000") THEN
88 ELSE
90 END IF;
91 END IF;
92 CLK_1 <= temporal1;
97 END PROCESS;
98 PROCESS(SEL)
99 BEGIN
100 COUNTERHOLD1 <= "1000";
101 COUNTERHOLD2 <= "00000000000000000000";
102 COUNTERHOLD3 <= "00000000";
103 COUNTERHOLD4 <= "00000000000000000000";
104 COUNTERHOLD5 <= "00000000";
105 case SEL(11 DOWNTO 9) is --LF MFCV
106 when "000" => COUNTERHOLD2 <= COUNTERMAX9;
112 when others => Null;
113 end case;
114 case SEL(8 DOWNTO 6) is --HF MFCV
123 end case;
124 case SEL(5 DOWNTO 3) is --LF IMDF
132 end case;
133 case SEL(2 DOWNTO 0) is --HF IMDF

142 end case;
143 END PROCESS;
144 END Behavioral;
195
Appendix C
C.1 Bit-Stream Cross-Correlator Modelling
Listed bellow is the MATLAB

® code for the selection of optimum window size and sampling
frequency for the bit-stream cross-correlator.
1 load results1 . mat
2 f_samp = 4000;
3 ch1aa = detrend(ch1(1.3e4−16:2.8e5−16));
4 ch2aa = detrend(ch2(1.3e4:2.8e5));
5
6 for l = 1:8%undersample and ovesample from 1K to 8K equivalent sampling rate. Also simulate
window size from 0.1sec to 1.5sec in 15 steps of 0.1
7 ch1a = resample(ch1aa,l,4);
8 ch2a = resample(ch2aa,l,4);
9 % %inserting linear delay into 1 channel and correlating with other
10 % %insert 10 msec delay in samples
11 insert_delay = 0.01*(f_samp*l/4);
12 step = oor(length(ch1aa)/insert_delay);
13 del_array(1:insert_delay) = 0;
14 out1 = 0;
15 for i =1:insert_delay
16 temp = [ch1aa((i*step)−step+1:i*step) ch1aa(i*step)];
17 out1 = [out1 temp];
18 end
C.1 Bit-Stream Cross-Correlator Modelling 196
19 out2 = [ch2aa del_array];
20 reference = 0.00;
21 out1b = zeros(1, length(out1)) ;
22 out2b = zeros(1, length(out2)) ;
23 for i =1:length(out1b)
24 if out1(i ) >= reference
25 out1b(i) = 1;
26 else
27 out1b(i) = 0;
28 end
29 if out2(i ) >= reference
30 out2b(i) = 1;
31 else
32 out2b(i) = 0;
33 end
34 end
35
36 for k =(0.1*f_samp*l/4):(0.1*f_samp*l/4):(1.5*f_samp*l/4)
37 window = k;
38 for i =1:oor(length(out1)/window)
39 r_emg = xcorr(out1(window*i−(window−1):window*i),out2(window*i−(window−1)
:window*i));
40 r_square = xcorr(out1b(window*i−(window−1):window*i),out2b(window*i−(
window−1):window*i));
41 [ v_emg(i),p_emg(i)] = max(r_emg);
42 [ v_square(i),p_square(i)] = max(r_square);
43 end
44 Delay_emg = (p_emg−window)/(f_samp*l/4);
45 Delay_Square = (p_square−window)/(f_samp*l/4);
46 DEL_XCORREMG{k/(0.1*f_samp*l/4),l} = Delay_emg;
47 DEL_XCORRSQUARE{k/(0.1*f_samp*l/4),l} = Delay_Square;
48 clear v_square p_emg p_square v_emg Delay_emg Delay_square
49 ch1b = out2b;
50 ch2b = out1b;
51 step = k
52 i =1;
53 t=1;
54 x=1;
55 length_samples = length(ch2b);
56 while i <=(length_samples−step)
57 temp1 = ch2b(i:i+step−1);
58 temp2 = ch1b(i:i+step−1);
59 size = length(temp1);
60 temp1 = [zeros(1,length(temp1)) temp1 zeros(1,length(temp1))];
61 for j =1:size*2+1
62 res = xor(temp1(size*2−(j−2) : size*3−j+1) , temp2);
63 out_temp(j) = sum(res(:) == 0);
64 t=t+1;
65 clear res
66 end
67 [~, index] = max(out_temp);
68 out(x) = step − index;
69 t=1;
70 x=x+1;
71 i =i+step;
72 clear index out_temp temp1 temp2 size
73 end
74 DEL_BITBASEN{k/(0.1*f_samp*l/4),l} = out./(f_samp*l/4);
75 clear out
76 end
77 clear ch2b ch2a ch1b ch1a out1 out2 out1b out2b
78 end
79
80 for l = 1:8
81 for i = 1:15
82 len1 = length(DEL_BITBASEN{i,l});
83 len2 = length(DEL_XCORREMG{i,l});
84 len3 = length(DEL_XCORRSQUARE{i,l});
85 y1 = 0:0.0095/len1:0.0095−0.0095/len1;
86 y2 = 0:0.0095/len2:0.0095−0.0095/len2;
87 y3 = 0:0.0095/len3:0.0095−0.0095/len3;
88 err_bitbasen(i, l ) = immse(y1,DEL_BITBASEN{i,l});
89 err_xcorremg(i,l) = immse(y2,DEL_XCORREMG{i,l});
90 err_xcorrsquare(i, l ) = immse(y3,DEL_XCORRSQUARE{i,l});
91 di1 = (y1−DEL_BITBASEN{i,l});
92 ratio1 = di1 ./DEL_BITBASEN{i,l};
93 ratio1 (isnan( ratio1 ))=0;
94 ratio1 ( isinf ( ratio1 ))=0;
95 MARD_bitbasen(i,l) = mean(abs(ratio1)).*100;
96 di2 = (y2−DEL_XCORREMG{i,l});
97 ratio2 = di1 ./DEL_XCORREMG{i,l};
100 MARD_xcorremg(i,l) = mean(abs(ratio2)).*100;
101 di3 = (y3−DEL_XCORRSQUARE{i,l});
102 ratio3 = di1 ./DEL_XCORRSQUARE{i,l};
105 MARD_xcorrsquare(i,l) = mean(abs(ratio3)).*100;
106 end
107 end
C.2 Noise Analysis 199
C.2 Noise Analysis
Listed bellow is the MATLAB

® code for the noise analysis.
1 data1 = detrend(data);
2 data2 = data1/100;% divide by di−gain in V/V
3 Fs = 100000;% dene sampling frequency in labchart/ADC
4 time = (1/100000):(1/100000):(length(data2)/100000);
5 LTime = length(time);
6 [ VOF_POWER,FREQ] = periodogram(data2,rectwin(LTime),LTime,Fs);
7 % if amplitude density labchart is used, skip periodogram and do:
8 % loglog(FrequencyHz,(Channel1VvHz))
9 gure
10 loglog (FREQ,sqrt(VOF_POWER), 'b')
11 grid on
12 % Caluculation of the integrated input refered noise
13 sum = 0;
14 Freq100m = nd(FREQ>0.1,1);
15 Unitfreq = nd(FREQ>2500,1);
16 for i =Freq100m:Unitfreq−1
17 sum=sum+(FREQ(i+1)−FREQ(i))*2*abs(VOF_POWER(i));
18 % if amplitude amplitude density labchart is used:
19 % sum=sum+(FrequencyHz(i+1)−FrequencyHz(i))*2*abs(Channel1VvHz(i)^2);
20 end
21 (sum^0.5)
22 % Plot single −sided amplitude spectrum.
23 title ( ' Single −Sided Amplitude Spectrum preamp')
24 xlabel ( ' Frequency (Hz)')
25 ylabel ( ' Power spectrum (V/sqrt(Hz))')
26 % to compare lab with cadence, load noise.csv and plot together :
27 % step1 = length(VNINX)/length(FREQ);
28 % newFreq = interp1((1:length(VNINX)), (VNINX),1:step1:length(VNINX));
29 % newMagq = interp1((1:length(VNINY)), (VNINY),1:step1:length(VNINY));

C.2 Noise Analysis 200
30 % loglog(FREQ(1:length(newFreq)),sqrt(VOF_POWER(1:length(newFreq))), 'r')
31 % hold on
32 % loglog(newFreq,newMagq, 'k')

Koutsos E 2017 PHD Thesis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Koutsos E 2017 PHD Thesis

Uploaded by

Copyright:

Available Formats

real time semg based muscle fatigue

monitoring using low power integrated

January 27, 2017

A thesis submitted for the degree of Doctor of Philosophy of Imperial

Department of Electrical and Electronic Engineering

edged in the text.'

Attribution Non-Commercial No Derivatives licence. Researchers are free to copy, distribute or

for his CAD contributions.

my vision of this project.

and all the good times we enjoyed.

encouraging me to be myself was the biggest help of all.

biofeedback parameters of EMG in rehabilitation, ergonomics and training, by using measured

technology facilitates localised real-time processing to achieve complete miniaturisation, resulting

primary design considerations.

1.2 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Thesis Organisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.2.1 Muscle contraction and EMG formation . . . . . . . . . . . . . . . . . . . . . 8

2.2.2 EMG detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.2.3 Standardised sEMG Detection Protocol . . . . . . . . . . . . . . . . . . . . . 14

2.3 Myoelectric Manifestations of Muscle Fatigue . . . . . . . . . . . . . . . . . . . . . . 17

2.3.1 Processing in Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.3.1.1 Root Mean Square, Spike Counting, Zero Crossing Rate . . . . . . . 19

2.3.1.2 Muscle Fibre Conduction Velocity . . . . . . . . . . . . . . . . . . . 20

2.3.2 Processing in Time-Frequency Domain . . . . . . . . . . . . . . . . . . . . . . 21

2.3.2.1 Short Time Fourier Transform . . . . . . . . . . . . . . . . . . . . . 22

2.3.2.2 Continuous Wavelet Transform . . . . . . . . . . . . . . . . . . . . . 22

2.3.2.3 Other Time-Frequency Methods . . . . . . . . . . . . . . . . . . . . 23

2.3.3 Fatigue Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

2.3.4 Repeatability of sEMG Muscle Fatigue Estimates . . . . . . . . . . . . . . . . 26

2.4 Current Approaches for Muscle Fatigue Monitoring . . . . . . . . . . . . . . . . . . . 26

2.4.1 Approach in this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.2 Methods for real-time iMDF tracking systems . . . . . . . . . . . . . . . . . . . . . . 45

3.3 Proposed IC iMDF Tracking System . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.3.0.1 iMDF Model Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 51

3.3.0.2 iMDF Model Simulations . . . . . . . . . . . . . . . . . . . . . . . . 53

3.4 iMDF Tracker Circuit Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3.5 Measurement Results and Reections on iMDF Tracker . . . . . . . . . . . . . . . . 65

3.5.1 Design Considerations for Future Improvements . . . . . . . . . . . . . . . . . 65

3.5.2 A novel Neuromorphic based iMDF Tracking System . . . . . . . . . . . . . . 67

4.2 Methods for Conduction Velocity Estimation . . . . . . . . . . . . . . . . . . . . . . 77

4.3 MFCV Tracking Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

4.3.1 Design Considerations for the CMOS Cross-Correlator . . . . . . . . . . . . . 83

4.3.2 Proposed Bit-Stream Cross-Correlator . . . . . . . . . . . . . . . . . . . . . . 88

4.4 MFCV Tracking ASIC Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

4.4.1 sEMG Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

4.5 MFCV Tracking ASIC Experimental Results . . . . . . . . . . . . . . . . . . . . . . . 97

4.5.1 System Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

4.5.2 Validation on Human Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . 101

4.5.2.1 Experimental Protocol . . . . . . . . . . . . . . . . . . . . . . . . . . 101

4.5.2.2 Results from Human Trials . . . . . . . . . . . . . . . . . . . . . . . 103

4.5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

4.6 Wearable Muscle Fatigue Monitor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

5.2 Median Frequency and sEMG Spike Duration . . . . . . . . . . . . . . . . . . . . . . 118

5.2.1 Algorithm for Spike Duration Based iMDF Approximation . . . . . . . . . . . 123

5.3 SDMF Tracking ASIC Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

5.3.1 sEMG Front End Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . 129

5.3.1.1 Choice of Electrodes . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

5.3.1.2 Analogue Front End . . . . . . . . . . . . . . . . . . . . . . . . . . . 131

5.3.1.3 sEMG Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

5.4 SDMF Tracking ASIC Experimental Results . . . . . . . . . . . . . . . . . . . . . . . 134

5.4.1 System Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

5.4.2 Physiological Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

3.5 Measurement Results and Reections on iMDF Tracker . . . . . . . . . . . . . . . . 65

thus increasing the insigh and eectiveness of biofeedback in the future. . . . . . . . . . . . . . 3

cycle of the muscle begins, based on the sliding lament model. . . . . . . . . . . . . . . . . . 11

fV CO . The output is low-pass ltered for a smoother response. . . . . . . . . . . . . . . . . . . 50

3.12 Harrison-amplier architecture [7] for a variable gain, subtracting amplier. . . . . . . . . . . . . 57

3.13 Dierential in/out switched-capacitor integrator with oset compensation. . . . . . . . . . . . . 58

3.15 FLF eliptic lter architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

3.17 OTA used in biquad lter with source degeneration. . . . . . . . . . . . . . . . . . . . . . . . 60

Thus, the eect of thresshold can be studied. . . . . . . . . . . . . . . . . . . . . . . . . . . 101

5.14 Second order, low pass lter, gm-C architecture. . . . . . . . . . . . . . . . . . . . . . . . . . 133

were detected and shown in detail in the left gure. . . . . . . . . . . . . . . . . . . . . . . . 138