Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
by
Oljeta Bida
A Thesis
Submitted to the Faculty
of the
WORCESTER POLYTECHNIC INSTITUTE
in partial fulfillment of the requirements for the
Degree of Master of Science
in
Electrical Engineering
January 2005
________________________________________
Professor Edward A. Clancy, Thesis Advisor
________________________________________
Professor Donald R. Brown, Comittee Member
________________________________________
Professor David Cyganski, Comittee Member
© 2005 OLJETA BIDA
ALL RIGHTS RESERVED
To my dearest family,
for their selfless sacrifices,
and for the continuous belief in my capabilities
regardless of my weak moments and the obstacles
that we have gone through.
iii
ABSTRACT
A number of studies have investigated the relationship between surface
electromyogram (EMG) and torque exerted about a joint. The standard deviation of the
recorded EMG signal is defined as the EMG amplitude. The EMG amplitude estimation
technique varies with the study from conventional type of processing (i.e. rectification
followed by low pass filtering) to further addition of different noise rejection and signal
tonoise ratio improvement stages. Advanced EMG amplitude processors developed
recently that incorporate signal whitening and multiplechannel combination have been
shown to significantly improve amplitude estimation. The main contribution of this
research is a comparison of the performance of EMGtorque estimators with and without
these advanced EMG amplitude processors.
The experimental data are taken from fifteen subjects that produced constant
posture, nonfatiguing, forcevarying contractions about the elbow while torque and
biceps/triceps EMG were recorded. Utilizing system identification techniques, EMG
amplitude was related to torque through a zerosonly (finite impulse response, FIR)
model. The incorporation of whitening and multiplechannel combination separately
reduced EMGtorque errors and their combination provided a cumulative improvement.
A 15
th
order linear FIR model provided an average estimation error of 6% of maximum
voluntary contraction (or 90% of variance accounted for) when EMG amplitudes were
obtained using a fourchannel, whitened processor. The equivalent singlechannel,
unwhitened (conventional) processor produced an average error of 8% of maximum
voluntary contraction (variance accounted for of 68%).
iv
This study also describes the occurrence of spurious peaks in estimated torque
when the torque model is created from data with a sampling rate well above the
bandwidth of the torque. This problem is anticipated when the torque data are sampled at
the same rate as the EMG data. The problem is resolved by decimating the EMG
amplitude prior to relating it to joint torque, in this case to an effective sampling rate of
40.96 Hz.
Keywords: EMG, EMG Amplitude, Torque, EMGtorque Model, Optimal Sampling
Rate, System Identification, and Linear Torque Model.
v
ACKNOWLEDGMENTS
I am deeply thankful to my advisor, Professor Edward Clancy, who initially introduced me
to the fascinating field of Electrical Engineering, during my freshman year while taking EE2011.
My respect toward him grew as I started to realize that the knowledge gained and the work ethics
developed in my introductory class brought me to this point. I feel honored to have had a chance
to work with him in this research for my Masters Degree and because of him to have reached so
far. Without his guidance and support, the completion of this thesis would not have been possible.
I thank Dr. Denis Rancourt from University of Shebrooke (Canada) for his help throughout
the course of this project and for extending his help even beyond this project completion. Heartfelt
thanks to my MQP advisor Professor David Cyganski and to Professor Rick Brown for their
presence in the research committee and their helpful advice in the project. Also thanks to my lab
mates Karthik and Hongfang for the joyful, kind, and sharing atmosphere created in our lab.
I owe everything I am and I have done to the selfless sacrifice and to the love which my
family raised me with. We have been together through many hardships and now it is time to enjoy
the fruits of our work. Thanks to my dearest sister Ana for the mutual love and loyal friendship.
“Falenderoj nga thellesia e zemres prinderit e mi te shtrenjte per sakrificen dhe durimin gjate
veshtiresive. Nuk mund ta imagjinoni dot mirenjohjen per dashurine, vullnetin, dhe vlerat
shpirterore me te cilat me kini rritur duke shpresuar qe nje dite te arrijme te gjitha endrrat e
thurura se bashku.”
Thanks to my husbandtobe Eno for being my best friend and for already standing by me
for better and for worst. With his continuous love and care, he has encouraged me to work harder
to reach our life dreams.
Beyond all and everything, I am thankful to the Almighty God, for answering my prayers, and for
always guiding me toward happiness and peace.
vi
TABLE OF CONTENTS
ABSTRACT................................................................................................................................................ III
ACKNOWLEDGMENTS............................................................................................................................V
TABLE OF CONTENTS........................................................................................................................... VI
LIST OF FIGURES................................................................................................................................. VIII
LIST OF TABLES...................................................................................................................................... IX
CHAPTER 1. INTRODUCTION................................................................................................................ 1
1.1. PROJECT MOTIVATION................................................................................................................. 1
1.2. THESIS CONTRIBUTION ................................................................................................................ 3
1.3. THESIS CONTENT ......................................................................................................................... 4
CHAPTER 2. PROJECT BACKGROUND............................................................................................... 6
2.1. EMG SIGNAL FUNDAMENTALS.................................................................................................... 6
2.1.1. Electrical Activity Generation ................................................................................................ 6
2.1.2. Origin and Character of EMG ............................................................................................... 8
2.1.3. Factors that Effect EMG Signal............................................................................................ 12
2.2. SURFACE EMG AMPLITUDE ESTIMATION TECHNIQUES............................................................. 15
2.2.1. Standard EMG Amplitude Estimation .................................................................................. 15
2.2.2. Advanced EMG Amplitude Estimation ................................................................................. 16
2.2.3. Noise Rejection Filters ......................................................................................................... 18
2.2.4. Adaptive Whitening............................................................................................................... 19
2.2.5. Multiple Channel Combination and Gain Normalization..................................................... 22
2.3. BIOMECHANICAL SYSTEM MODELING TECHNIQUES .................................................................. 22
2.3.1. Overview of Modeling Techniques ....................................................................................... 23
2.3.2. Parametric System Identification.......................................................................................... 24
2.3.3. EMGTorque Relationship Modeling ................................................................................... 27
CHAPTER 3. SURFACE EMG TO TORQUE MODEL DESIGN....................................................... 34
3.1. EMGTORQUE MODEL DESIGN ................................................................................................. 34
3.1.1. Physical Interpretation of the EMGTorque Model.............................................................. 35
3.1.2. Mathematical Modeling for EMGtorque............................................................................. 38
3.2. MODEL SOLUTION...................................................................................................................... 40
CHAPTER 4. DATA COLLECTION AND ANALYSIS METHODS .................................................. 43
4.1. EMG DATA COLLECTION........................................................................................................... 43
4.1.1. Noise Reduction Precautions................................................................................................ 44
4.1.2. Apparatus and Experimental Procedure ............................................................................. 46
4.2. EMG AMPLITUDE ESTIMATION METHOD .................................................................................. 51
4.3. SYSTEM IDENTIFICATION PROCEDURE ....................................................................................... 53
4.3.1. Data PreProcessing ............................................................................................................ 53
4.3.2. Torque Estimation Procedure............................................................................................... 55
4.3.3. Model Performance Measures.............................................................................................. 57
CHAPTER 5. PROJECT RESULTS........................................................................................................ 60
5.1. DECIMATION.............................................................................................................................. 60
5.2. COMPARISON OF EMG AMPLITUDE PROCESSORS...................................................................... 64
CHAPTER 6. DISCUSSION AND CONCLUSIONS ............................................................................. 71
vii
6.1. DISCUSSION OF RESULTS............................................................................................................ 71
6.1.1. Advances to EMGtorque Estimation ................................................................................... 71
6.1.2. Study Limitations and Future Suggestions ........................................................................... 72
6.2. SUMMARY AND CONCLUSIONS................................................................................................... 74
REFERENCES ........................................................................................................................................... 76
APPENDICES: ADDITIONAL INFORMATION, PLOTS, AND FIGURES...................................... 82
I. LBXXXX EXPERIMENT DATA FILE DESCRIPTION ......................................................................... 82
II. OPTIONAL PROPERTIES FOR AMPLITUDE ESTIMATION ALGORITHM ............................................... 84
III. EXTRA FIGURES AND PLOTS....................................................................................................... 86
IV. PAPER SUBMITTED TO THE JOURNAL OF BIOMECHANICS ........................................................... 92
viii
LIST OF FIGURES
FIGURE 2.1: MUSCLE FIBERS COMPOSITION [PERRY AND BEKEY, 1981] ....................................................... 7
FIGURE 2.2: GENERATION OF ELECTRIC FIELD IN MUSCLE FIBERS [PERRY AND BEKEY, 1981] ..................... 8
FIGURE 2.3: OBSERVED MOTOR UNIT ACTION POTENTIAL, MUAP [BASMAJIAN AND DE LUCA, 1985]. ....... 9
FIGURE 2.4: EMG SIGNAL ORIGIN BLOCK DIAGRAM [BASMAJIAN AND DE LUCA, 1985]............................ 11
FIGURE 2.5: SIX STAGES MULTICHANWHIT EMGAMP PROCESSOR [CLANCY ET AL., 2001]..................... 17
FIGURE 2.6: SINGLECHANWHIT PROCESS FOR EMGAMP ESTIMATION [CLANCY ET AL., 2004]................. 18
FIGURE 2.7: MODEL OF EMG USED FOR ADAPTIVE WHITENING FILTERS [CLANCY AND FARRY, 2000]........ 20
FIGURE 2.8: ADAPTIVE WHITENING OF EMGAMP ESTIMATION [CLANCY AND FARRY, 2000] ...................... 21
FIGURE 2.9: SYSTEM IDENTIFICATION PROBLEM (BLACKBOX TYPE OF MODELING) .................................... 24
FIGURE 2.10: GENERIC DYNAMIC SYSTEM BLOCK DIAGRAM (DISCRETE TIME SIGNALS) ............................. 25
FIGURE 3.1: RAW SURFACE EMG TO TORQUE MODEL [CLANCY AND HOGAN, 1997] ................................. 37
FIGURE 4.1: EMG ELECTRODE PLACEMENT [DE LUCA, 2002]..................................................................... 44
FIGURE 4.2: BIODEX EXERCISE MACHINE FOR THE EXPERIMENT [BOUCHARD, 2001] ................................. 47
FIGURE 4.3: SURFACE EMG ELECTRODES AND ACQUISITION BOX [BOUCHARD, 2001]............................... 48
FIGURE 4.4: SUBJECT DURING EXPERIMENT [BOUCHARD, 2001] ................................................................. 50
FIGURE 4.5: BLOCK DIAGRAM OF EMG DATA PREPROCESSING FOR SYSTEM ID ALGORITHM................... 54
FIGURE 4.6: SYSTEM IDENTIFICATION PROCEDURE [CREATED BASED ON LJUNG, 1999].............................. 58
FIGURE 5.1: CHANGES OF PREDICTED TORQUE WHILE INCREASING DECIMATION RATE.............................. 61
FIGURE 5.2: SIGNAL POWER ACCUMULATION (AVERAGE PSD TORQUE) VS. FREQUENCY ........................... 62
FIGURE 5.3: DECIMATION RATE EVALUATION PLOT.................................................................................... 64
FIGURE 5.4: RAW EMG (FLEXION & EXTENSION) AND TORQUES................................................................. 65
FIGURE 5.5: MEDIAN (LEFT) AND MEAN (RIGHT) OF % VAF AND % MAE FOR FAST TRACKING.................. 66
FIGURE 5.6: PSD OF ERROR ACCUMULATION RATE..................................................................................... 68
FIGURE 5.7: THE AVERAGE PSD OF ERROR AS ESTIMATED FROM WELSH PERIODOGRAM........................... 68
FIGURE 5.8: HIGH DC OFFSET ERROR ON ESTIMATED TORQUE ................................................................... 69
FIGURE 0.1: ESTIMATION ERROR PSD (WELSH PERIODOGRAM) FOR ALL 4 PROCESSORS............................ 86
FIGURE 0.2: SYSTEM PERFORMANCE (% VAF & MAE) USING QR FACTORIZATION (FAST TRACKING) ...... 87
FIGURE 0.3: SYSTEM PERFORMANCE (% VAF & MAE) USING PSEUDOINVERSE (SLOW TRACKING) ........ 88
FIGURE 0.4: SYSTEM PERFORMANCE (% VAF & MAE) USING AC PART OF EMG AMPLITUDES (FAST
TRACKING + PINV) ............................................................................................................................. 89
FIGURE 0.5: COEFFICIENTS FREQUENCY RESPONSE FOR A TYPICAL EMGTORQUE MODEL (SLOW
TRACKING) .......................................................................................................................................... 90
FIGURE 0.6: COEFFICIENTS FREQUENCY RESPONSE FOR A TYPICAL EMGTORQUE MODEL (FAST TRACKING)
............................................................................................................................................................ 91
ix
LIST OF TABLES
TABLE 2.1: FACTORS THAT INFLUENCE SURFACE EMG [FARINA, MERLETTI, AND ENOKA, 2004] .............. 14
TABLE 2.2: COMMON BLACKBOX MODELS, SIMPLIFICATION OF GENERAL EXPRESSION ........................... 26
TABLE 4.1: SUBJECT INFORMATION (CODE, AGE, AND GENDER) ................................................................. 49
TABLE 4.2: A/D ELECTRODE CHANNELS FROM THE EXPERIMENTAL DATA ................................................. 52
TABLE 4.3: FOUR PROCESSORS TYPES (PROCESSOR 14) ............................................................................. 52
TABLE 5.1: DISTRIBUTION INFO OF % VAF VALUES FOR EACH PROCESSOR (FAST TRACKING) ................. 67
TABLE 5.2: DISTRIBUTION INFO OF % MAE VALUES FOR EACH PROCESSOR (FAST TRACKING)................. 67
TABLE 0.1: TRIAL ID NAME CODES............................................................................................................. 82
TABLE 0.2: A/D CHANNEL NAME CODES..................................................................................................... 83
1
CHAPTER 1. INTRODUCTION
“Electromyography is a seductive muse because it provides easy access to physiological
processes that cause the muscle to generate force, produce movement, and accomplish
the countless functions that allow us to interact with the world around…To its
detriment, electromyography is too easy to use and consequently too easy to abuse.”
[Carlo J. De Luca, 1993]
The contraction of muscle fibers generates electrical activity that can be measured by
electrodes affixed to the skin surface on top of the muscle group. The recorded spikes of
electrical activity are referred to as the electromyogram signal or “raw” EMG. The
surface EMG signal recorded using large electrodes (e.g., diameter 5 mm) that monitor
the activity of multiple muscle fibers can be well modeled as a zeromean timevarying
stochastic process. Motor units are the smallest functional muscle group. It is observed
that the standard deviation of the raw EMG signal is monotonically related to the number
of the activated motor units and the rate of their activation. This standard deviation is
used to approximate the magnitude of the muscular electrical activity referred to as EMG
amplitude [Clancy and Hogan, 1997]. EMG amplitude has a variety of applications, such
as a control signal for myoelectrical prostheses, ergonomic assessments, biofeedback
systems, and it is used to approximate the torque about a joint [De Luca, 1993; Thelen et
al., 1994; Gottlieb and Agarwal, 1977; ValeroCuevas et al., 2003].
1.1. PROJECT MOTIVATION
After obtaining high quality estimates of EMG amplitudes, a common practice is
relating them to the tension of individual muscles via mathematical models, even though
2
there are limitations to this method. The tension produced by individual muscles can not
be measured noninvasively, thus there is no direct mechanical method to validate the
model predictions. In addition, the existence of crosstalk (defined as the interfering
electrical activity from the surrounding muscles) and the inability to measure this effect
add to the difficulties of creating this model.
Considering the mentioned limitations, many researchers [Gottlieb and Agarwal,
1977; Clancy and Hogan 1997; Thelen et al., 1994] have focused their efforts on relating
the EMG amplitude to the torque about a joint as the next logical and practical
alternative. The effect of crosstalk may be automatically canceled or minimized in the
case of the torque about the joint [Clancy et al., 2001]. Total net torque about a joint can
be easily verified via mechanical measurements. Furthermore, considering coactivation
effects on underlying group muscles, the system model performance is evaluated against
the net joint torque contribution, rather than the individual ones that are impossible to
distinguish.
Over the last few years, there are clear advances in estimating EMG amplitude yet the
EMGtorque modeling has not benefited from this progress. If EMG is a useful indicator
of the muscular tension, it is necessary to develop accurate means of quantification, both
in terms of properly measuring and interpreting EMG and in creating mathematical
models relating EMGtorque. Amplitude estimation accuracy influences the performance
of EMGtorque models, because torque about a joint (tension exerted in muscles) is the
outcome of proper EMG signal interpretation and consequently its careful treatments.
The importance of EMG signal processing can not be emphasized enough, since
3
electromyography is such a powerful and physiologically easily obtained tool, therefore
as expressed by De Luca, its misusage can lead to fatal mistakes [De Luca, 1993].
Demonstrating the benefit of utilizing advanced EMG amplitude processing, Clancy
and Hogan (1997) showed that the torque estimation error is reduced when using
improved EMG amplitude processors. The experiment results were obtained using a
linear model to relate EMG amplitude from biceps/triceps to the elbow joint torque in the
case of constantposture and constantforce contractions. Additionally, encouraging
results were also obtained in less constrained conditions (slowly varying force), but
several trial combinations that lead to unrealistic model performance (considered as
model nonconvergence) were an obstacle that needed further investigation [Bouchard,
2001]. The result of the previous research inspired the focus of this project: relating the
EMG amplitudes from biceps/triceps to the torque about the elbow and proving that
better EMG amplitude processing leads to better torque predictions during dynamic
experimental tasks (forcevarying contractions).
1.2. THESIS CONTRIBUTION
The goal of this project was to demonstrate that the usage of high fidelity processing
techniques (inclusion of whitening filters and multiple channels) for EMG amplitudes
leads to improvements in the accuracy of estimating torque. To achieve this main
objective, it was necessary to develop a model to relate EMG amplitude to torque and
compare the model performance, as the EMG amplitude processors were varied among
four different types. These four types of processors were obtained using the combination
of multiple channel recordings with the addition of adaptive whitening. The four
processors created were: singlechannelunwhitened, singlechannelwhitened, multiple
4
channelunwhitened, and multiplechannelwhitened. Further accomplishment was to
determine the decimation rate required for the data prior to applying them to the system
identification algorithms. Decimation solved some model nonconvergence problems
encountered during prior research [Bouchard, 2001].
In conclusion, there are several important deliverables from the completion of this
project work. The first is the model used to relate extension/flexion EMG amplitudes to
torque about the elbow. The second is the data preprocessing routine (decimation)
required to achieve the maximum performance from the model. Finally, the thorough
documentation of the results and the steps achieving them, along with the
recommendations for improvements will serve as starting point for future research.
1.3. THESIS CONTENT
The content of this paper is presented in a logical and chronological, order as
appropriate in order to explain the process involved in completing the project.
CHAPTER 2 provides background information about the EMG signal starting from its
recording to amplitude processing techniques, focusing on the adaptive whitening filters
as a new step that has revolutionized the existing processing methods. There is also a
review of some of the most common system identification models. The chapter ends
with a brief review of the literature on EMGtorque modeling techniques. Following the
background, CHAPTER 2 is the model design development chapter. It includes all
physiological concepts and thoughts that were poured into quantifying the EMG to torque
relationship, reaching into the linear (ARX) model used in this project. The model then
is solved, describing most of the algebraic steps involved into obtaining a linear least
squares error solution.
5
CHAPTER 4 explains the data collection method and the process of obtaining EMG
amplitudes from the four different processors. EMG amplitude estimation, decimation,
and truncation are part of a preprocessing routine used prior to system identification.
The system identification procedure involves two main steps, training and validation.
During training, a coefficient vector is fit to the input data based on the least squares error
minimization. Model validation requires utilizing a distinct dataset to estimate the output
using the optimal coefficients. The details of the traintest paradigm along with
definitions of model performance quantifiers are also explained in this chapter. The
subsequent CHAPTER 5 describes the results obtained after following the tests explained
in the previous methodology chapter. The chapter includes general observations and
hypotheses derived through experimental data interpretation to validate the observations.
Interpretation and the study limitations are discussed in detail in the last chapter
(CHAPTER 6). This chapter summarizes the main contribution of this research and it
lays out some suggestions for future work, based on the conclusions drawn. Finally, the
document ends with the APPENDICES: that includes additional information on the
experimental data and some additional plots that were not crucial to the results, but
support their interpretations.
6
CHAPTER 2. PROJECT BACKGROUND
The content of this chapter is intended to provide background information necessary
to understand the subsequent sections that describe the specific thesis contribution. The
chapter starts with a brief introduction of the physiological raw EMG signal, focusing on
the random character of EMG. Then, it continues with a brief description of the
techniques used to process the EMG amplitude. At the end, there is a review of
achievements in relating the surface EMG amplitude to the torque about a joint following
a summary of modeling techniques.
2.1. EMG SIGNAL FUNDAMENTALS
The electromyogram (EMG) is the recording of the electrical activity produced within
the muscle fibers. The relation of surface EMG to torque makes EMG an attractive
alternative to direct muscle tension measurements, necessary in many physical
assessments. However, the complexity of the EMG signal origin has been a barrier for
developing a quantitative description of this relation. The EMG signal origin and
character is necessary background to understand the difficulty of establishing a
relationship between surface EMG and torque. The description in this section is brief and
selective; the reader is suggested to review Basmajian and De Luca (1985) for more
details.
2.1.1. Electrical Activity Generation
Electrical activity in the muscles arises from the contraction of the muscle fibers,
the structure of which is shown in Figure 2.1. Each muscle fiber contains a bunch of
7
myofibrils (long chains of contractile units). The myofibrils contain long chains of
contractile units called sarcomeres, which contribute to the force exerted within the
muscles.
Figure 2.1: Muscle Fibers Composition [Perry and Bekey, 1981]
Each of the myofibrils is chemically activated by local neurons, generating an electrical
charge that moves up and down the myofibril, activating the chains of sarcomeres (Figure
2.2). The charge motion generates an electromagnetic field that induces volume
conduction, which enables recording of an electrical signal both internally at the muscle
and externally at the surface over it. The detected waveform resulting from the
depolarization of the wave propagating between the motoneuron and end plate is called
the muscle fiber action potential (MAP). MAPs are not commonly seen in the general
EMG literature, because they are recorded using microelectrodes, and can not be picked
up by the noninvasive surface electrodes.
8
Figure 2.2: Generation of Electric Field in Muscle Fibers [Perry and Bekey, 1981]
The muscle fibers contract in groups that are controlled by the central nervous
system via nerve fibers (axons) transmitting the signal to the ending neurons. To
simplify analysis and mathematical interpretation of EMG, the smallest controllable
functional unit of muscle fibers is defined as a motor unit (MU). The motor unit consists
of a single motoneuron, its neuromuscular junction, and the muscle fibers that it excites.
The number of the muscle fibers contained in a MU varies with the size of the muscle
within which a MU belongs. Smaller muscles have MUs that contain 310 myofibrils,
while larger ones contain up to 2000 myofibrils. It is important to emphasize the similar
structure of the muscles, regardless of the scaling on the size and the number of
myofibrils [Perry and Bekey, 1981; Lamb and Hobart, 1998].
2.1.2. Origin and Character of EMG
Microelectrodes on the cell surface are not the only way to measure motor action
potentials. The living tissues act as volume conductors, therefore a potential at the
motoneuron source is spread away via ion movements throughout the entire unit volume.
Applying the same principles of conduction described above, the action potential
9
propagates along the motoneuron to the endplate of the muscle fibers (Figure 2.3). The
electrical potential surrounding the muscle fibers changes, because the geometry of the
conducting volume changes. Therefore, the conduction times of the muscle fibers in a
motor unit are different. The spatiotemporal summation of the individual myofibril
action potentials recorded by the electrode is called a motor unit action potential
(MUAP). Figure 2.3 represents the motor unit action potential as the superposition of
MUs generated by each of the myofibrils. Each muscle fiber within the MU (on the left
of the figure) contributes to the surface potential (on the right of the figure). [Basmajian
and De Luca, 1985].
Figure 2.3: Observed Motor Unit Action Potential, MUAP [Basmajian and De
Luca, 1985].
10
The recorded MUAP is an attenuated version of the action potential generated in
the muscle fibers because of the filtering effect that is due to the transmission line
between the motoneuron and electrode. In particular, the tissue acts as a low pass filter
with a cutoff frequency proportional to the distance of the electrode to the signal source
[Lindstrom and Magnusson, 1977]. Usually, the individual MUAP is recorded using fine
wire electrodes, although under certain conditions surface electrodes can be used. The
duration of the MUAPs can vary from a few milliseconds to 14 ms, and their amplitudes
vary from microvolt ranges to a maximum of 5 mV. Typical surface EMG electrodes are
used to record the myoelectric activity of the skeletal muscle as a whole, rather than
individual MUAPs. Generally, the pickup area of an electrode includes more than one
motor unit, because muscle fibers of different motor units are mixed throughout the entire
muscle [Lamb and Hobart, 1992].
The MUAP is the response of the motor unit MU to a single motoneuron
excitation. If the stimulus is modeled as an impulse dirac function, δ(t), then the MUAP
is considered the impulse response h(t). The repetitive sequence of stimulations to the
motor units results into a series of impulse responses referred to as the motor unit action
potential train (MUAPT). Each of the motor unit responses to the impulse train is
independent from the sequence and the total series response has a random character.
Therefore, the superposition of the MUAPTs is the physiological EMG signal and can be
modeled as stochastic process (sum of independent random variables).
11
Figure 2.4: EMG Signal Origin Block Diagram [Basmajian and De Luca, 1985]
A schematic representation of the EMG generation is shown in Figure 2.4. The
symbol m
p
(t, F), myoelectric signal as a function of time (t) and the number of firings (F),
represents the physiological EMG and it is not recordable or measured. The detected
EMG signal that is utilized in the research is the observed signal m(t, F) that is
12
contaminated with electronic noise (almost white) and has lost some of the high
frequency components due to the filtering effects at the electrodes [De Luca, 1993].
To conclude, considering the EMG signal as a time varying stochastic process
gives the possibility to model it as a zeromean Gaussian distribution, because EMG is
the sum of a large number of MUAPs [Papoulis, 2002]. This random character of the
EMG signal enables the later described approximation of EMG amplitude as the square
root of the detected signal’s variance. In addition, the recorded EMG signal is dependent
on the type, geometry, and position of the recording electrodes. The depolarization wave
also causes chemical changes that result in a mechanical twitch, which is slower than the
electrical response, and delayed by 50100 msec. This mutual relation of EMG and
mechanical activity to the MUAPs inspires the establishment of an EMGtorque
relationship that will be discussed in detail in the upcoming chapters.
2.1.3. Factors that Effect EMG Signal
There are many factors identified in the research as having a great influence on
EMG interpretation. Even though they all are important, a common practice among
researchers has been to focus on the effects that have the most impact on the application
for which the EMG signal is used [DeLuca, 1993; Farina, Merletti, and Enoka, 2004;
Perry and Bekey, 1981; Lamb and Hobart, 1992]. This section also will follow the same
rule, and briefly describe some of the factors that directly effect the EMG signal
interpretation and analysis when estimating torque. Quantifying the factors that effect
EMG signals is a complex task, because there is not enough information to validate the
assumptions. Considering the varieties in the structure of electrodes and living tissues, it
also is impossible to generalize the observations over all subjects and cases.
13
De Luca (1993) categorizes the factors that effect EMG signal and force into three
groups: causative, intermediate and deterministic factors. The causative factors are the
basis of EMG signal and they are both intrinsic and extrinsic. The extrinsic factors are
related to the electrode structure and its placement on the skin overlying the muscle.
Such instances include the electrode configuration, location, and the orientation of
detection surfaces relative to the muscle fibers. On the other hand, the intrinsic causative
factors are related to the physiological, anatomical and biochemical character of EMG
signals. These factors can not be controlled, but their knowledge and understanding help
with the accuracy of EMG interpretation. The causative intrinsic factors include the
number of active MUs at the time, the pH level in the muscle fibers, the blood flow, and
geometry of the fibers. The intermediate factors (i.e. crosstalk, conduction volume and
velocity, superposition, etc.) are the effects that are influenced by the causative factors
and in consequence they influence the deterministic factors (i.e. number of MUs
activated, MU firing rate, MUAP shape and duration, etc.). The amount of the effect that
the deterministic and the intermediate factors have on EMG is an applicationbased
evaluation.
Table 2.1 from Farina et al. (2004) represents a summary of the known effects to
EMG interpretation. The presence of subcutaneous fatty tissues becomes a significant
factor, because the loss of the high frequency components reduces the spectrum of the
EMG signal. Besides the stability of the position of the electrodes and the stability of the
MU firing rate, the issue of crosstalk is always present. Crosstalk is defined as the
interference pattern recorded from a distant muscle when the electrodes are intended to
monitor another muscle. Crosstalk is an issue that can be misleading when EMG is
14
explained by the properties of volume conduction. Simulation and analyses have shown
that the crosstalk can neither be measured nor eliminated with the existing technology.
Therefore, it should be recognized while utilizing EMG to estimate muscle forces
[Farina, Merletti and Enoka, 2004].
Table 2.1: Factors that Influence Surface EMG [Farina, Merletti, and Enoka, 2004]
Nonphysiological
Anatomic
Detection System
Geometrical
Physical
Physiological
Fiber membrane
properties
Motor unit properties
Shape of the volume conductor
Thickness of the subcutaneous tissue layers
Distribution of the MUs territories in the muscle
Size of the motor unit territories
Distribution and the number of fibers in the MU territories
Length of the fibers
Spread of the endplates and tendon junction within MUs
Spread of the innervations zones and tendon regions among MUs
Presence of more than one pinnation angle
Skin electrode contact (impedance or noise)
Spatial filter for signal detection
Interelectrode distance
Electrode size and shape
Inclination of the detection system relative to the fiber
orientation
Location of the electrodes over the muscle
Muscle fiber shortening
Shift of the muscle relative to the detection system
Conductivities of the tissues
Amount of the crosstalk from the nearby muscles
Average muscle fiber conduction velocity
Distribution of the MU conduction velocities
Distribution of the conduction velocities within in MUs
Shape of the intracellular action potential
Number of recruited MUs
Distribution of motor unit discharge rates
Statistics and coefficient of variation for discharge rate
MU synchronization
15
2.2. SURFACE EMG AMPLITUDE ESTIMATION TECHNIQUES
If the EMG amplitude is defined as the standard deviation of the raw EMG signal,
then it can be estimated by applying standard statistical techniques [Clancy and Hogan,
1997]. Since raw EMG is a stochastic process in nature, its statistical processing can be
used for predictive purposes. The estimation of the EMG amplitude has been refined and
improved since the early EMG amplitude developed from a simple rectifier and lowpass
filtering [Imnan et al., 1952]. The state of art EMG amplitude processing includes six
stages that will be discussed separately after a brief presentation of the complete process.
2.2.1. Standard EMG Amplitude Estimation
The most common technique of detection for EMG amplitude is the rectification
process followed by a smoothing step. According to Hof and Van Den Berg (1981), the
recorded EMG signal is described as the product of a zeromean stochastic process with
the timevarying EMG intensity. Therefore the intensity of the EMG signal (EMG
amplitude) can be obtained by proper rectification and smoothing [Hof and Van Den
Berg, 1981]. The early researchers in the field studied and utilized nonlinear analog
circuits, such as a full wave rectifier and a low pass filter made of simple passive
components (resistors and capacitors), to detect the signal [Bigland and Lippold, 1954].
This method eventually led to the use of the statistical moving average mean absolute
value (MAV) and the moving average root mean square (RMS).
Moving Average Mean Absolute Value:
∑
+ − =
=
t
N t i
i t
x
N
MAV
1
1
(2.1)
Moving Average Root Mean Square:
∑
+ − =
=
t
N t i
i t
x
N
RMS
1
2
1
(2.2)
16
where in both expressions N is the number of samples in each smoothing window
of the moving average filter; t is the time at which this interval starts; and x
i
is the
signal being smoothed in the timedomain.
EMG amplitude can also be computed in software using either one of the above
formulae. The amplitude estimates found using RMS and MAV calculations exhibit very
similar performance. However, the MAV method has been initially used more than the
RMS, because of the lesser amount of time necessary for computations. Currently, the
computation time is not as problematic especially when processing is performed offline.
The process of detection is followed by smoothing and relinearization. The
method of accomplishing the two last steps differs between RMS and MAV. In the case
of RMS, the detection of the signal is achieved by squaring all the terms. The resulted
squared terms are smoothed by taking their average and then relinearized by taking the
square root of the mean. The detection for the MAV method is done by taking the
absolute value of the terms. The result is smoothed by taking the average of these terms.
In this case, there is no need to relinearize.
2.2.2. Advanced EMG Amplitude Estimation
The EMG signal processing is a crucial factor in the way that EMG amplitude is
interpreted and used in different applications. Therefore, specifying and understanding
the steps involved in the processing technique is extremely important. The estimator has
evolved from the use of a simple rectification and a lowpass filter. An advanced EMG
amplitude estimator consists of the following six stages (Figure 2.5):
1. Noise rejection filter
2. Adaptive whitening
3. Multiple Channel Combination and Gain Normalization
17
4. Rectification and Demodulation
5. Smoothing
6. Relinarization
Figure 2.5: Six Stages MultiChanWhit EMGamp Processor [Clancy et al., 2001]
In the above figure, inputs m
k
(k = 14) are the recorded signals from the surface
electrodes placed on top of each of the muscle groups. The output ) ( ˆ t s is the estimated
EMG amplitude (EMGamp). The pictorial presentation of the signal transformation for
each of the channels is given in Figure 2.6.
Each of the surface EMG signal m
k
is transformed to the EMG amplitude
k
sˆ after
passing through all the stages of the processor. In the first stage, motion artifact is
attenuated with a highpass filter. In the second stage, the signal is whitened. The
adaptive whitening has demonstrated better performance for lowamplitude levels. Stage
three rectifies the signal and then raises it to a power to make it nonlinear. In stage four,
the demodulated samples are averaged (smoothed). In stage five, the signal is
18
relinearized by raising it to the inverse of the power applied previously. During the
“Detect” and “Relinearize” stages, d=1 for MAV and d=2 for RMS.
Figure 2.6: SingleChanWhit process for EMGamp estimation [Clancy et al., 2004]
The smoothing step is omitted when the EMG amplitude obtained is used to estimate
torque. Additional detail of these steps is given in the following sections.
2.2.3. Noise Rejection Filters
High pass filters, prior to RMS and MAV, are used to eliminate the noise from
motion artifact. The power density of motion artifact is mostly below 20 Hz; therefore, a
19
highpass filter with cutoff frequency between 1020 Hz is sufficient to reduce/eliminate
these effects. Cutoff frequencies greater than 20 Hz can cause loss of EMG signal,
considering that the rolloff of the real filters can coincide with the median frequency of
the EMG signal, especially during fatigue [Clancy, Morin, and Merletti, 2002]. The high
pass filter can be analog incorporated into the hardware instrumentation and/or digital
implemented in software. The advantage of using digital filters is the ease of
implementing high order filters to achieve sharp rolloff and eliminate more of the noise
power ensuring that the loss of useful information is minimal. In some cases, analog
filters are used in addition to digital filters to prevent saturation caused if the EMG signal
is corrupted by large amplitude motion artifact.
2.2.4. Adaptive Whitening
The whitening step is recently included in EMG signal processing software
algorithms. The term whitening originates from the power of the white light spectrum
spreading out uniformly over all frequencies. Whitening an EMG signal is the process of
decorrelating the neighboring samples in the time domain. Doing so, the statistical
bandwidth increases therefore the approximation of standard deviation is more accurate
[Bendat and Piersol, 1986]. The adaptive whitening removes the additive noise described
in the physiological model of EMG created by Clancy and Farry (2000) presented in
Figure 2.7.
20
Figure 2.7: Model of EMG used for adaptive whitening filters [Clancy and Farry,
2000]
A more detailed description of the model and the math behind it can be found in the
original source. Briefly, the signal w
i
is a zero mean Gaussian random process of unit
variance that serves as a start for modeling the EMG. This signal is passed through a
shaping filter, H
time
that creates the lowpass effect of the tissues and skin layers on real
EMG signal while still maintaining unit standard deviation. The output (n
i
) is then
multiplied by the amplitude of EMG (s
i
) resulting to the noisefree EMG (r
i
). The signal
v
i
is a zeromean random process representing additive electronic noise and random noise
from the electrodeskin interface that is summed with r
i
to complete the model of the
measured surface EMG, m
i
. Recalling the physiological description of EMG in the
previous section, this model is consistent with the character of the raw EMG in Figure
2.4.
The shape of adaptive whitening filters is formed based on the power spectral
density (PSD) of the noiseless signal and the additive noise. Briefly, the shape of the
original whitening filter is the inverse square root of the PSD taken from the true EMG
signal. The adaptive whitening involves incorporating a noise attenuation stage that
operates based on the relative power of the signal over the existing noise. Adaptive
whitening is necessary, because it is observed that the noise exhibits a larger relative
w
i
n
i
) (
jw
time
e H
s
i
r
i
m
v
i
Σ
21
magnitude during low level contractions, where the relative EMG intensity is lower. The
time duration of the whitening filter is short; hence, the EMG amplitude remains
essentially constant during that period, making the adaptive whitening process quasi
stationary [Clancy and Bouchard, 2001].
Figure 2.8: Adaptive whitening of EMGamp estimation [Clancy and Farry, 2000]
The whitening process proposed by Clancy and Farry (2000) includes three stages
used to improve amplitude estimation (Figure 2.8). Without including the details (they
can be found in the mentioned source) the first stage of this process whitens the noiseless
EMG amplitude s
i
, but also a filtered version of the additive independent noise v
i
. The
second stage optimally estimates the noisefree whitened signal i m by adaptively
removing the noise through a Wiener filter. The third stage applies an adaptive gain
determined based on the transformations of EMG signal from the two previous stages.
This step is used to maintain the variance of the EMG signal throughout the complete
whitening stage [Clancy and Farry, 2000].
22
2.2.5. Multiple Channel Combination and Gain Normalization
This step involves the combination of EMG recordings obtained from several
electrodes placed adjacent to each other, on the skin overlying the same muscle. The
reason for the combination of multiple channels is that the SNR improves with the
increase in the volume of muscles recorded. Since the gain and the distance from the
muscle differ from electrode to electrode, the combination of the recordings is followed
by the gain normalization process. This ensures equal contribution from each of the
recordings, and can be considered as decorrelation of the signal spatially. Research has
shown that using several electrodes for measuring the EMG from a muscle results in
more accurate EMG amplitude estimation. SNR performance improvements of up to
91% have been observed using multiple channels, as compared to the results from a
single channel processing [Hogan and Mann, 1980b]. There are also some disadvantages
to the multiple electrode recording combination including that the chance of defects that
may arise due to noise, shorted electrodes, etc. is increased with the number of channels
[Hogan and Mann, 1980a].
2.3. BIOMECHANICAL SYSTEM MODELING TECHNIQUES
System identification is a study of the dynamics and physical behavior of systems
under external disturbances. Specifically, it is a set of standardized guides on building
system mathematical models based on observations made on system reactions. The
external data that can be manipulated and measured by the user are referred to as inputs
and others as disturbances, even though most of the time their difference does not affect
the modeling process. The measured/observed response of the system is referred to as its
23
output. This section gives a brief description of the system identification. Detailed
reference of the models and system identification techniques are found in Ljung (1999).
The dependence of the recorded EMG signal and muscle tension on mutual
physiological factors inspires ongoing research work to develop mathematical models
relating EMG to torque. The experimental studies have explored both linear and
nonlinear models to achieve better accuracy. Some researchers have even built complex
models that describe the details of muscles, however little or no improvement is seen in
doing so. Keeping in mind the ultimate goal of this research, this section also illustrates
the forgoing theory of mathematical modeling with some EMGtorque examples found in
the literature.
2.3.1. Overview of Modeling Techniques
Modeling of the complex relationship between muscular activity and torque has
been approached in two different methods; a priori (morphological) and a posteriori
(black box) type of modeling techniques [Westwick, 1995]. The morphological modeling
technique involves designing a model based on the physical characteristics of the system.
The parameters are flexible and well adapted to the system itself. The drawback of this
method is the large number of parameters that result in a high level of complexity.
Additionally, it requires a thorough understanding of the system structure, while most of
the times, the system is unknown and it is considered as a black box (Figure 2.9). The
black box type of modeling is referred to as system identification, and it is used to obtain
a relationship between inputs and outputs, rather than determining the structure of the
system. Although this modeling technique is more practical than the first one, the results
require careful interpretation and validation with the physical concepts.
24
Figure 2.9: System Identification Problem (blackbox type of modeling)
The construction of a model via system identification commonly involves three
steps. The first step is input/output data collection, which is mostly completed through
prior experiments. The second stage is narrowing the model choice to several that fit the
system physical capabilities. The last step is model validation, which involves
performance error measures. If this last step fails to achieve the error requirements, than
the steps are repeated until the desired results are obtained. The data collection process is
explained in detail in another chapter, the following sections describe basics of system
identification standard models. The types of models described in this study are linear
time invariant. Even though these types of systems are limited, the theory developed
through them can be used to approximate real systems.
2.3.2. Parametric System Identification
The parametric model is a set of differential or difference equations that describe
the operation of the system in terms of inputs and outputs. These equations also include a
number of parameters that can be varied to alter the behavior of the model. The values of
the parameters are numerically estimated to give the best agreement between the
experimentally measured output and the model estimated output. The matching criterion
is usually the minimization of the squared error, where error is defined as the difference
between the measured and predicted outputs.
25
Parametric system identification is basically a simplification of general standard
equations for dynamic systems. Figure 2.10 shows a general block diagram of a dynamic
system. Although many are tempted to use a large number of parameters to describe the
system, the number of parameters to identify should be small. The accuracy of
coefficients estimation decreases with the number of the parameters to be estimated
[Ljung, 1999].
Figure 2.10: Generic Dynamic System Block Diagram (discrete time signals)
The general equation (Ztransform) for the dynamic system is:
) (
) (
) (
) (
) (
) (
) ( ) (
1
1
1
1
1
k e
z D
z C
k u
z F
z z B
k y z A
d
−
−
−
− −
−
+ = (2.3)
where,
A(z
1
) = 1 + a
1
z
1
+ … + a
na
z
na
B(z
1
) = b
1
z
1
+ … + b
nb
z
nb
C(z
1
) = 1 + c
1
z
1
+ … + c
nc
z
nc
D(z
1
) = 1 + d
1
z
1
+ … + d
nd
z
nd
F(z
1
) = 1 + f
1
z
1
+ … + f
nf
z
nf
Σ
) (
) (
1
1
−
− −
z F
z z B
d
) (
1
1 −
z A
) (
) (
1
1
−
−
z D
z C
u(k)
e(k)
y(k)
26
The polynomials represent the components used to find the transfer functions (eq. 2.4)
derived from the state space equation of the system behavior. The shift operator z
1
is
consistent with the ztransform and the negative power represents the right shift in
sampletime. In equation 2.3 the term z
d
next to coefficient matrix [B] represents the
time lag between input and output which means that some leading coefficients of [B] are
zero when there is a delay in the system. The order of the polynomials is described by
na, nb, nc, nd and nf. The values of these variables are determined in the process of the
system identification, to better match the behavior of the system. If both sides of the
equation are divided by the feedback term A(z
1
), then:
) (
) ( ) (
) (
) (
) ( ) (
) (
) (
1 1
1
1 1
1
k e
z A z D
z C
k u
z A z F
z z B
k y
d
− −
−
− −
− −
+ = (2.4)
where the input terms next to u(k) can be grouped to form the transfer function G(z
1
) and
disturbance terms next to e(k) form H(z
1
). In other words, G(z
1
) and H(z
1
) are the
transformations of the inputs and disturbances, respectively to obtain the output [Ljung,
1999 Chapter 4].
Table 2.2: Common BlackBox Models, Simplification of General Expression
POLYNOMIALS USED NAME OF THE MODEL
B(z
1
) FIR – Finite Impulse Response (n
a
= 0)
A(z
1
); B(z
1
) ARX – Auto Regressive with eXogenous input
A(z
1
); B(z
1
); C(z
1
) ARMAX  Auto Regressive Moving Average with eXogenous output
A(z
1
); C(z
1
) ARMA  Auto Regressive Moving Average
A(z
1
); B(z
1
); D(z
1
) ARARX  Auto Regressive Auto Regressive with eXogenous output
A(z
1
); B(z
1
); C(z
1
); D(z
1
) ARARMAX – combination of ARARX with Moving Average
B(z
1
); F(z
1
) OE  Output Error
B(z
1
); F(z
1
); C(z
1
); D(z
1
) BJ – Box Jenkins
27
Simplifying the general equation 2.3 or 2.4, there are several types of standard
models that can be developed. Table 2.2 summarizes the special case of a priori type of
modeling techniques. System identification has no restriction on the number of inputs
and outputs to the model. The common use of single/multiple input and output systems
has created a specific nomenclature for each of the cases.
SISO – Single Input, Single Output
MISO – Multiple Inputs, Single Output
SIMO – Single Input, Multiple Outputs
MIMO – Multiple Inputs, Multiple Outputs
The system identification literature describes the solutions and techniques for the single
input, single output models (SISO); however, superposition enables the use of the
techniques for any case.
2.3.3. EMGTorque Relationship Modeling
There are many applications that the tension exerted by the muscle group during
the various activities is useful, however direct measurements are unnatural, invasive,
expensive, and they may also not be possible presently. The assumption of torque being
related to the nervous excitation of the individual muscle or the muscle group, relates
torque to the magnitude of electrical muscle activity (EMG signal). A relation between
EMG and torque simplifies the situation, because EMG is readily obtained by either
surface or wire electrodes depending upon whether the muscle group or individual
muscle measurements are needed [Perry and Bekey, 1981]. Although many studies have
made a great impact in the EMG field, there is no consensus on a standardized set of
28
models that relate a specific muscle (muscle group) to tension (torque). In addition, the
progress in obtaining EMG amplitudes is not yet incorporated into the existing models.
The development of generic prediction models has been less successful, perhaps
due to variations in muscle composition. However, different procedures used to record
and analyze EMG also need to be considered when determining the relationship between
muscular forces and the EMG signal. Several investigators have agreed that it is
necessary to incorporate the control strategy for the muscles being investigated,
including: the force generation rate, joint angle, muscle length, and muscular co
activation [Solomonow et al, 1990]. It is also determined that changes in recording
procedures, including variations in electrode placement, recording configuration and limb
position, significantly alter the EMGtorque relationship of the biceps and triceps brachii
[Woods and BiglandRitchie, 1983].
The interaction of muscles during contractions must be accounted for during
analyses. Principal components have been used to minimize the effects of crosstalk, the
overlapping affects of independent variables, but generalization may not be possible due
to the large number of assumptions and originality of the situations examined [Hughes
and Chaffin, 1997]. In general, predicting torque is difficult because so many factors can
influence the resulting exertion. The muscle being investigated, procedures
implemented, and the form of the forceEMG relationship are vital components for
accurately determining force levels. Various approaches have utilized relatively simple
models under controlled conditions to determine the torque produced by different
muscles groups about different joints.
29
Studies of the relationship between surface EMG and force have found that there
exist both linear and nonlinear relationships. Woods and BiglandRitchie (1983)
investigated the degree of linearity in the torque to EMG relationship and found that
linearity existed for muscles such as the adductor pollicis and soleus. They have also
found that other muscles, such as the biceps and triceps, behaved nonlinearly from 0
30% MVC (maximum voluntary contraction), and then linearly above this range. On the
other side, Moritani and DeVries (1978) determined that a linear relationship existed
between the electrical muscle activity of the biceps brachii and the muscular tensions
produced during exertions. Others have concluded that surface EMG, after processing
using rectification and integration, varies linearly with tension generated at a constant
muscle length or during contractions with constant velocity [MilnerBrown and Stein,
1975].
Characteristics of the muscle of interest may also influence the EMG to torque
relationship. Muscles of uniform fiber composition exhibit a linear relationship while a
random noneven composition of fibers behaves more nonlinearly [Woods and Bigland
Ritchie, 1983]. The main fiber type can also influence the linearity with slow twitch
muscles behaving more linearly as compared to the nonlinear characteristics of fast
twitch fibers [Zuniga and Simons, 1969]. Furthermore, the muscles display nonlinear
behavior at lower torque levels due to selective recruitment of motor units at different
distances from the electrodes. In addition, the dependence on frequency coding (the
frequency of the incoming action potentials) for force modulation in the muscles results
in linearity while muscles such as the bicep brachii recruit throughout the total range of
30
force and behave nonlinearly, with the discontinuity at approximately 30% of the
maximum voluntary contraction [Woods and BiglandRitchie, 1983].
The degree of linearity is dependent on the muscle being investigated, but other
factors must be also considered. MilnerBrown and Stein (1975) suggest sampling bias,
synchronization, and tension nonlinearity also influence the behavior of EMG to torque
relationship. Frequency coding has been shown to increase the linearity [Ray and Guha,
1983], whereas tension, length, and velocity characteristics within muscles are
nonlinearities that affect the overall relationship [Perry and Bekey, 1981]. Moreover,
Zuniga and Simons (1969) determined that there is a nonlinear relationship between
averaged EMG potential and muscle tension. In addition to muscle characteristics, the
electrode arrangement, type of measurement, fatigue, and level of physical conditioning
level may influence the apparent EMG to torque relationship [Zuniga and Simons, 1969].
Recent advances in the research field have demonstrated that linear models can
predict shoulder forces during isometric contractions [Laursen et al., 1998].
Additionally, MilnerBrown and Stein (1975) concluded that there was a simple linear
relationship between surface EMG and force within the first dorsal interosseus muscle of
the hand. However, on the other side, Woods and BiglandRitchie (1983) have found
that under isometric conditions, the relationship between integrated, smoothed, or
rectified EMG and muscle force depends on the physiological characteristics of the
muscle. If the muscle mechanics are known, they can be incorporated into a Hilltype
model that can be used to predict muscle forces [Dowling, 1997]. Linear algebraic
equations may not suffice when attempting to explain dynamic situations. The velocity
of contractions and the tension produced can be related using Hill’s hyperbolic equation
31
[Perry and Bekey, 1981]. All the above show that the degree of linearity depends on the
muscle being investigated, many muscles seem to exhibit a linear relationship between
force and EMG, but nonlinear models seem to capture more of the physiological
behavior. The usage of linear or nonlinear model depends on the focus of the research
work and it is really a matter of perspective of the researchers.
Although clear progress has not been made toward development of generic
models, some of the models developed for specific cases have made impact in the field.
Armstrong et al. (1982) used rectified EMG signals of the forearm flexor muscles to
predict the finger forces produced during tasks involving pinching, grasping and pressing.
Grant et al. (1994) predicted grip force from EMG measures and ratings of perceived
exertions, and reported that as much as 74% of the variation could be explained.
Sommerich et al. (1998) studied typing tasks in an attempt to determine a doseresponse
relationship for general hand intensive tasks and create generic biomechanical
assessments. Buchanan et al. (1993) used surface EMG and anatomical parameters to
estimate isometric muscle forces about the wrist using an EMG coefficient method.
Although there are limitations with this model, including the lack of repeatability and
restriction to “static isometric conditions,” torque at the wrist could be estimated with
coefficients of variation less than 10%.
Several studies have examined muscle torques produced about the elbow. A
multichannel surface EMG approach by Clancy and Hogan (1997) was used to develop a
third order polynomial algebraic relation with an estimation error of approximately 3% to
predict torques about the elbow. Furthermore, a model created by Wyss and Pollak
(1984) approximated muscle forces about the elbow with 10% error. The EMGtorque
32
relationship of abdominal muscles required quadratic regression but still did not account
for all of the variation around a linear regression line [Stokes et al., 1989]. Extensive
work has been conducted on the lumbar musculature during static and dynamic situations
with EMG based models being in the focus [Hughes et al., 1994; McGill, 1992;
Nussbaum et al., 1995].
In summary, there is a substantial amount of work investigating surface EMG to
torque models which confirms the importance of utilizing EMG as a physiologically
powerful tool. The above experimental studies are not constrained only to static
conditions, individual progress has been made establishing both linear and nonlinear
relationships for quasiisotonic (slowly force varying) and even extending to fully
dynamic conditions. While the accomplishments have made an impact in the field, there
are clear problems that still exist in some of these studies. First, most of the earlier (more
than two decades ago) investigators assumed that the antagonist muscle can be safely
neglected, since the mechanical activities of agonist and antagonist muscles are
considered independent from each other. Secondly, even though calibration is a common
practice nowadays, some researchers neglect the importance of it while some others go
beyond and suggest calibrating to each subject separately [Hasan and Enoka, 1985].
The most relevant factor that should remain from this review is the necessity for a
method to obtain accurate torque estimation. There is no consensus on the degree of
linearity, because the findings are influenced by many factors, such as the dynamic range,
the level of the force contraction, and the size of the muscles. The level of the details and
the type (physiological or blackbox) on the various EMGtorque models is also relative
to the focus of the study and it is driven by the main objective, improving the accuracy of
33
torque predictions. Although the same objective is intended, several important
contributions in the literature, such as combination of multiple surface EMG recordings
to improve the SNR and the adaptive whitening filters to improve the statistical
bandwidth, have not yet been used to improve torque estimations.
The experimental data used for the present research thesis are carried over from
previous research. This prior research also modeled EMG amplitude to torque
incorporating both agonist and antagonist muscles during tasks that involved force
varying contraction. However, it avoided several of the earlier mentioned difficulties by
examining constant posture efforts (similar to strict isometric conditions), whereas many
of the above researchers examined fully dynamic tasks. These simplifications were
intended to allow for an assessment of the newly developed EMG amplitude processors.
The results were positive demonstrating a clear improvement in torque prediction when
advanced processing techniques were used to obtain EMG amplitudes. This present
research is anticipated to further investigate the results for more profound knowledge and
to reexamine the encountered model convergence problems.
34
CHAPTER 3. SURFACE EMG TO TORQUE MODEL DESIGN
In the previous chapter there was a brief review of the literature achievements on
EMGtorque relationship. As mentioned, there is not any generic model established yet,
therefore researchers create models that best fit their design application or that are
derived from earlier experimental work (Hilltype model). The primary focus of this
research thesis was not to find the best model, but rather demonstrate the importance of
incorporating the advances of EMG amplitude processing into each model. Hence, the
results presented in later chapters will display the improvements in EMGtorque model
performance as a function of EMG amplitude method. Since the EMGtension
relationship for each individual muscle is not possible, the EMGtorque relation can
alleviate some issues such as measurements for mechanical verification, cocontraction,
and crosstalk. This chapter describes the design process of a linear EMGtorque model
that will be used to compare four different types of processors.
3.1. EMGTORQUE MODEL DESIGN
This section is a summary of the theory involved to design the EMGtorque model.
The concept of torque for the skeletal muscles is derived from the motion of the bones
about a joint due to muscle contractions. Since surface EMG signal measures the activity
of the skeletal muscles, a mathematical relationship can be established between the EMG
amplitude and net joint torque. Using the principles of system identification, the model is
standardized to a parametric type ARX (FIR) model.
35
3.1.1. Physical Interpretation of the EMGTorque Model
The level of the tension produced in the muscle is controlled through the
recruitment of motor units and their firing rate adjustments. The motor recruitment is
hypothetically
1
done orderly based on the size of the muscle fibers. For tasks that involve
slow force variations, as the tension level varies from low to high, the low frequency
motor units are the first to be activated while the ones with high minimum frequency are
the last [Hannerz, 1974]. The tension developed by the muscle also depends on both the
conduction velocity and the geometry of the muscle fibers. The assumption that
muscular force depends only on the firings of the motor units (rate and number of units)
makes the relation between EMG to force nonlinear in a sense that the number of MUs
recruitment is higher for the higher contraction levels. As mentioned, the dependence on
the conduction velocity and the geometry dependent parameters also contribute to the
nonlinearity.
The surface EMG is a noninvasive, easily obtained, measure of electrical activity
in the skeletal muscle. Since both electrical and mechanical activities are mutually
related through several mentioned physiological parameters, relating EMG amplitude to
muscle tension would be ideal. However, there are two fundamental issues with this
model. First, it is extremely difficult to measure EMG from only one muscle. In
practice, surface mounted electrodes capture EMG generated from the muscles in a
surrounding area, which are not necessarily the ones under investigation. This condition,
known as crosstalk and already discussed in Section 2.1.3, is one of the main factors
affecting EMG signal interpretation. Additionally, there is no practical method for
1
Based on the EMG models created for isometric low force level contractions
36
accurately measuring the tension provided by an individual muscle. These two factors
prevent the use of EMG amplitude to force models in terms of isolated muscle
contribution as a method for evaluating performance of an EMG amplitude estimator
[Clancy and Bouchard, 2001].
Another mechanical activity commonly related to EMG is the torque produced
about a joint as muscular force is exerted. The problem of crosstalk is still present.
Although unlike in the case of EMG to individual tension relation, it is not as influential
to the net torque estimates [Clancy, Morin, and Merletti, 2002]. In addition, the model
performance can be easily quantified by comparing the torque estimates to the actual
torque about the joint that can be measured using a dynamometer. The prediction of the
net torque requires the usage of both agonist and antagonist muscles. Muscles that
perform a desired action are known as agonist muscles, whereas those that oppose the
action are antagonist.
Some researchers have separated the contributions of agonist and antagonist
muscles, assuming that the agonist muscles are inhibited while the antagonist ones are
contracted. Doing so, the net torque is a result of the inhibition or agonist muscles
[Lawrence and De Luca, 1983; Vredenbregt and Rau, 1973; Woods and Biggland
Ritchie, 1983; Zuniga and Simons, 1969]. On the other hand, Hasan and Enoka (1985)
have experimentally determined the existence of cocontraction in contraction levels
exceeding 20% MVC. Therefore, in the cases of 50% MVC it is necessary to
acknowledge the contribution of both agonist and antagonist muscles.
37
Figure 3.1: Raw Surface EMG to Torque Model [Clancy and Hogan, 1997]
Four surface electrodes affixed on top of the muscles (biceps and triceps) record the EMG signals. After amplified,
filtered, and sampled they are applied to the EMG amplitude processor. The EMG amplitude estimations for
flexion F(n) and extension E(n) are decimated to obtain F(k) and E(k) respectively. The amplitude estimates are
used as two inputs to a system identification algorithm to predict the net torque about the joint (T).
EMG Extension 4
EMG Extension 3
EMG Extension 2
EMG Extension 1
EMG Flexion 4
EMG Flexion 3
EMG Flexion 2
EMG Flexion 1
System ID Estimated Torque EMG amplitude
Raw EMG Recordings
F (k)
E (k)
Flexor
EMG
Amplitude
Decimator
Extensor
EMG
Amplitude
Decimator
F (n)
E (n)
Flexor
EMG
Amplitude
Estimator
Extensor
EMG
Amplitude
Estimator
Flexor
EMG
Amplitude to
Torque
Estimator
Extensor
EMG
Amplitude to
Torque
Estimator
Σ
T
F
T
T
E
+

38
Figure 3.1 shows a block diagram modified from the diagram from Clancy and Hogan
(1997) and represents the model that is used to predict torque about the elbow joint from
biceps and triceps muscle groups considering both agonist and antagonist muscles. Even
though the contributions of flexion and extension are attributed to agonist and antagonist
muscles, the model is designed based on their algebraic sum, rather than their
independent contributions. The EMG amplitude processing (first stage from the left of
Figure 3.1) and the data preprocessing (second stage in the same figure) stages will be
discussed in detail in subsequent chapters. For now, it is assumed that the data are
available and ready to be used in the EMGtorque model.
3.1.2. Mathematical Modeling for EMGtorque
Several studies, as mentioned in section 2.3.3, have determined that there exists a
mathematical relationship between EMGtorque (linear or nonlinear). The internal
change in the muscles may be produced by processing the EMG signal [Perry and Bekey,
1981]. It is not clear whether nonlinear or linear models are the best choice. The linear
models are widely used because of the simplicity in their design associated with the linear
least squares solution [Inman et al, 1952; Thelen et al. 1994; Clancy et al., 2001],
whereas the researchers using nonlinear models argue that they better describe the EMG
physiological nature [Solmonow et. al, 1986; Vredenbregt and Rau, 1973; Woods and
BigglandRitchie, 1983; Zuniga and Simons, 1969].
Without generalizing the model results, Gottlieb and Agarwal (1977) related EMG to
torque using the following transfer function:
1
1
1
1
) (
) (
) (
2 1
+
⋅
+
= =
s T s T
k
s E
s F
s G
RA
(3.1)
39
where E
RA
is the EMG amplitude obtained from the averaged rectifier output [Gottlieb
and Agarwal, 1977]. This model is the Laplace transform of the differential equation
with two degrees of freedom (time constants) and it assumes a continuous time domain
system/signal, but it is agreed that it is easier to work with difference equations that
represent the system for discrete time signals (samples). The conversion amongst them is
straightforward if the sampling rate is known. Another limitation is that the model can
be used only for “isometric” tasks, because it does not include any physiological
parameters (e.g. joint angle).
The model used for the purpose of this thesis is created by modifying the EMG
torque (ARX) model used by Stephan Bouchard (2001). The proposed model is an FIR
(zerosonly) model and the degrees of freedom depend on the order of the system.
Bouchard’s model includes the feedback matrix of the ARX model, which is equivalent
to the poles of a system. Even though it may require higher order to capture all the
dynamics of a system, a zerosonly model is mathematically possible and easier to
implement. The model form is written as:
( ) ( ) ( ) ( )
( ) ( ) ( ) nb k F f k F f k F f
nb k E e k E e k E e k T
nb
nb
− + + − + − +
− − − − − − − =
2 1
2 1
2 1
2 1
(3.2)
Referring to Figure 3.1 that shows the complete EMGtorque modeling process and the
above equation, the decimated flexion [F(k), where k is the decimated discretetime
sample index] and extension [E(k)] EMG amplitudes were related to torque [T(k)] via the
dynamic, linear, FIR model [Ljung, 1999, pp. 80–83]. In equation 3.2, the e
i
represent
the extensor model coefficients, the f
i
are flexor model coefficients. The nb is the model
order as defined in section 2.3.2. The model mathematical expression in 3.2 does not
40
change for the timevarying EMG amplitudes that are not decimated, therefore only the
final results are presented.
3.2. MODEL SOLUTION
The model represented in equation 3.2, is the FIR version of the ARX model. This
model is obtained using the standard linear difference equation of the ARX model that is
found using the information given in Table 2.2 [Ljung, 1999]. Ljung writes the general
model as:
y(t) + a
1
y(t1)+… + a
na
y(tna) = b
1
u(t1)+...+b
nb
u(tn
nb
) + e(t). (3.3)
The adjustable parameters in the equation are put in matrix form
θ = [a
1
a
2
… a
na
b
1
b
2
… b
nb
]
T
.
(3.4)
The FIR model used for EMGtorque is the special case of ARX, obtained by setting the
parameter na=0. The system is dual input – single output, therefore the equation is
adapted for the MISO case. In equation 3.2, the coefficients e
1

nb
and f
1

nb
represent
Ljung’s b
1nb
coefficients. Analogous coefficient matrix is given by:
θ = [e
1
 e
2
… e
nb
f
1
f
2
… f
nb
]
T
(3.5)
The vector of the known input EMG amplitude samples (decimated) are represented in the
φ(k) vector. If the measurements are repeated over time then the vector φ(k) becomes a
matrix with N rows, where N is the number of samples. In general N >> nb for the system
identification to be possible. The inputs to this EMGtorque model φ(k) are written as:
− + − + + − + − + +
− + + − + +
− − − −
=
) ( ... )... 1 ( ) ( ) ( .. )... 1 ( ) (
... ... ... ... .... ....
) 1 ( ... )... ( ) 1 ( ) 1 ( .... )... ( ) 1 (
) ( ... )... 1 ( ) ( ) ( .... )... 1 ( ) (
) (
nb N k F N k F N k F nb N k E N k E N k E
nb k F k F k F nb k E k E k E
nb k F k F k F nb k E k E k E
k ϕ
41
The output matrix T(k) for N samples is given by:
+
+
=
) (
....
) 1 (
) (
) (
N k T
k T
k T
k T (3.6)
The problem involves solving the equation 3.7 to obtain the coefficient matrix.
T
k k T θ ϕ ⋅ = ) ( ) ( (3.7)
The solution can not be obtained by taking the inverse of the data matrix, [φ(k)]
1
, because
the inverse of the matrix exists only for square matrices (in this case the number of the
unknowns exceeds the number of the existing linear equations). Therefore there is not a
unique solution, but rather a method referred to as linear squares error minimization of
error can be used to obtain the “bestfit” coefficient matrix. The optimal value for θ is
referred to as θ
ˆ
and is given as:
) ) ( ) ( ( min arg
ˆ
θ ϕ θ
θ
⋅ − = k k T
)] ) ( ) ( ( ) ) ( ) ( [( min arg θ ϕ θ ϕ
θ
⋅ − ⋅ − = k k T k k T
T
] ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( [ min arg θ ϕ ϕ θ ϕ θ θ ϕ
θ
k k k T k k k T k T k T
T T T T T T
+ − − = (3.8)
Applying vector calculus concepts, the minimum is denoted as the value at which the
gradient of the matrix is zero. Hence, the minimum value of θ
ˆ
is found by computing
the gradient of the difference between the measured output and the predicted output
T(k)φ(k)θ. Therefore the coefficient matrix producing minimum error is equal toθ
ˆ
.
The gradient of the error matrix is expressed as:
θ ϕ ϕ ϕ θ ) ( ) ( 2 ) ( ) ( ) ( ) ( ) ( k k T k k T k k T E
T T T
+ − − = ∆ (3.9)
42
Equating the expression to zero, and solving for θ, yields to the minimization matrix in
the least squared sense.
1
)] ( ) ( [ )] ( ) ( [
ˆ
) ( ) ( 2
ˆ
) ( ) ( 2
0
ˆ
) ( ) ( 2 ) ( ) ( ) ( ) (
−
⋅ =
= ⋅
= ⋅ + − −
k k k k T
k k T k k
k k k k T k k T
T T
T T
T T T
ϕ ϕ ϕ θ
ϕ θ ϕ ϕ
θ ϕ ϕ ϕ ϕ
(3.10)
After calculatingθ
ˆ
, an estimate of the torque about a joint can be calculated as the
product of the EMG amplitude estimates from both extensor and flexor muscles andθ
ˆ
.
In practice, the coefficient matrix θ
ˆ
is computed using a “training” dataset, then the
resulting θ
ˆ
is used to estimate the torque produced by another independent “test” dataset.
The difference between the estimated torque and the real torque in the test dataset is the
error and it is used to quantify the model performance. More on the testing procedure
will be described in subsequent chapters.
43
CHAPTER 4. DATA COLLECTION AND ANALYSIS METHODS
This chapter explains the data collection method that followed some of the described
literature suggestions. Then, it continues by describing the process of obtaining EMG
amplitudes from four different processors. The EMG amplitude estimates are also
decimated to improve the model performance by eliminating erroneous spikes existent in
estimated torque as will be in detail discussed later. The entire preprocessing procedure
is presented in a block diagram form and it is utilized as necessary during performance
testing. System identification involves two main steps, training and validation. During
training, a coefficient vector is fit to the input data based on the least squares error
minimization. Model validation requires utilizing a distinct dataset to estimate the output
using the optimal coefficients. The traintest along with model performance measures
procedure is explained in detail in the present chapter.
4.1. EMG DATA COLLECTION
Recording the EMG signal using surface electrodes faces challenges related to the
signal power and external/internal noise. As explained earlier, some of the internally
generated noise found in EMG system can neither be eliminated nor reduced. However,
errors due to recording devices, electrode placement, skin effects, and some other
external error sources (2.1.3) can be reduced by taking the necessary precautions. This
data collection section explains some of the potential noise sources along with some
suggestions for minimization and then it continues by describing the experiment
conducted to collect the EMG data used in this project. The description of the apparatus
and experimental procedure is brief, since the experiment is not part of this thesis
44
contribution. More details can be found elsewhere [Clancy, 1999; Bouchard, 2001;
Clancy and Farry, 2000].
4.1.1. Noise Reduction Precautions
One of the most important sources of error in EMG recording is the placement of
the electrodes. The tight spacing of the electrodes in a multiple channel recording can
produce correlated signals, increases the chance of short circuits, and enables electrical
coupling. In addition, the PSD of the recorded EMG signal depends on the location of
the electrode in relation to the innervation zone (MUAP generation site) the myotendon
zone and the lateral edge of the muscle [De Luca, 2002]. The center position between the
innervation on the top and myotendons on the bottom is preferred (see Figure 4.1).
Figure 4.1: EMG Electrode Placement [De Luca, 2002]
45
The principle of EMG electrode functionality derives from a layer of charge
created in the interface between the metal of the electrode and an electrolyte solution.
The charge layer creates a potential gradient that translates into voltage picked up by the
electrode. The voltage is dependent on the type of the electrode material, hence it is
crucial to use electrodes of the same material to minimize the potential difference
(especially in a multichannel application). In addition, some of the power of the initial
EMG signal is lost due to the effect of skinelectrode impedance. To reduce its
impedance, skin is prepared using conductingpaste, rubbing alcohol, or lipid solvents.
The amplifier is also chosen to have input impedance at least 100 times more than
expected electrodeskin impedance, thus reducing the power loss even more. [Clancy,
Morin, and Merletti 2002]
Surface electrodes suffer from motion artifact due to displacement and
deformation (stretch) of the underlying skin. Both effects are minimized by cleansing the
skin with solvents, rubbing a conductive paste, and affixing the electrodes carefully prior
to recording EMG. Motion artifacts can also be reduced through signal conditioning both
online and offline. Since the typical power density of these types of motion artifacts is
below 20 Hz, they can be largely attenuated using highpass filter at that frequency either
in the processing software or integrated in the hardware [Clancy, Morin, and Merletti
2002].
The cables of the electrodes have an intrinsic capacitance, which if exposed to a
varying magnetic field (electric field) can produce alternating currents (150Hz). The
cumulative effect of the voltage created by the product of the skinelectrode interface
impedance and the displacement current in addition to the magnetic coupling in some
46
cases is comparable to the magnitude of the real EMG. Shielding the cables can improve
the magnetic environment, although electrostatic discharges spreading through
measurements can be a harmful side effect [Clancy, Morin, and Merletti 2002]. The
usage of active electrodes is a better solution, because the voltage buffer transforms high
impedance at the electrode to low impedance at the output where the signal is fed to the
cables [De Luca, 2002].
The interference of power line current and its harmonics (60 Hz in U.S. and 50 Hz
in Europe) can have power densities larger than EMG itself and are a commonly
appearance in EMG recordings. The effect of the power lines can be reduced by notch
filters at 60Hz or by differential amplifiers that have a high common mode rejection ratio
(CMRR). The second is preferred, because the notch filters omit real signal in addition to
the noise. A CMRR of 90 dB is generally advised in the literature, even though current
technology can provide a CMRR of 120dB. The reasons for not using amplifiers with
120dB at CMRR are the price, their stability, and the power line signal may not be in
phase [De Luca, 2002].
4.1.2. Apparatus and Experimental Procedure
ii
The experiment conducted to collect this set of EMG data was consistent with the
suggestions mentioned in the previous section (4.1.1) and literature [De Luca, 2002;
Merletti, 1999; Clancy, Morin and Merletti, 2002] for noise minimization. The data
collection experiment was conducted previously; thus, the experiment description is
summarized from other sources [Clancy, 1999; Bouchard, 2001; Clancy and Farry,
2000]. The subjects had signed consent for their participation and proper human studies
ii
Parts of this section are taken from the paper submitted to the Journal of Biomechanics (attached to
Appendices) authored by Clancy, E.A; Bida, O.; Rancourt, D.
47
permission was taken prior to the experiment. The apparatus (Figure 4.2) used to collect
the EMG data was a Biodex exercise machine (Biodex Medical Systems, Inc., Shirley,
NY). Each of the subjects were seated in a lightly cushioned seat and secured using
seatbelts. The subject's right arm was positioned in the plane parallel to the floor, with
the shoulder abducted 90
o
, the forearm oriented in the parasaggital plane, the wrist fully
supinated and the elbow flexed 90
o
(Figure 4.4). The wrist was rigidly attached to the
Biodex dynamometer with a cuff at the styloid process. The position of the dynamometer
was maintained throughout the entire experiment [Clancy, 1999].
Figure 4.2: Biodex Exercise Machine for the Experiment [Bouchard, 2001]
48
The skin above the investigated muscles was cleaned with an alcohol wipe and a
small amount of conducting paste was used to rub the subject’s arm. The EMG active
amplifying electrodes (Liberating Technologies Inc. model MYO115, Holliston, MA)
were placed over each of the biceps and triceps muscles, midway between the elbow and
the midpoint of the upper arm, centered on the muscle midline. The two contacts (4 mm
diameter, stainless steel, separated 15 mm centertocenter) of each electrodeamplifier
(Figure 4.3) were oriented along the muscle’s long axis. Adjacent electrodeamplifier
centers were spaced 1.75 cm apart, transversely across the arm. The ground electrode
was applied over the acromion process. Each electrodeamplifier had a gain of 725, a
common mode rejection ratio of 90 dB at 60 Hz, and a secondorder 10–2000 Hz
bandpass filter.
Figure 4.3: Surface EMG Electrodes and Acquisition Box [Bouchard, 2001]
The output from each of the electrodeamplifier was electrically isolated,
amplified, and low pass filtered (fourthorder filter at 2000 Hz). Amplification stage
49
contained a negative gain configuration standard opamp with a selectable gain (1 to 25).
Recordings with the two contacts of each electrodeamplifier shorted gave a measure of
equipment noise, which averaged 2.1±1.7% of the root mean square EMG at 50%
maximum voluntary contraction (MVC). The EMG and dynamometer signals were
sampled at 4096 Hz using a 16bit A/D converter (Computer Boards model CIO
DAS1600/16, Mansfield, MA).
Fifteen healthy subjects (eight male, seven female; aged 23–65 years) each
completed one experiment.
Table 4.1: Subject Information (Code, Age, and Gender)
SUBJECT CODE (Exp. LB) AGE GENDER
02 31 F
03 49 M
05 29 F
07 65 M
08 43 F
09 60 M
10 41 F
12 62 F
13 50 M
16 28 M
17 58 F
18 41 M
19 31 F
20 23 M
21 65 M
Subjects initially performed two 2 second MVCs each in flexion and extension, the
averages of which were used as the subject’s MVCs for the experiment. Next, they
performed a 0% MVC (rest contraction) and separate flexion and extension 50% MVCs
for five seconds, utilizing force feedback on a computer screen. These contractions were
used to calibrate the advanced EMG amplitude processors [Clancy and Farry, 2000].
50
The subjects then performed dynamic (constantposture, forcevarying) target
tracking contractions positioned as in Figure 4.4. A computer screen displayed either
their elbow joint torque (the dynamometer signal) or the algebraic difference between
realtime biceps and triceps EMG amplitude, as a biofeedback signal. The EMG
amplitude difference provided a biofeedback signal that was similar in character to the
torque feedback, albeit with increased variance. The computer produced a second
“pursuit” display on the screen which moved randomly between 50% MVC extension
and 50% MVC flexion. The random pursuit profile followed a uniform random
distribution with a bandwidth of either 0.25 Hz (slow tracking) or 1 Hz (fast tracking).
Figure 4.4: Subject During Experiment [Bouchard, 2001]
51
Subjects completed 15 slow tracking trials (three sets of five) and 15 fast tracking trials
(three sets of five), each of 30 s duration. The subject’s arm was removed from the wrist
cuff between all recording trials to allow 2–3 minutes of rest to avoid fatigue.
4.2. EMG AMPLITUDE ESTIMATION METHOD
For the present study, all data analysis was performed offline using MatLab (The
Mathworks, Natick, MA). In order to determine the influence of amplitude processors on
torque estimation, four different EMG amplitude (EMGamp) processors were compared.
In each case, an amplitude estimate was produced in six stages separately for the biceps
and triceps muscle groups (Figure 2.5). The design of the EMG amplitude processor is
explained in detail in the background chapter, while this section describes the experiment
in a more practical perspective. Before creating amplitude estimates from the raw EMG
data (stored in forms of A/D channels for each of the electrode locations), it was
necessary to calibrate the noiserejection, adaptive whitening, and spatial uncorrelation.
The calibration requires the additive noise signal, which in this case is the 0% MVC
signal recorded while subjects were fully resting. This recording is assumed to capture
the 60 Hz noise as well. The electrode locations encoded in the channel number are
given in Table 4.2 and further description is found in the APPENDICES: (I). The A/D
channels 14 are used to estimate biceps EMG amplitude and channels 811 are used to
estimate triceps EMG amplitude in multiple channel processing. In the case of single
channel processing, channels 2 and 9 are used for flexion and extension EMG
amplitudes, respectively.
52
Table 4.2: A/D Electrode Channels from the Experimental Data
A/D Channel Contents
1 EMG: Biceps, most lateral location
2 EMG: Biceps, lateral center
3 EMG: Biceps, medial center
4 EMG: Biceps, most medial location
8 EMG: Triceps, most medial location
9 EMG: Triceps, medial center
10 EMG: Triceps, lateral center
11 EMG: Triceps, most lateral
16 Dynamometer
For all processors (Table 4.3), the EMG data were first highpass filtered at 15 Hz using a
noncausal, effective 10
th
order, FIR filter. The detection was performed with an absolute
value operation (MAV, d=1) and the smoothing stage was omitted, since smoothing was
incorporated within the subsequent preprocessing step.
Table 4.3: Four Processors Types (Processor 14)
Single Channel Unwhitened (SCHUNWHIT)
Single Channel Whitened (SCHWHIT)
Multiple Channel Unwhitened (MCHUNWHIT)
Multiple Channel Whitened (MCHWHIT)
Processor type 1 is the singlechannel, unwhitened processor. The raw EMG signal for
each muscle group from one of the electrodes located centrally on the muscle was
digitally highpass filtered and rectified. Processor 2 is a singlechannel, whitened
processor obtained by adaptively whitening the same electrode channel as Processor 1
before the rectification stage. The adaptive whitening technique is explained briefly in
section 2.2.4 and it has been implemented in a standalone MATLAB toolbox [Clancy,
2004]. Processor 3 is a fourchannel, unwhitened processor. After rectification, the four
EMG signals from a muscle group were normalized in magnitude and ensemble
53
averaged. Processor 4 is a fourchannel, whitened processor. Each of the four channels
from Processor 3 was adaptively whitened prior to rectification. The settings for optional
properties for the amplitude estimation function can be found in APPENDICES: (II).
4.3. SYSTEM IDENTIFICATION PROCEDURE
After estimating the amplitudes offline using the EMG toolbox, the data were further
processed as necessary to improve the model performance [Ljung, 1999 pp. 386]. The
torque estimates were then obtained using a train and test procedure, where the
coefficients were computed using a train dataset and were validated using another dataset
referred to as the test dataset. The performance of the model design was evaluated using
conventional measures that will be discussed in detail in this section.
4.3.1. Data PreProcessing
Some of the data manipulation to achieve the desired results involves
normalization to % MVC and A/D offset subtraction for torque as well as decimation and
truncation of both EMGamp(s), and torque (Figure 4.5). Normalization to maximal
voluntary contraction % MVC is a common practice when trying to relate EMG to torque
[Merletti, 1999]. The EMG amplitudes estimates were to % MVE (where 100% MVE 
EMG amplitude level corresponding to 100% MVC) in extension/flexion units; hence, it
was only necessary to normalize torque to %MVC according to either flexion or
extension.
54
Figure 4.5: Block Diagram of EMG Data Preprocessing for System ID Algorithm
The next step common to both torque and EMG amplitudes (EMGamps) in Figure
4.5 is decimation. Because the raw EMG data were sampled at 4096 Hz, the EMG
amplitude estimates were also produced at 4096 Hz. The flexion (biceps) and extension
(triceps) EMGamps are the inputs to a system identification model in which elbow joint
torque is the output. The torque output, however, has signal power over a much lower
band of frequencies than raw EMG. Thus, the high sampling rate is unnecessary.
Therefore, the EMGamps are decimated. The decimation rate was found by evaluating
various integervalued decimation rates from 1–900. In each case, the data were also low
pass filtered with a cutoff frequency equal to half the new sampling rate (8
th
order
Butterworth filter applied in the forward, then the backward time directions to achieve
zerophase) prior to downsampling.
The data truncation was used to eliminate the transient effect of the filters on the
signals (a.ka. a “startup” transient). The effect is seen as corruption only at the beginning
if causal filters are used and in the case of noncausal filters both at the beginning and at
Raw EMG
(4 channels)
Flexion
Raw EMG
(4 channels)
Extension
Measured
Torque
Amplitude
Estimator
(EMGamp)
Amplitude
Estimator
(EMGamp)
Normalize
to %MVC
(flex. / ext.)
Decimation
&
Truncation
Decimation
&
Truncation
A/D Offset
Subtraction
&
Decimation
System ID
&
Truncation
T(k)
55
the end. Adding up the effect based on the orders of filters in each of the amplitude
processing stages, it is necessary to subtract 500 ms on each side prior to system
identification (ID) and 500 ms after it. The number of samples subtracted was based on
the new decimated sampling rate. Additionally, A/D converter offset was subtracted
from the torque signal to account for imbalance in the dynamometer output.
4.3.2. Torque Estimation Procedure
The EMG data used for this project were previously collected as described. Data
were collected from fifteen subjects using a multiple channel data acquisition system. In
addition to four channels of EMG amplitudes for each of the muscle groups, the torque
about the elbow joint was measured using a dynamometer. As mentioned, for each
subject, three sets of data recordings were performed. Within each set of three, data were
recorded five times using various feedback mechanisms for controlling elbow torque.
Each of these five subsets was recorded at two different tracking speeds (0.25 Hz and 1
Hz) totaling to 3 × 5 × 2 = 30 sets of EMG data for each subject.
Referring to the model solution in section 3.2, it is required to fit a set of
coefficients to the input data matrix (EMGamps for flexion and extension) to estimate
torque. For convenience the model mathematical expression is repeated here:
( ) ( ) ( ) ( )
( ) ( ) ( ) nb k F f k F f k F f
nb k E e k E e k E e k T
nb
nb
− + + − + − +
− − − − − − − =
2 1
2 1
2 1
2 1
(4.1)
The optimal fit coefficients were determined using a linear least squares solution. The
input data matrix φ(k) in section 3.2 was created using the toeplitz routine to increment
the samples of amplitude estimation (flexion and extension) and the model order as in
equation 4.1. The system solution involves matrix inversions that can be implemented
56
using QRfactorization (qr) or the pseudoinverse (pinv). QRfactorization expresses the
matrix as the product of a real orthonormal or complex unitary matrix and an upper
triangular matrix. PseudoInverse is the process of inverting an overdetermined linear
system (matrix has more rows than columns). The advantage of using the QR
factorization method is computation simplicity due to elimination of redundancy
[Kolman and Hill, 2001].
Since the results and the computation time for both methods were the same, only
the results from pseudoinverse techniques will be discussed. The model coefficients are
determined utilizing a traintest evaluation routine. Specifically, coefficients were fit to
the input data from a training trial and then used to compute the torque via equation 4.1
from a different test trial. The error signal is defined as the difference between the
measured and the model estimated torques. If all possible combinations are used for the
data combination, there would be a total of 30 × 29 combinations of training and test data
sets. Because of the large computation time, a comprehensive crossvalidation is
sufficient for the scope of this research. Therefore, the 15 trials at a given tracking speed
were organized as three sets of five contractions. (The two speeds are evaluated
separately) Within a set, optimal coefficients were trained to one trial, and then tested on
the four remaining trials. There were a total of 180 error signals available per type of
processor (15 subjects × 3 sets per subject × 1 training trial per set × 4 test trials per
training trial). As mentioned, one second of data from the beginning and end of each
error signal was removed (trimmed), since these data were corrupted by the startup
transients of the various processing filters.
57
In summary, the following parameters were varied in the computations performed
using MatLab (R13, Version 6.5):
• EMG amplitude estimator type: singlechannelunwhitened, singlechannel
whitened, multiplechannelunwhitened, multiplechannelwhitened
• Integervalued decimation rate subjectively chosen from 1 to 900
• Matrix inversion method either pinv. or qr factorization
• EMGtorque model order from 1 to 60
The performance of the model implemented for this research matches the performance
obtained using the builtin ARX parametric model function in MatLab System ID
toolbox. The toolbox is suggested to be utilized in the future, when more complicated
models are necessary.
4.3.3. Model Performance Measures
The resulting EMGtorque error signal from 180 combinations for each processor
type was investigated in several ways. The error mathematical expression used is:
( )
∑
=
− =
N
k
k T k T k Error
1
) ( ) (
ˆ
) ( (4.2)
where N is the sample duration of the truncated estimated torque ) (
ˆ
k T and measured
torque T(k). The measured torque was truncated to equate the number of samples with
the estimated torque. All errors were normalized to twice the torque at 50% flexion
MVC, denoted %MVC
F
. Throughout this thesis results of model performance will be
evaluated using two main time domain error measures: the percent mean absolute error
(%MAE) as computed for each trial
( ) ) ( 100 % k Error mean MAE ⋅ = , (4.3)
58
and the percent variance accounted for (%VAF), defined by Kearney [Kirsch, Kearney,
Crago, 1994] as:
( )  
( )




¹

\

− ⋅ =
∑
∑
=
=
N
k
N
k
k T
k Error
VAF
1
2
1
2
1 100 %
.
(4.4)
The power spectral density (PSD) of each error sequence was another method used to
evaluate the error frequency distribution and identify the character of the error. The PSD
was estimated using Welch periodograms (Hamming window, 1024point FFT, 50%
overlap).
Figure 4.6: System Identification Procedure [Created based on Ljung, 1999]
START
Experimental
Model Design
System ID
Data PreProcessing
Model Structure
Parameter Estimation
Model
Validation
END
YES
?
NO
Does model meet criteria?
Physiological Interpretation
& Performance Measures:
%VAF, %MAE, PSD of ERROR
59
The final stage of the system identification (System ID) procedure is the model
validation. During this stage, the model performance results (%VAF and %MAE) are
interpreted in the physiological sense of the real system behavior. If the expectations are
not satisfied, then there is need for model redesign (Figure 4.6).
60
CHAPTER 5. PROJECT RESULTS
There were several important results derived after completing the described test
procedure. This chapter starts with the decision on the final decimation rate as a solution
to model nonconvergence problems that were observed during prior research. Then, it
continues with a graphical description of the model performance where the outcomes are
contrasted for the four types of processors.
5.1. DECIMATION
The process of decimation includes lowpass filtering and downsampling.
Decimation is used if the system is oversampled and if it contains high frequency noise
components. The necessity for decimation prior to system identification was determined
because some of the prediction torque sequences exhibited a few large “spikes.” The
observed errors were referred to as spikes because they had much larger amplitude
compared to the test torque and lasted only a few samples (see Figure 5.1). The spikes
occurred infrequently (~15% to 20% of combinations) but their magnitude caused the
overall %VAF (and MAE) to be unrealistic. During prior EMGtorque work, the trials
exhibiting the described error were considered to be “nonconvergent” [Clancy et al.,
2001]. Figure 5.1 shows a typical example of observed torque spikes. The time domain
torque plots are on the left of the picture. The amplitude of torques spiked to 20, 000 %
MVC
F
when there was no decimation (c.f., time sample about 2*10^4 when d=1).
61
0 2 4 6 8 10 12
x 10
4
2
0
2
x 10
4
Subject 07, Trial 47, order nb=15
T
a
n
d
T

h
a
t
0 500 1000 1500 2000
100
0
100
200
EXTENSION Coeffcients Frequency Response
C
f

G
a
i
n
(
d
b
)
0 1000 2000 3000 4000 5000
2000
1000
0
1000
T
a
n
d
T

h
a
t
0 20 40 60 80 100
100
0
100
200
d = 20 and VAF = 3431732.8592 C
f

G
a
i
n
(
d
b
)
0 500 1000 1500
0.5
0
0.5
1
T
a
n
d
T

h
a
t
0 5 10 15 20 25 30 35
100
50
0
50
d = 60 and VAF = 87.7499 C
f

G
a
i
n
(
d
b
)
0 200 400 600 800 1000
0.5
0
0.5
1
Time (samples)
T
a
n
d
T

h
a
t
0 5 10 15 20
100
50
0
50
d = 100 and VAF = 87.755
Frequency (Hz)
C
f

G
a
i
n
(
d
b
)
Real
Estimate
Torque %MVCF (Time Domain)
d = 20 and VAF = 3431732.8592
d = 60 and VAF = 87.7499
d = 100 and VAF = 87.755
d = 1 and VAF = 134277469.6018
d = 1 and VAF = 134277469.6018
Figure 5.1: Changes of Predicted Torque while Increasing Decimation Rate
The figure above shows the transformations of the torque predictions as a function of the
sampling rate. On the left side of the figure, the torques both predicted (red dashed line)
and measured (blue solid line) are plotted versus sample time for four decimation rates
(top to bottom: 1, 20, 60, and 100). On the right side of the figure, the gains of the
coefficients computed during the system identification (extension) are plotted for each of
the decimation rates. The purpose of this figure is the graphical presentation of the
observed spikes. When the decimation rate is d=1 the torque amplitude exceeds 10
4
and
the coefficients gain exceed 100 dB outside the system band (~ 410 Hz). As the
62
sampling rate is increased to d = 100 the model performance improves (%VAF =
87.755), the spikes disappear, and the left side gain plot shows the expected LPF like
performance with a max gain of 0 dB. During previous research, where the spikes were
initially observed, a decimation rate of 20 was utilized, but as seen in this figure that rate
was not enough.
Further investigation demonstrated that the errors were related to sampling rate. The
raw EMG sampling rate of 4096 Hz was more 1000 times the natural bandwidth of the
EMGtorque system that contained 99.9% of the power within about 4 Hz, even at fast
tracking bandwidth. Figure 5.2 shows the average torque power accumulation rate for
fast tracking bandwidth.
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
30
40
50
60
70
80
90
100
Frequency (Hz)
%
A
c
c
u
m
u
l
a
t
e
d
o
f
T
o
r
q
u
e
P
o
w
e
r
Fast Tracking: Average Accumulation PSD of Torque
Figure 5.2: Signal Power Accumulation (average PSD torque) vs. Frequency
63
The occurrence of the noise spikes was associated with the cases that the system was
calibrated with torque trials (training sessions to determine fit coefficients) that contained
no power beyond 4 Hz (99% chance). If the test trials were contaminated with noise
extended to larger frequencies, the system produces models with unrealistically high gain
at frequencies above 4 Hz. The high gains beyond the system band are shown for a
typical example on the right side of the picture (Figure 5.1). Hence, even a small amount
of noise power at frequencies above 4 Hz in a test trial caused a noise spike in the
predicted torque. Although the occurrence was infrequent because most of the torque
power was contained within 4 Hz, they were an obstacle to further improvements. This
phenomena encountered with oversampling is also described as common in the system
identification literature [Ljung, 1999].
The solution to this problem was to decimate the EMG amplitude signals prior to
performing the system identification. Progressively lowering the effective sampling rate,
the spikes reduced both in occurrence rate and magnitude. A factor of decimation of 100
(effective sampling rate of 40.96 Hz) extinguished all observed spikes. As seen from the
%VAF values plot in Figure 5.3 the performance of the EMGtorque saturates for rates
above 100 and eventually decreases as the decimation rate exceeds 400. This eventual
drop in the performance is due to the decrease of the cutoff frequency of the lowpass
filter included in the decimation process. The low cutoff frequencies eventually remove
signal power within the natural bandwidth. The optimal sampling rate of 40.96 Hz is
approximately 10 times the highest signal frequency, which is consistent with the rate
recommended by Ljung (1999) and captured all of the signal power. Concluding, the
decimation rate of 100 is used to obtain the consequent results.
64
All Vaf(s)
1.00E+02
8.00E+01
6.00E+01
4.00E+01
2.00E+01
0.00E+00
2.00E+01
4.00E+01
6.00E+01
8.00E+01
1.00E+02
0 100 200 300 400 500 600 700 800 900
decimation
V
A
F
VAF1
VAF2
VAF3
VAF4
LEGEND % VAF1 % VAF2 % VAF3 % VAF4
Subject:
Train Trial:
Test Trial
02
25
28
07
45
47
20
65
69
04
60
61
Figure 5.3: Decimation Rate Evaluation Plot
5.2. COMPARISON OF EMG AMPLITUDE PROCESSORS
Applying decimation with a factor of 100 and the additional preprocessing steps
(truncation and A/D offset subtraction) described, the EMG amplitudes obtained from
four different processors were used to estimate torque. The method of estimation is
explained earlier hence this section presents the drawn results.
65
5 10 15 20 25
400
200
0
200
400
Extension (Triceps M/C)
Time (sec)
R
a
w
E
M
G
i
n
%
M
V
E
E
5 10 15 20 25
400
200
0
200
400
Flexion (Biceps L/C)
Time (sec)
R
a
w
E
M
G
i
n
%
M
V
E
F
5 10 15 20 25
50
0
50
Real Torque and Predicted Torque Normalized in %MVC
Time (sec)
T
o
r
q
u
e
A
m
p
l
i
t
u
d
e
%
M
V
C
Positive Torque  Flexion
Negative Torque  Extension
Real Torque: Measured
Predicted Torque, Unwhitened EMG
Predicted Torque, Whitened EMG
Figure 5.4: Raw EMG (flexion & extension) and Torques
Figure 5.4 shows the raw EMG signal and the predicted torque from two different
processors in a typical example. The predicted torque in both cases captures most of the
dynamics exhibited in the actual torque.
The values obtained for performance evaluating expressions, % MAE in eq. 4.3
and % VAF in eq. 4.4, were averaged across the 180 combinations for each processor.
Figure 5.5 shows mean and median values of %VAF and MEA, as a function of the
system identification model order, for each of the four EMG amplitude processors. The
plot shows the results obtained for the fast tracking speed using PINV to invert matrices,
whereas the slow tracking speed and the use QRfactorization performance results are
shown in APPENDICES: (III)
66
0 5 10 15 20 25 30
50
60
70
80
90
Median of %VAF, Fast Tracking Speed
System Identification Order, nb
%
V
A
F
0 5 10 15 20 25 30
4
5
6
7
8
9
10
11
12
Median of MAE, Fast Tracking Speed
System Identification Order, nb
M
A
E
(
P
e
r
c
e
n
t
)
0 5 10 15 20 25 30
50
60
70
80
90
Mean of %VAF, Fast Tracking Speed
System Identification Order, nb
%
V
A
F
0 5 10 15 20 25 30
4
5
6
7
8
9
10
11
12
Mean of MAE, Fast Tracking Speed
System Identification Order, nb
M
A
E
(
P
e
r
c
e
n
t
)
SCHUNWHIT
SCHWHIT
MCHUNWHIT
MCHWHIT
SCHUNWHIT
SCHWHIT
MCHUNWHIT
MCHWHIT
SCHUNWHIT
SCHWHIT
MCHUNWHIT
MCHWHIT
SCHUNWHIT
SCHWHIT
MCHUNWHIT
MCHWHIT
Figure 5.5: Median (left) and Mean (right) of % VAF and % MAE for fast tracking
Assuming that better EMGtorque performance is indicated by higher %VAF and lower
% MAE, for all EMGtorque processors the plots show a progressive increase in
performance as model order is increased up to about 10–15
th
order. The improvement
seems to stop passing the 20
th
order and it is expected to eventually decay as the model
order starts fitting into the noise. The distribution characteristics (mean, median, and
standard deviation) for these data are shown in the following tables.
67
Table 5.1: Distribution Info of % VAF Values for Each Processor (Fast Tracking)
Processor 14 % VAF n = 1 n = 2 n = 3 n = 4 n = 5 n = 10 n = 15 n = 20
SCHUNWHIT MEDIAN 51.32 57.56 63.86 69.55 73.64 83.29 84.70 84.94
MEAN 49.17 53.03 56.48 59.65 62.60 68.50 69.07 69.07
STDEV 43.03 44.17 45.71 47.40 49.31 54.54 56.99 57.41
MCHUNWHIT MEDIAN 55.49 62.65 69.05 73.41 77.57 84.80 85.51 85.79
MEAN 54.58 59.89 63.47 66.66 69.33 74.73 75.37 75.52
STDEV 33.73 32.40 32.19 32.13 32.19 32.66 33.03 32.95
SCHWHIT MEDIAN 59.98 65.23 69.99 74.64 78.70 88.31 89.33 89.35
MEAN 56.91 61.98 66.06 69.81 73.15 80.35 81.39 81.47
STDEV 31.42 29.99 29.36 28.93 28.75 28.62 29.16 29.49
MCHWHIT MEDIAN 61.99 69.61 75.72 79.60 82.88 89.45 89.98 90.14
MEAN 61.31 67.92 71.97 75.50 78.33 84.09 84.83 84.97
STDEV 26.27 22.71 21.21 20.02 19.18 17.84 17.86 17.92
Table 5.2: Distribution Info of % MAE Values for Each Processor (Fast Tracking)
Processor 14 % MAE n = 1 n = 2 n = 3 n = 4 n = 5 n = 10 n = 15 n = 20
SCHUNWHIT MEDIAN 10.64 9.90 9.42 8.77 8.16 6.50 6.24 6.20
MEAN 11.21 10.73 10.27 9.82 9.36 8.25 8.04 8.01
STDEV 3.97 3.99 4.09 4.20 4.35 3.97 5.03 5.10
MCHUNWHIT MEDIAN 10.25 9.41 8.74 8.03 7.56 6.34 6.10 6.08
MEAN 10.51 9.89 10.51 8.99 8.59 7.61 7.45 7.42
STDEV 3.66 3.59 3.63 3.68 3.75 4.01 4.08 4.10
SCHWHIT MEDIAN 9.98 9.40 8.68 8.02 7.36 5.65 5.43 5.40
MEAN 10.15 9.55 9.01 8.49 7.98 6.63 6.37 6.33
STDEV 3.27 3.13 3.09 3.27 3.27 3.35 3.47 3.52
MCHWHIT MEDIAN 9.49 8.62 7.88 7.11 6.56 5.31 5.12 5.02
MEAN 9.53 8.74 8.20 7.69 7.25 6.12 5.94 5.90
STDEV 3.01 2.71 2.63 2.59 2.58 2.71 2.76 2.77
As seen, the performance of the processors follows a ranking order with the whitened
multiplechannel processor providing the best performance, followed by multiplechannel
unwhitened, then singlechannel whitened, then singlechannel unwhitened. It is also
noticed that the median results are higher than the mean, suggesting that there are still a
few influential large errors that drag the mean downward.
Finally the PSD of the error is graphed to gain an insight on the origin of the
error. The plots show that 8090% of the error power was within the first 0.5 Hz (shown
in the figure below).
68
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
10
20
30
40
50
60
70
80
90
100
PSD of Total Error Cumulated vs. Frequency for each Processor Type
Frequency in Hz
%
P
S
D
o
f
E
r
r
o
r
C
u
m
u
l
a
t
e
d
SCHUNWHIT
SCHWHIT
MCHUNWHIT
MCHWHIT
Figure 5.6: PSD of Error Accumulation Rate
0.5 0 0.5 1 1.5 2
0
0.01
0.02
0.03
0.04
0.05
0.06
Frequency (Hz)
A
v
e
r
a
g
e
P
S
D
o
f
E
r
r
o
r
%
M
V
C
F
Error Spectral Distribution over Total Error
SCHNWHIT
MCHWHIT
Figure 5.7: The Average PSD of Error as Estimated from Welsh Periodogram
69
Even though Figure 5.6 suggests that almost all error is accumulated at low frequencies,
the windowing effect and the spectral leakage in the Welsh periodogram do not allow
further resolution (Figure 5.7). Therefore, it is hard to determine the origin of the
remaining error.
After a closer study of the trials with the largest error, it was identified that the
error was due to DC shift on the estimated torque as compared to measured one.
Moreover, some trials matched were almost perfectly, except for the existence of a
positive/negative offset (Figure 5.8).
0 2 4 6 8 10 12
x 10
4
2
1
0
1
DCPROBLEMS: Subject 16, Trial 64, order nb=15
T
a
n
d
T

h
a
t
0 500 1000 1500 2000
20
0
20
40
EXTENSION Coefficients Frequency Response
C
f

G
a
i
n
(
d
b
)
0 1000 2000 3000 4000 5000
2
1
0
1
T
a
n
d
T

h
a
t
0 20 40 60 80 100
20
0
20
d = 20 and VAF = 752.2972
C
f

G
a
i
n
(
d
b
)
0 500 1000 1500
2
1
0
1
T
a
n
d
T

h
a
t
0 5 10 15 20 25 30 35
20
0
20
d = 60 and VAF = 876.7484
C
f

G
a
i
n
(
d
b
)
0 200 400 600 800 1000
2
1
0
1
Time (samples)
T
a
n
d
T

h
a
t
0 5 10 15 20
20
0
20
40
d = 100 and VAF = 911.0364
Frequency, Hz
C
f

G
a
i
n
(
d
b
)
Real
Estimate
d = 100 and VAF = 911.0364
d = 60 and VAF = 876.7484
d = 20 and VAF = 752.2972
Time Domain Torque in % MVC
F
d = 1 and VAF = 532.72
Figure 5.8: High DC Offset Error on Estimated Torque
70
The figure shows the changes of the estimated torque as the decimation rate is increased
(1, 20, 60, 100). The performance of the system indicated by % VAF is getting
negatively larger and the DC shift remains as the only apparent error as the AC portion of
the estimated torque starts looking exactly like the measured one. Because most of the
error power is contained at such low frequencies and the observation of DC shift
suggested the same, it can be hypothesized that there is a large amount of error remaining
at approximately DC. More investigation of the error plots is required to ensure the
hypothesis is correct.
In conclusion, the optimum EMGtorque model (15
th
order) produced an average
error of 6% MVC
F
with a %VAF of 90% when the EMG amplitude estimates were
obtained from a multiple channel whitened processor. The typical model performance
was shown previously in Figure 5.4. The results for the slow tracking speed data were
similar and are shown in the appendices. The next largest error in torque estimation
seems to be due to a DC offset.
71
CHAPTER 6. DISCUSSION AND CONCLUSIONS
iii
This last chapter contains two main sections, the discussion and conclusions. The
first part contains the interpretation of the previously anticipated results and then in the
second part conclusions are derived from them. In addition, this chapter includes the
study limitations and the suggestions for future work.
6.1. DISCUSSION OF RESULTS
The objective of this research was to demonstrate that incorporation of the recent
advances in EMG amplitude processors into EMGtorque estimation model produce
lower torque prediction errors. The results presented are achieved considering several
assumptions on physiological characteristics of the EMGtorque system.
6.1.1. Advances to EMGtorque Estimation
Keeping in mind the main objective, various analyses were conducted to
demonstrate that improvement in EMG amplitude processing reduces the estimation error
in the torque prediction models. The results derived were consistent with the
expectations showing that both whitening and multiplechannel combination of the EMG
lead to reduced EMGtorque prediction errors. Even lower errors are obtained when the
two techniques are used in combination. Another important result was identifying and
resolving the problem of oversampling. This problem arises when EMG data (typical
bandwidth from 20–500 Hz) and torque data (typical bandwidth ≈ 510 Hz) are
simultaneously sampled at the highest required rate. Thus, to preserve information in the
iii
Parts of this chapter are taken from the paper submitted to the Journal of Biomechanics (attached to
Appendices) authored by Clancy, E.A; Bida, O.; Rancourt, D.
72
EMG data, the torque data are oversampled. If the so overasmpled torque data are used
to calibrate EMGtorque models, the resultant transfer functions produce unrealistic gains
at frequencies above the typical band of torque data. This issue was observed and solved
by decimating the data to a rate of 40.96 Hz that correspond to approximately ten times
the highest torque frequency, as recommended by Ljung (1999).
6.1.2. Study Limitations and Future Suggestions
Several assumptions were necessary to complete this project. First, the
experimental design consisted of constantposture nonfatiguing contractions about the
elbow. Most practical contractions are more fully dynamic (posture varying). Second,
the mechanical model for the elbow treated the joint as a simple hinge, with only one
agonist and one antagonist muscle group. This assumption enables the possibility of
obtaining one single EMG amplitude estimate using electrodes that are placed anywhere
on the skin overlying that muscle group. Consequently, the combination of multiple
channels improves the EMG amplitude estimation. In addition, there are only two inputs
to the model (EMGamp flexion and EMGamp extension) as opposed to several EMG
amplitudes, representative of each individual muscle recording. Increasing the number of
muscles in the model would not necessarily improve the EMG amplitude estimates and
would increase the complexity of system identification. Another important assumption to
consider is that each muscle group contributed only to a certain torque component, such
as extensor (flexor likewise) torque was a result of extension (flexion likewise) only.
Otherwise, the crossterms of EMG amplitude had to be incorporated in the model
requiring corresponding torque components for mechanical validation.
73
The EMGtorque model was identified using a zerosonly system design,
corresponding to a standard parametric model in the literature (FIR type of ARX model).
The reason for selecting a standard linear model was the level of simplicity in the
solution and the information available on the system identification techniques. Nonlinear
models can be more accurate, for example including hysteresis to capture any systematic
differences in the EMGtorque relation between concentric and eccentric contractions.
Another advantage is that nonlinear models can potentially capture additional subtle
behavior in an EMGtorque relationship such as the electromechanical delay between
action potential activation and muscle fiber contraction. Electromechanical delay (EMD)
is defined as the temporal delay (26131 ms) that exists between the onset of muscular
activity and the generation of force [Strojnik and Komi, 1998]. In the case of abrupt
changes in muscle activation from 0%MVC to 50% MVC, the EMD is a dynamic
parameter, hence it is dependent on the number of motor units activated and the fatiguing
effects. However, during slow varying force tasks (25% MVC to 50% MVC) the
electromechanical delay is constant (60 ms); therefore, its inclusion in the model can be
potentially neglected [Vint, McLean, and Harron, 2001].
Although the EMGtorque model for this study included both “AC and DC
characteristics” of the system, it is highly recommended in the literature to separate them
as part of the data preprocessing routine [Ljung 1999, pp. 458460]. The investigation of
the trials displaying large errors demonstrated that the operating point (DC bias)
contaminated in the inputs (EMGamp) and output (torque) data influenced the dynamics
of the system, resulting in a DC shift in the estimated torque. Even though many system
identification paradigms remove the DC component of each signal prior to identification
74
so that only the system dynamics (AC portion) are identified, a complete output response
(torque) is found by adding the estimated AC portion to a separate estimate of the output
DC value. Further studies are suggested, since a complete system response requires
separate torque data for model validation and should not have influenced the relative
comparison of EMG amplitude processors
The selection of contraction bandwidth that sufficiently excites the system is one
of the main issues that need to be addressed in the future. Although tracking targets were
at low frequency, 1 Hz is about as fast as subjects were able to track. Faster speeds
would require ballistic force trajectories. Therefore, the contraction bandwidth was
limited to a 1 Hz tracking target. If higher frequencies are required in the future, the
tracking target will need to produce a deterministic trajectory so that ballistic movements
(which are inherently faster) can be utilized.
Finally, precautions were taken when recording the data in order to avoid
hardware problems such as electrode failures, the distances between the electrodes, and
their placement on the skin. The skin was properly prepared to reduce the impedance
prior to placing the electrodes; active electrodes rejected the cable motion artefact; digital
high pass filters were utilized to eliminate noise due to physiological motion artefact; and
the lowpass filters along with decimation attenuated other interferences such as UV
lights and power harmonics.
6.2. SUMMARY AND CONCLUSIONS
This research focused on demonstrating that advances in EMG amplitude processors
result in EMGtorque model performance improvements. Advances in EMG amplitude
estimation were applied to the EMGtorque problem for constantposture, nonfatiguing,
75
forcevarying contractions about the elbow. Results from 15 subjects showed that EMG
whitening and multiplechannel combination both reduce EMGtorque errors and their
combination provides an additive benefit. The dynamic relationship between EMG
amplitude and joint torque was formulated as a standard linear least squares problem.
Using 15
th
order and higher linear FIR models, EMGtorque errors with a fourchannel,
whitened processor produced an average error of 6% MVC
F
(%VAF of 90%) at the fast
tracking speed. By comparison, the singlechannel, unwhitened (conventional) processor
produced an average error of 8% MVC
F
(%VAF of 68%).
Accomplishing the objective, the issue of nonconvergent trials was isolated and
resolved by decimating and low pass filtering the data prior to system identification. The
EMG amplitude sampling rate was reduced to 40.96 Hz and both EMG amplitude and
torque were lowpass filtered at the Nyquist rate, using an 8
th
order, zerophase,
Butterworth filter. The power spectral density analysis showed that the chosen cutoff
frequency of the filter preserved 99.9% of the system power.
Concluding, the primary interest was the influence of different EMG amplitude
processors on EMGtorque prediction performance. As such, the study was limited in
several manners in order to be able isolate the effect of EMG amplitude without the
complexity of less restrictive EMGtorque models. The expectation is that the benefits
shown here of improved EMG amplitude processors would transfer to many other EMG
torque modeling problems. Certainly, it now seems justified to progressively release
these restrictions and validate these benefits in more general applications in future work.
76
REFERENCES
[Armstrong et al., 1982] T. J. Armstrong, J. A. Foulke, B. S. Joseph, S. A. Goldstein,
“Investigation of cumulative trauma disorders in a poultry processing plant.” Amer.
Industrial Hyg. Assoc. J., vol. 43, no. 2, pp. 103116, 1982.
[Agarwal and Gottlieb, 1977] G.C. Agarwal, G.L. Gottlieb,“Compliance of the human
ankle joint.” J Biomech Eng, vol. 99, pp.166–170, 1977.
[Basmajian and DeLuca, 1985] J.V. Basmajian and C.J. De Luca, Muscles Alive: Their
Functions Revealed by Electromyography, Williams & Wilkins Company, 1985.
[Bigland and Lippold, 1954] B. Bigland, O.C.J Lippold, “The relation between force,
velocity and integrated electrical activity in human muscles.” J. Physiol. (Lond), vol.
123, pp. 214224, 1954.
[Bendat and Piersol, 1986] J.S. Bendat, A.G. Piersol, Random Data: Analysis and
Measurement Procedures, John Wiley, N.Y, 1986.
[Bobet and Norman, 1990] “LeastSquares Identification of the dynamic relation
electromyogram and joint moment.” J. Biomech., vol. 23, no. 12, pp. 12756, 1990.
[Bouchard, 2001] S. Bouchard, “Relation dynamique entre les signaux
electromyogrphiques et le couple au coude lors de contarctions a angles constants.”
M.S. thesis, Dept. Mech. Eng., Laval Univ., Quebec, Canada 2001.
[Buchanan et al., 1993] T. S. Buchanan, M. J. Moniz, J. P. A. Dewald, W. Z. Rymer,
“Estimation of muscle forces about the wrist joint during isometric tasks using an
EMG coefficient method.” J. Biomech., vol. 26, no. 45, pp. 547560, May 1993.
[Clancy, 1991] E.A. Clancy, “Stochastic modeling of the relationship between the
surface electromyogram and muscle torque.” PhD thesis, Dept. Elect. Eng. Comp.
Sci., MIT, Cambridge, MA 1991.
[Clancy, 1999] E.A. Clancy, “Electromyogram amplitude estimation with adaptive
smoothing window length.” IEEE Trans. Biomed. Eng., vol. 46, no. 6, 1999.
[Clancy, 2004] E.A. Clancy. EMG amplitude estimation toolbox: User’s guide. Alpha
version 0.04. Available online at: http://ece.wpi.edu/~ted/emg_tool.htm. Accessed
January 30, 2005.
[Clancy et al., 2001] E.A. Clancy, S. Bouchard, D. Rancourt, “Estimation and
application of EMG amplitude during dynamic dontractions.” IEEE Eng. Med. Bio.
Mag., vol. 20, no.6, Nov.  Dec. 2001.
77
[Clancy et al., 2004] E.A. Clancy, D. Farina, G. Filligoi, “Single Channel Techniques for
Information Extraction from the Surface EMG,” in Electromyography: Physiology,
Engineering, and Noninvasive Applications, Roberto Merletti and Philip A. Parker,
eds., John Wiley & Sons, Inc., Hoboken, New Jersey, pp. 133−168, 2004.
[Clancy and Farry, 2000] E.A. Clancy and K.A. Farry, “Adaptive whitening of the
electromyogram to improve amplitude estimation.” IEEE Trans. Biomed. Eng., vol.
47, no. 6, June 2000.
[Clancy and Hogan, 1995] E.A Clancy, N. Hogan, “Multiple site electromyograph
amplitude estimation.” IEEE Trans. Biomed. Eng., vol. 42, pp. 203–211, 1995.
[Clancy and Hogan, 1997] E.A. Clancy and N.Hogan, “Relating agonistantagonist
electromyograms to joint torque during isometric, quasiisotonic, nonfatiguing
contractions.” IEEE Trans. Biomed. Eng., vol. 44, no. 8, June 2000.
[Clancy, Morin, and Merletti 2002] E.A. Clancy, E.L. Morin, R. Merletti, “Sampling,
noisereduction and amplitude estimation issues in surface electromyography.” J.
Electromyog. Kines., vol. 12, pp 1–16, December 2002.
[DeLuca, 2002] C.J. De Luca, “Surface electromyography: detection and recording”
DelSys Inc. Copyrighted 2002
[Dowling, 1997] J.J Dowling, “The use of electromyography for the noninvasive
prediction of muscle forces.” Sports Med., vol. 24, no. 2, pp. 8296, Aug. 1997.
[Farina, Merletti, and Enoka, 2004] D. Farina, R. Merletti, R.M. Enoka, “The
extraction of neural strategies from the surface EMG.” J. Appl. Physiol. vol. 96 pp.
14861495, 2004.
[Grant et al., 1994] K. A. Grant, D. J. Habes, V. PutzAnderson, “Psychophysical and
EMG correlates of force exertion in manual work.” Int. J. Industrial Ergonomics, vol.
13, pp. 3139, 1994.
[Gottlieb and Agarwal, 1971] G.L. Gottlieb and G.C. Agarwal, “Dynamic relationship
between isometric muscle tension and the electromyogram in man.” J. Appl. Physiol.,
vol. 30, pp. 345 – 351, 1971.
[Gottlieb et al., 1997] G.L. Gottlieb, Q. Song, G.L. Almeida, D. Hong, D. Corcos,
“Directional control of planar human arm movement.” J. Neurophys., vol. 78, pp.
29852998, 1997
[Hannerz, 1974] J. Hannerz, “Discharge properties of motor units in relation to
recruitment order in voluntary contraction.” Acta Physiol. Scand., vol. 91, pp. 374–
384, 1974.
78
[Hasan and Enoka, 1985] Z. Hasan, R. M. Enoka, “Isometric torque angle relationship
and movement related activity of human elbow flexors: Implications for the
equilibriumpoint hypothesis.” Exp. Brain Res., vol. 59, pp. 441450, 1985.
[Hof and Van Den Berg, 1981] A.L. Hof, Jw. Van den Berg, “EMG to force processing
I: An electrical analogue of the Hill muscle model.” J. Biomechanics vol. 14, no. 11,
pp. 747–758, 1981.
[Hogan and Mann, 1980a] N. Hogan and R.W. Mann, “Myoelectric signal processing:
optimal estimation applied to electromyographyPart I: Derivation of the optimal
myoprocessor.” IEEE Trans. Biomed Eng., vol. BME27, pp. 382395. July 1980.
[Hogan and Mann, 1980b] N. Hogan and R.W. Mann, “Myoelectric signal processing:
optimal estimation applied to electromyographyPart I: Derivation of the optimal
myoprocessor.” IEEE Trans. Biomed Eng., vol. BME27, pp. 396410. July 1980.
[Hughes and Chaffin, 1997] R. E. Hughes, D. B. Chaffin, “Using principalcomponents
regression to stabilize EMGmuscle force parameter estimates of torso muscles.”
IEEE Trans. on Biomed. Eng., vol. 44, no. 7, pp. 639642, 1997.
[Inman et al., 1952] V.T. Inman, H.J. Ralston, J.B.C.M Saunders, B. Feinstein, E.W.
Wright, “Relation of human electromyogram to muscular tension.” Electroencephol.
and Clin. Neurophys. vol. 4, 187–194, 1952.
[Kirsch, Kearney, and Crago, 1994] R.F. Kirsch, R.E. Kearney, P.E. Crago, “Dynamic
stiffness measurements and control movement.” IEEE 16th Ann. Int. Conf. Proc.,
vol.1, pp. 428429, November 1994.
[Kolman and Hill, 2001] B. Kolman, D. R. Hill, “Introductory Linear Algebra with
Applications”, 7
th
ed. Prentice Hall, New Jersey, pp.327339, 2001.
[Laursen et al., 1998] B. Laursen, B. R. Jensen, G. Nemeth, G. Sjogaard, “A model
predicting individual shoulder muscle forces based on relationship between
electromyographic and 3D external forces in static position.” J. Biomech., vol. 31, pp.
731739, 1998.
[Lawrence and DeLuca, 1983] J.H.Lawrence, C.J De Luca, “Myoelectric signal versus
force relationship in different human muscles.” J. Appl. Physiol., vol. 54, pp. 1653
1659, 1983.
[Lamb and Hobart, 1992] R. Lamb, D. Hobart, “Selected topics in surface
electromyography for use in the occupational settings: exepert perspectives”, (CH 2),
U.S. Department of Health and Human Service (CDC) – NIOSH, March 2002.
79
[Lindstrom and Magnusson, 1977] J. Lindstrom, R. T. Magnusson, “Interpretation of
myoelectric power spectra: A model and its application.” Proceedings of the IEEE,
vol. 65, pp 653662, 1977.
[Ljung, 1999] L. Ljung, System Identification: Theory for the User, Upper Saddle River,
NJ: PrenticeHall, 1999.
[MilnerBrown and Stein, 1975] H. S. MilnerBrown, R. B. Stein, “The relation
between the surface electromyogram and muscular force.” J. Physiol., vol. 246, pp.
549569, 1975.
[McGill, 1992] S.M. McGill, “A myoelectrically based dynamic threedimensional
model to predict loads on lumbar spine tissues during lateral bending.” J. Biomech.,
vol. 25, no. 4, pp. 395414, Apr.1992.
[Merletti, 1999] R. Merletti, “Standards for Reporting EMG Data.” J. Electromyography
Kinesiology, 1999.
[Moritani and DeVries, 1978] T. Moritani, H. DeVries, “Reexamination of the
relationship between the surface integrated electromyogram (IEMG) and force of
isometric contraction.” Amer. J. Phys. Med., vol. 57, no. 6, pp. 263277, 1978.
[Nussbaum, Chaffin, and Martin, 1995] M. A. Nussbaum, D. B. Chaffin, B. J. Martin,
“A backpropagation neural network model of lumbar muscle recruitment during
moderate static exertions.” J. Biomech., vol. 28, no. 9, pp. 10151024, 1995.
[Papoulis, 2002] A. Papoulis, Probability, Random Variables, and Stochastic Processes,
4th edition, McGraw Hill, 2002.
[Perry and Bekey, 1981] J. Perry, G. A Bekey, “EMGforce relationships in skeletal
muscle.” CRC Critical Rev. in Biomed. Eng., pp. 122, December 1981.
[Ray and Guha, 1983] G.C. Ray, S.K. Guha, “Relationship between the surface e.m.g.
and muscular force.” Med. Biol. Eng. Comput., vol. 21, no. 5, pp. 579586, Sept. 1983
[Solmonow et al., 1986] M. Solomonow, A. Guzzi, R. Baratta, H. Shoji, R. D’Ambrosia,
“EMGforce model of the elbow’s antagonistic muscle pair: The effect of joint
position, gravity and recruitment.” Am. J. Physic. Med., vol. 65, pp. 223–244, 1986.
[Solmonow et al., 1990] M. Solomonow, R. Baratta, H. Shoji, R. D’Ambrosia, “The
EMGforce relationships of skeletal muscle; dependence on contraction rate, and
motor units control strategy.” Electromyog.r Clin. Neurophysiol., vol. 30, no. 3, pp.
141152, 1990.
80
[Sommerich et al., 1998] C. M. Sommerich, W. S. Marras, M. Parnianpour, “A method
for developing biomechanical profiles of handintensive tasks.” Clinic. Biomech., vol.
13, pp. 261271, 1998.
[Stokes, Moffroid, Rush, and Haugh, 1989] I. A. F. Stokes, M. Moffroid, S. Rush, L.
D. Haugh, “EMG to torque relationship in rectus abdominis muscle.” Spine, vol. 14,
no. 8, pp. 85761, 1989.
[Strojnik and Komi, 1998] V. Strojnik, P. V. Komi, “Neuromuscular fatigue after
maximal stretchshortening cycle exercises.” J. Appl. Physiol., Vol. 84, Issue 1, pp.
344350, 1998.
[Thelen et al., 1994] D.G. Thelen, A.B. Schultz, S.D. Fassois, J.A. AshtonMiller,
“Identification of dynamic myoelectric signaltoforce models during isometric
lumbar muscle contractions.” Journal of Biomechanics, vol. 27, pp. 907–919, 1994.
[Thelen, 2003] D.G. Thelen, “Adjustment of muscle mechanics model parameters to
simulate dynamic contractions in older adults.” J. Biomech. Eng., vol. 125, pp. 7077,
2003.
[ValeroCuevas et al., 2003] F.J ValeroCuevas, M.E. Johanson, J.D. Towles, “Towards
a realistic biomechanical model of the thumb: the choice of kinematic description
may be more critical than the solution method or the variability/uncertainty of
musculoskeletal parameters.” J Biomech, vol. 36, pp. 10191030, 2003.
[Vint, McLean, and Harron, 2001] P.F. Vint, S.P. McLean, G.M. Harron,
“Electromechanical delay in isometric actions initiated from nonresting levels.”
Med. & Sci. in Sports & Exercise, vol. 33, issue 6 pp. 978983. 2001.
[Vredenbregt and Rau, 1973] J. Vredenbregt, G. Rau, “Surface EMG in relation to
force, muscle length and endurance.” New Develop. EMG Clin. Neurophys., vol. 1,
pp. 607622, 1973.
[Westwick, 1995] D.T. Westwick, Methods for the Identification of MultipleInput
Nonlinear Systems. Dept. EE Biomed. Eng., McGill University, Montreal, Quebec,
Canada, 1995. http://www.enel.ucalgary.ca/People/Westwick/Publications/
[Woods and BiglandRitchie, 1983] J.J. Woods, B. BiglandRitchie, “Linear and non
linear surface EMG/force relationships in human muscles: An anatomical/functional
argument for the existence of both.” Am. J. Phys. Med., vol. 62, no. 6, pp. 287299,
1983.
[Wyss and Pollak, 1984] J.P. Wyss, V.A. Pollak “Surface electromyogram/muscle force:
a muscle model based on RMG peaks.” Eng. Med., vol. 13, no. 1, pp. 2733, 1984.
81
[Zuniga and Simons, 1969] E.N. Zuniga, D.G. Simons, “Nonlinear relationship between
averaged electromyogram potential and muscle tension in normal subjects.” Arch.
Phys. Med. and Rehab., 613620, November 1969.
82
APPENDICES: ADDITIONAL INFORMATION, PLOTS, AND FIGURES
I. LBXXXX EXPERIMENT DATA FILE DESCRIPTION
The file name code is LBSSTT where SS stands identifies the subject (0205, 0710, 12,
13, 16, 17, 18, 20, 21) and TT identifies the experiment trial (2029, 4049, 6069).
There are 15 subjects and 3 sets of 10 trials per subject. The different trials within a set
were obtained using different feedback mechanisms and bandwidths for the tracking
signal. One out of four trials was used to train the EMGtorque model the other four to
test it.
Table 0.1: Trial ID Name Codes
TRIAL ID (TT) Feedback Mechanism Target Bandwidth (Hz)
20 (train) EMG, Single Channel Unwhitened 0.25
21 (test) EMG, Multiple Channel Unwhitened 0.25
22 (test) EMG, Single Channel Whitened 0.25
23 (test) EMG, Multiple Channel Whitened 0.25
24 (test) Dynamometer 0.25
25 (train) EMG, Single Channel Unwhitened 1
26 (test) EMG, Multiple Channel Unwhitened 1
27 (test) EMG, Single Channel Whitened 1
28 (test) EMG, Multiple Channel Whitened 1
29 (test) Dynamometer 1
40 (train) EMG, Single Channel Unwhitened 0.25
41 (test) EMG, Multiple Channel Unwhitened 0.25
42 (test) EMG, Single Channel Whitened 0.25
43 (test) EMG, Multiple Channel Whitened 0.25
44 (test) Dynamometer 0.25
45 (train) EMG, Single Channel Unwhitened 1
46 (test) EMG, Multiple Channel Unwhitened 1
47 (test) EMG, Single Channel Whitened 1
48 (test) EMG, Multiple Channel Whitened 1
49 (test) Dynamometer 1
60 (train) EMG, Single Channel Unwhitened 0.25
61 (test) EMG, Multiple Channel Unwhitened 0.25
62 (test) EMG, Single Channel Whitened 0.25
83
63 (test) EMG, Multiple Channel Whitened 0.25
64 (test) Dynamometer 0.25
65 (train) EMG, Single Channel Unwhitened 1
66 (test) EMG, Multiple Channel Unwhitened 1
67 (test) EMG, Single Channel Whitened 1
68 (test) EMG, Multiple Channel Whitened 1
69 (test) Dynamometer 1
Each of the files contained 16 channels from the DAQ describing the electrode positions
as follows. The table lists only the channel used for this study.
Table 0.2: A/D Channel Name Codes
A/D Channel Contents
1 EMG: Biceps, most lateral location
2 EMG: Biceps, lateral center
3 EMG: Biceps, medial center
4 EMG: Biceps, most medial location
8 EMG: Triceps, most medial location
9 EMG: Triceps, medial center
10 EMG: Triceps, lateral center
11 EMG: Triceps, most lateral
16 Dynamometer
84
II. OPTIONAL PROPERTIES FOR AMPLITUDE ESTIMATION ALGORITHM
EMG amplitude estimation as taken from the EMG toolbox User Manual.
Syntax : EMGamp = e_amp(EMGin, EMGinfo)
EMGinfo – obtained from calibration
EMGin – input EMG channel (CH 2 for flexion and CH 9 for extension) or Multiple
Channels (CH 1:4 for flexion and CH 8:11 for extension)
Calibration for EMG amplitude estimation as in EMG toolbox User Manual.
Syntax: EMGinfo = e_cal(SandNmat, sCal, NoiseMat, SampFreq [, 'PropertyName',
PropertyValue, ...]
SandNmat  Noisy signal Matrix (Input Channel). This signal is specifically recorded
for calibration process. Trial 10 is used for flexion and trial 12 for extension. The
channels are the same as the input channels shown for EMGin.
NoiseMat – Noise recorded per trial, rest Trial 15 (take corresponsing channel) For
example: NoiseMat for flexion is trial 15 (CH 2 for singlechannel and CH 1:4 for
multiplechannel) and for extension is still trial 15 but different channels.
Properties as set of this project:
sCal: 0.5
SampFreq: 4096 Hz
High Pass Filter Settings
Causality 'Flag': Default value is 'Noncausal'.
HpassFlag 'Flag': Default value is 'Filter'.
HpassOrder Order: Default value is set by e_h_pass() and Order = 5.
HpassWn: Default value is set to correspond to 15 Hertz.
Whitening Filter Settings
WhiteFlag 'Flag': Depends, SCHWHIT and MCHWHIT is ‘ON’ else ‘OFF’.
85
WhiteEdges EdgeString: Default is set by e_whiten().
WhiteMaxGain MMaxGain: Default is set by e_cal_wh().
WhiteNfft NNfft: Default is set by e_cal_wh().
WhiteOrder Order: Default is set by e_cal_wh().
White_sSpec sSpec: Default is set by e_cal_wh().
WhiteSum SumString: Default is set by e_whiten().
WhiteSmFilt FiltOption: Not Used.
WhiteSmFilt = 'MAV':
WhiteSmFixWin Window = 1024 = 250ms.
Multiple Channel Combination Setting
UncorrFlag 'Flag': Default is 'GainOnly', if multiple channels.
Demodulation Settings
DemodFlag 'Flag': Default value is 'On'.
DemodM M: Default d = 1 (MAV)
Smoothing Filter Setting
SmoothFilt FiltOption: ‘butter’
SmoothEdges EdgeString: Default value is set by e_smooth().
SmoothWn: 0 < Wn < 1= > Wn = 20/2048 which is 20Hz
SmoothOrder: Order = 8
For example EMGinfo for Extension in the case of Single Channel Processor
EMGinfoExt = e_cal(SandNmatE, 0.5, NmatE, 4096, 'WhiteFlag', Wflag, 'WhiteSmFilt',
'MAV', 'WhiteSmFixWin', 1024);
The code looks exactly the same for all the other cases. For more information on the
settings and the EMG toolbox functionalities refer to the User Manual created and
maintained by Clancy (2004) found in the website:
http://ece.wpi.edu/~ted/emg0_04/front.html
86
III. EXTRA FIGURES AND PLOTS
0 0.5 1 1.5 2
0
0.02
0.04
0.06
Frequency
0 0.5 1 1.5 2
0
0.02
0.04
0.06
0.08
0 0.5 1 1.5 2
0
0.01
0.02
0.03
0.04
0.05
0 0.5 1 1.5 2
0
0.02
0.04
0.06
Frequency
Frequency
Frequency
Average of Error PSD (%MVC
F
) SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
Figure 0.1: Estimation Error PSD (Welsh Periodogram) for all 4 Processors
The power density spectra (PSD) of estimation error are approximated using Welsh
periodogram for all four processors. Figure 0.1 is used to support the argument that the
estimation error for all processors behaves similarly in frequency domain. Figure 5.7
shows only two of the processors (schwhit, mchwhit).
87
0 20 40 60
40
50
60
70
80
90
100
Median of %VAF, fast tracking speed
%
V
A
F
0 20 40 60
0.04
0.06
0.08
0.1
0.12
Median of MAE, fast tracking speed
SI order, nb
M
A
E
0 20 40 60
40
50
60
70
80
90
100
Mean of %VAF, fast tracking speed
SI order, nb
%
V
A
F
0 20 40 60
0.04
0.06
0.08
0.1
0.12
Mean of MAE, fast tracking speed
SI order, nb
M
A
E
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
QR Method
SI order, nb
Figure 0.2: System Performance (% VAF & MAE) using QR Factorization (Fast
Tracking)
The system performance versus system ID order using QR factorization to compute the
inverse of the matrices is given to support the argument that the pseudoinverse and the
QR factorization give equal results. Figure 0.2 is contrasted with Figure 5.5 and the
results seem to match perfectly.
88
0 20 40 60
50
60
70
80
90
100
Median of %VAF, slow tracking speed
SI order, nb
%
V
A
F
0 20 40 60
0.04
0.06
0.08
0.1
0.12
Median of MAE, slow tracking speed
SI order, nb
M
A
E
0 20 40 60
50
60
70
80
90
100
Mean of %VAF, slow tracking speed
SI order, nb
%
V
A
F
0 20 40 60
0.05
0.06
0.07
0.08
0.09
0.1
0.11
Mean of MAE, slow tracking speed
SI order, nb
M
A
E
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
Figure 0.3: System Performance (% VAF & MAE) using PseudoInverse (Slow
Tracking)
Figure 0.3 is contrasted with Figure 5.5 and the processor performance ranking order
seems to follow the same pattern. The model gives better results for slow target tracking
tasks which is consistent with the statement that the subjects have trouble tracking
ballistic random trajectories. Regardless, the positive influence of EMG amplitude
processor on the EMGtorque model performance is proven for both cases (fast and slow
bandwidth signal target).
89
0 20 40 60
50
60
70
80
90
100
Median of %VAF, fast tracking speed
SI order, nb
%
V
A
F
0 20 40 60
0.04
0.06
0.08
0.1
0.12
Median of MAE, fast tracking speed
SI order, nb
M
A
E
0 20 40 60
50
60
70
80
90
100
Mean of %VAF, fast tracking speed
SI order, nb
%
V
A
F
0 20 40 60
0.04
0.06
0.08
0.1
0.12
Mean of MAE, fast tracking speed
SI order, nb
M
A
E
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
SCHUNWHIT
MCHUNWHIT
SCHWHIT
MCHWHIT
AC Model ONLY
Figure 0.4: System Performance (% VAF & MAE) using AC part of EMG
Amplitudes (Fast Tracking + PINV)
Figure 0.4 is contrasted with Figure 0.3 and Figure 5.5. The processor performance
ranking order seems to follow yet the same pattern. The model gives even better results
when only the AC portion of EMG recordings and measured torque are used in the
system. This is another observation toward solving the remaining problems and leading
again to the assumption that the inclusion of DC in the models is erroneous.
90
0 5 10 15 20
0
1
2
3
Extension Coefficient Freq. Response
O
r
d
e
r
=
2
0 5 10 15 20
0
1
2
3
Flexion Coefficient Freq. Response
O
r
d
e
r
=
2
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
1
0
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
1
0
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
1
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
1
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
2
0
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
2
0
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
2
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
2
5
0 5 10 15 20
0
1
2
3
Frequency (hz)
O
r
d
e
r
=
3
0
0 5 10 15 20
0
1
2
3
Frequency (hz)
O
r
d
e
r
=
3
0
Subject 03 Trial 40 Test 42
Figure 0.5: Coefficients Frequency Response for a typical EMGTorque model
(Slow Tracking)
The most important factor observed in this figure is the shape change as the number of
zeros (model order) is increased. Notice that as the order of the system passes the value
ten little or no improvement is seen within 5Hz (bandwidth of the system). Higher orders
seem to add shape to frequencies outside the system bandwidth. This observation is
consistent with the system saturation after 15
th
order seen in many of the plots shown
above.
91
0 5 10 15 20
0
1
2
3
Extension Coefficient Freq. Response
O
r
d
e
r
=
2
0 5 10 15 20
0
1
2
3
Flexion Coefficient Freq. Response
O
r
d
e
r
=
2
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
1
0
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
1
0
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
1
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
1
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
2
0
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
2
0
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
2
5
0 5 10 15 20
0
1
2
3
O
r
d
e
r
=
2
5
0 5 10 15 20
0
1
2
3
Frequency (hz)
O
r
d
e
r
=
3
0
0 5 10 15 20
0
1
2
3
Frequency (hz)
O
r
d
e
r
=
3
0
Subject 03 Trial 45 Test 47
Figure 0.6: Coefficients Frequency Response for a typical EMGTorque model
(Fast Tracking)
This plot is equivalent with Figure 0.5 for fast target tracking tasks. Similarly, as order
increases more shape is added to the response. In this plot it is also observed that during
low order the fast later zeros do not leave time for the previous ones to create the dip
effect. This is observed on the highpass look alike response for second and fifth orders.
92
IV. PAPER SUBMITTED TO THE JOURNAL OF BIOMECHANICS
Influence of advanced electromyogram (EMG) amplitude processors on EMGto
torque estimation during constantposture, forcevarying contractions
by
E.A. Clancy
a,
*, O. Bida
b
, D. Rancourt
c
a
Department of Electrical and Computer Engineering, Department of Biomedical
Engineering, Worcester Polytechnic Institute, 100 Institute Rd., Worcester, MA,
01609, USA
b
Department of Electrical and Computer Engineering, Worcester Polytechnic Institute,
Worcester, MA, USA
c
Department of Mechanical Engineering, University of Sherbrooke, Sherbrooke, PQ,
Canada
Edward (Ted) A. Clancy (Corresponding Author)
Department of Electrical and Computer Engineering
Worcester Polytechnic Institute
100 Institute Road, Worcester, MA 01609 USA
Tel. (508) 8315778, Fax. (508) 8315491, Email: ted@wpi.edu
93
Abstract
Numerous studies have investigated the relationship between surface EMG and
torque exerted about a joint. These studies have used conventional EMG amplitude
(EMGamp) processing, such as rectification followed by low pass filtering, to pre
process the EMG before relating it to torque. Recently, advanced EMGamp processors
that incorporate signal whitening and multiplechannel combination have been shown to
significantly improve EMGamp processing. In this study, we compared the performance
of EMGamptorque estimators with and without these advanced EMGamp processors.
Fifteen subjects produced constantposture, nonfatiguing, forcevarying contractions
about the elbow while torque and biceps/triceps EMG were recorded. EMGamp was
related to torque using a linear FIR model. Both whitening and multiplechannel
combination reduced EMGtorque errors and their combination provided an additive
benefit. Using a 15
th
order linear FIR model, EMGtorque errors with a fourchannel,
whitened processor averaged 6% of maximum voluntary contraction (or 90% of variance
accounted for). By comparison, the equivalent singlechannel, unwhitened
(conventional) processor produced an average error of 8% of maximum voluntary
contraction (variance accounted for of 68%). In addition, the study describes the
occurrence of spurious peaks in estimated torque when the torque model is created from
data with a sampling rate well above the bandwidth of the torque. This problem occurs
when the torque data are sampled at the same rate as the EMG data. The problem is
corrected by decimating the EMGamp prior to relating it to joint torque, in our case to an
effective sampling rate of 40.96 Hz.
Keywords: EMG, EMG amplitude, Torque, EMGtorque model, Optimal sampling rate
© 2005 OLJETA BIDA ALL RIGHTS RESERVED
To my dearest family, for their selfless sacrifices, and for the continuous belief in my capabilities regardless of my weak moments and the obstacles that we have gone through.
Utilizing system identification techniques. unwhitened (conventional) processor produced an average error of 8% of maximum voluntary contraction (variance accounted for of 68%). The experimental data are taken from fifteen subjects that produced constantposture. forcevarying contractions about the elbow while torque and biceps/triceps EMG were recorded. EMG amplitude was related to torque through a zerosonly (finite impulse response. The incorporation of whitening and multiplechannel combination separately reduced EMGtorque errors and their combination provided a cumulative improvement. The standard deviation of the recorded EMG signal is defined as the EMG amplitude. whitened processor. The EMG amplitude estimation technique varies with the study from conventional type of processing (i. A 15thorder linear FIR model provided an average estimation error of 6% of maximum voluntary contraction (or 90% of variance accounted for) when EMG amplitudes were obtained using a fourchannel.ABSTRACT A number of studies have investigated the relationship between surface electromyogram (EMG) and torque exerted about a joint. nonfatiguing.e. rectification followed by low pass filtering) to further addition of different noise rejection and signaltonoise ratio improvement stages. Advanced EMG amplitude processors developed recently that incorporate signal whitening and multiplechannel combination have been shown to significantly improve amplitude estimation. iii . The main contribution of this research is a comparison of the performance of EMGtorque estimators with and without these advanced EMG amplitude processors. The equivalent singlechannel. FIR) model.
Keywords: EMG. in this case to an effective sampling rate of 40. EMG Amplitude. This problem is anticipated when the torque data are sampled at the same rate as the EMG data. Torque. and Linear Torque Model.96 Hz. The problem is resolved by decimating the EMG amplitude prior to relating it to joint torque. EMGtorque Model. System Identification. Optimal Sampling Rate.This study also describes the occurrence of spurious peaks in estimated torque when the torque model is created from data with a sampling rate well above the bandwidth of the torque. iv .
vullnetin. I owe everything I am and I have done to the selfless sacrifice and to the love which my family raised me with. Nuk mund ta imagjinoni dot mirenjohjen per dashurine. Denis Rancourt from University of Shebrooke (Canada) for his help throughout the course of this project and for extending his help even beyond this project completion.” Thanks to my husbandtobe Eno for being my best friend and for already standing by me for better and for worst. With his continuous love and care. the completion of this thesis would not have been possible. during my freshman year while taking EE2011. Also thanks to my labmates Karthik and Hongfang for the joyful. We have been together through many hardships and now it is time to enjoy the fruits of our work. who initially introduced me to the fascinating field of Electrical Engineering. Thanks to my dearest sister Ana for the mutual love and loyal friendship. and sharing atmosphere created in our lab. I am thankful to the Almighty God. Beyond all and everything. he has encouraged me to work harder to reach our life dreams. I feel honored to have had a chance to work with him in this research for my Masters Degree and because of him to have reached so far. I thank Dr. Without his guidance and support. kind. for answering my prayers. dhe vlerat shpirterore me te cilat me kini rritur duke shpresuar qe nje dite te arrijme te gjitha endrrat e thurura se bashku. Heartfelt thanks to my MQP advisor Professor David Cyganski and to Professor Rick Brown for their presence in the research committee and their helpful advice in the project. and for always guiding me toward happiness and peace. Professor Edward Clancy. My respect toward him grew as I started to realize that the knowledge gained and the work ethics developed in my introductory class brought me to this point.ACKNOWLEDGMENTS I am deeply thankful to my advisor. “Falenderoj nga thellesia e zemres prinderit e mi te shtrenjte per sakrificen dhe durimin gjate veshtiresive. v .
.......................2.............................................. SYSTEM IDENTIFICATION PROCEDURE ............................... BIOMECHANICAL SYSTEM MODELING TECHNIQUES ........... 34 3...................III ACKNOWLEDGMENTS....................................2............ INTRODUCTION........................... 51 4.............................3....... 1........................ Apparatus and Experimental Procedure ......................................................................2.........TABLE OF CONTENTS ABSTRACT ...... 5........................................................................................... 44 4............... DATA COLLECTION AND ANALYSIS METHODS ... Multiple Channel Combination and Gain Normalization........ PROJECT BACKGROUND......... 6 2..... 15 2.........................2.....................1..5......................................................................................................................................V TABLE OF CONTENTS ........1................................. Noise Reduction Precautions.............1............................... 23 2........... 6 2...............1................................................ PROJECT RESULTS.2................................................................................................................................................................................ Advanced EMG Amplitude Estimation ....................2....................................................2.. Parametric System Identification.......... DECIMATION ........................................................................................................................................................................................................................................................................................... EMGTorque Relationship Modeling ...3.................. PROJECT MOTIVATION ............................................................................................................................................ Electrical Activity Generation ........................................................................................................ MODEL SOLUTION ........4.1....... 53 4..... SURFACE EMG AMPLITUDE ESTIMATION TECHNIQUES ..3........2.................................... 22 2..... 38 3.....................................................................1.3............. 6 2.........1.................................................................................... 16 2....................................... 71 vi .............1.... 1 THESIS CONTRIBUTION .............................................1.......... Adaptive Whitening..........................3............................................. Physical Interpretation of the EMGTorque Model.......... 3 THESIS CONTENT ...............1..............................1.....................2.......................... Torque Estimation Procedure.............................3...........3............................ 8 2................................................................3......... EMGTORQUE MODEL DESIGN ....................................3....................................... Origin and Character of EMG ............................................. 43 4............................................................... IX CHAPTER 1................................................................................................................................ 64 CHAPTER 6......3............................................................... 18 2.............. Factors that Effect EMG Signal.. VIII LIST OF TABLES................................................ Noise Rejection Filters .................2......2....................................................................... SURFACE EMG TO TORQUE MODEL DESIGN ........3............ 1....... VI LIST OF FIGURES........ 40 CHAPTER 4............2.................................................................................................................................................................................1.................... Overview of Modeling Techniques ..... Mathematical Modeling for EMGtorque......2..................................................................................3...................... EMG DATA COLLECTION......................... 34 3.............. 24 2.1........ Data PreProcessing ............................ Standard EMG Amplitude Estimation ..............3................................................................. 43 4......................1................................... 15 2...... 60 COMPARISON OF EMG AMPLITUDE PROCESSORS ......................................................... 22 2................................................ 57 CHAPTER 5............................1............................ 55 4................1..... 19 2........... 60 5........... 35 3................................. EMG AMPLITUDE ESTIMATION METHOD .................................2................................. 53 4....... Model Performance Measures.......1................. 27 CHAPTER 3.. 1 1.............................................................................. 4 CHAPTER 2.................................... 12 2.................................................................................2........................................................................... 46 4..... EMG SIGNAL FUNDAMENTALS ............1....... DISCUSSION AND CONCLUSIONS ...........................2..
................... 86 PAPER SUBMITTED TO THE JOURNAL OF BIOMECHANICS ........................................ LBXXXX EXPERIMENT DATA FILE DESCRIPTION ......2......................... SUMMARY AND CONCLUSIONS .............................................................. 84 EXTRA FIGURES AND PLOTS ..................................................................................................................1. AND FIGURES.............. 71 6. 72 6.....1.......... Study Limitations and Future Suggestions ..................... 71 6................................................................... II........ PLOTS................... 74 REFERENCES ............ 76 APPENDICES: ADDITIONAL INFORMATION........................................................... III...... 82 I........................ Advances to EMGtorque Estimation ...1..... IV.........1.............................................6........... 82 OPTIONAL PROPERTIES FOR AMPLITUDE ESTIMATION ALGORITHM ................................................... DISCUSSION OF RESULTS.....................................................................................................................................................................2.... 92 vii ...................................
4: RAW EMG (FLEXION & EXTENSION) AND TORQUES ........................ 89 FIGURE 0............................................... MUAP [BASMAJIAN AND DE LUCA..................................................3: SURFACE EMG ELECTRODES AND ACQUISITION BOX [BOUCHARD.... 58 FIGURE 5. 91 viii ......................................... 2001].......................... 25 FIGURE 3.......8: HIGH DC OFFSET ERROR ON ESTIMATED TORQUE . 17 FIGURE 2............... 24 FIGURE 2............................7: MODEL OF EMG USED FOR ADAPTIVE WHITENING FILTERS [CLANCY AND FARRY.............6: SINGLECHANWHIT PROCESS FOR EMGAMP ESTIMATION [CLANCY ET AL.........3: SYSTEM PERFORMANCE (% VAF & MAE) USING PSEUDOINVERSE (SLOW TRACKING) ....... 65 FIGURE 5..................... 1999]...............9: SYSTEM IDENTIFICATION PROBLEM (BLACKBOX TYPE OF MODELING) .......10: GENERIC DYNAMIC SYSTEM BLOCK DIAGRAM (DISCRETE TIME SIGNALS) .... 18 FIGURE 2................ 50 FIGURE 4.....5: MEDIAN (LEFT) AND MEAN (RIGHT) OF % VAF AND % MAE FOR FAST TRACKING......... 54 FIGURE 4...........5: SIX STAGES MULTICHANWHIT EMGAMP PROCESSOR [CLANCY ET AL. 87 FIGURE 0................7: THE AVERAGE PSD OF ERROR AS ESTIMATED FROM WELSH PERIODOGRAM........ 1981] ..1: EMG ELECTRODE PLACEMENT [DE LUCA..................................3: OBSERVED MOTOR UNIT ACTION POTENTIAL.............................................................. 1997] ........... FREQUENCY ...............3: DECIMATION RATE EVALUATION PLOT . 2004]...............2: GENERATION OF ELECTRIC FIELD IN MUSCLE FIBERS [PERRY AND BEKEY.......................................................4: SYSTEM PERFORMANCE (% VAF & MAE) USING AC PART OF EMG AMPLITUDES (FAST TRACKING + PINV) ...........2: BIODEX EXERCISE MACHINE FOR THE EXPERIMENT [BOUCHARD.....4: EMG SIGNAL ORIGIN BLOCK DIAGRAM [BASMAJIAN AND DE LUCA..........1: CHANGES OF PREDICTED TORQUE WHILE INCREASING DECIMATION RATE .................. 21 FIGURE 2..................5: BLOCK DIAGRAM OF EMG DATA PREPROCESSING FOR SYSTEM ID ALGORITHM .................................................................................................. 7 FIGURE 2............ 8 FIGURE 2............. 37 FIGURE 4...... 2001] ........... 2001] ............2: SIGNAL POWER ACCUMULATION (AVERAGE PSD TORQUE) VS... 68 FIGURE 5..... 20 FIGURE 2............................1: RAW SURFACE EMG TO TORQUE MODEL [CLANCY AND HOGAN....................................................................... 62 FIGURE 5................................................................... 61 FIGURE 5.................... 66 FIGURE 5.......................... 68 FIGURE 5.. 86 FIGURE 0............ 1981] . 9 FIGURE 2................................ 2000]................................ 47 FIGURE 4...............................................1: MUSCLE FIBERS COMPOSITION [PERRY AND BEKEY. 90 FIGURE 0...................................6: COEFFICIENTS FREQUENCY RESPONSE FOR A TYPICAL EMGTORQUE MODEL (FAST TRACKING) ......................... 1985]............6: PSD OF ERROR ACCUMULATION RATE ........5: COEFFICIENTS FREQUENCY RESPONSE FOR A TYPICAL EMGTORQUE MODEL (SLOW TRACKING) ...................................1: ESTIMATION ERROR PSD (WELSH PERIODOGRAM) FOR ALL 4 PROCESSORS ...2: SYSTEM PERFORMANCE (% VAF & MAE) USING QR FACTORIZATION (FAST TRACKING) ........... 2002]........... 69 FIGURE 0... 48 FIGURE 4.................4: SUBJECT DURING EXPERIMENT [BOUCHARD.. 64 FIGURE 5..............LIST OF FIGURES FIGURE 2... 88 FIGURE 0....... 2000] .....6: SYSTEM IDENTIFICATION PROCEDURE [CREATED BASED ON LJUNG..................................................8: ADAPTIVE WHITENING OF EMGAMP ESTIMATION [CLANCY AND FARRY............ 2001].... 44 FIGURE 4.................... 11 FIGURE 2. 1985]...................
.......... 52 FOUR PROCESSORS TYPES (PROCESSOR 14) ........................... MERLETTI........ AND GENDER) ....................... AND ENOKA.............................2: TABLE 4.2: FACTORS THAT INFLUENCE SURFACE EMG [FARINA. AGE.............. SIMPLIFICATION OF GENERAL EXPRESSION .............................. 26 SUBJECT INFORMATION (CODE.......2: TABLE 4.............................................1: TABLE 4.. 49 A/D ELECTRODE CHANNELS FROM THE EXPERIMENTAL DATA ... 2004] ........... 14 COMMON BLACKBOX MODELS.................. 67 DISTRIBUTION INFO OF % MAE VALUES FOR EACH PROCESSOR (FAST TRACKING)..................................1: TABLE 0....................................1: TABLE 2.... 52 DISTRIBUTION INFO OF % VAF VALUES FOR EACH PROCESSOR (FAST TRACKING) ............................. 82 A/D CHANNEL NAME CODES ............ 67 TRIAL ID NAME CODES...............................................2: TABLE 0......LIST OF TABLES TABLE 2.............3: TABLE 5........... 83 ix ................................................................1: TABLE 5.......
CHAPTER 1. INTRODUCTION
“Electromyography is a seductive muse because it provides easy access to physiological processes that cause the muscle to generate force, produce movement, and accomplish the countless functions that allow us to interact with the world around…To its detriment, electromyography is too easy to use and consequently too easy to abuse.” [Carlo J. De Luca, 1993] The contraction of muscle fibers generates electrical activity that can be measured by electrodes affixed to the skin surface on top of the muscle group. The recorded spikes of electrical activity are referred to as the electromyogram signal or “raw” EMG. The surface EMG signal recorded using large electrodes (e.g., diameter 5 mm) that monitor the activity of multiple muscle fibers can be well modeled as a zeromean timevarying stochastic process. Motor units are the smallest functional muscle group. It is observed that the standard deviation of the raw EMG signal is monotonically related to the number of the activated motor units and the rate of their activation. This standard deviation is used to approximate the magnitude of the muscular electrical activity referred to as EMG amplitude [Clancy and Hogan, 1997]. EMG amplitude has a variety of applications, such as a control signal for myoelectrical prostheses, ergonomic assessments, biofeedback systems, and it is used to approximate the torque about a joint [De Luca, 1993; Thelen et al., 1994; Gottlieb and Agarwal, 1977; ValeroCuevas et al., 2003].
1.1. PROJECT MOTIVATION
After obtaining high quality estimates of EMG amplitudes, a common practice is relating them to the tension of individual muscles via mathematical models, even though 1
there are limitations to this method. The tension produced by individual muscles can not be measured noninvasively, thus there is no direct mechanical method to validate the model predictions. In addition, the existence of crosstalk (defined as the interfering electrical activity from the surrounding muscles) and the inability to measure this effect add to the difficulties of creating this model. Considering the mentioned limitations, many researchers [Gottlieb and Agarwal, 1977; Clancy and Hogan 1997; Thelen et al., 1994] have focused their efforts on relating the EMG amplitude to the torque about a joint as the next logical and practical alternative. The effect of crosstalk may be automatically canceled or minimized in the case of the torque about the joint [Clancy et al., 2001]. Total net torque about a joint can be easily verified via mechanical measurements. Furthermore, considering coactivation effects on underlying group muscles, the system model performance is evaluated against the net joint torque contribution, rather than the individual ones that are impossible to distinguish. Over the last few years, there are clear advances in estimating EMG amplitude yet the EMGtorque modeling has not benefited from this progress. If EMG is a useful indicator of the muscular tension, it is necessary to develop accurate means of quantification, both in terms of properly measuring and interpreting EMG and in creating mathematical models relating EMGtorque. Amplitude estimation accuracy influences the performance of EMGtorque models, because torque about a joint (tension exerted in muscles) is the outcome of proper EMG signal interpretation and consequently its careful treatments. The importance of EMG signal processing can not be emphasized enough, since
2
electromyography is such a powerful and physiologically easily obtained tool, therefore as expressed by De Luca, its misusage can lead to fatal mistakes [De Luca, 1993]. Demonstrating the benefit of utilizing advanced EMG amplitude processing, Clancy and Hogan (1997) showed that the torque estimation error is reduced when using improved EMG amplitude processors. The experiment results were obtained using a linear model to relate EMG amplitude from biceps/triceps to the elbow joint torque in the case of constantposture and constantforce contractions. Additionally, encouraging
results were also obtained in less constrained conditions (slowly varying force), but several trial combinations that lead to unrealistic model performance (considered as model nonconvergence) were an obstacle that needed further investigation [Bouchard, 2001]. The result of the previous research inspired the focus of this project: relating the EMG amplitudes from biceps/triceps to the torque about the elbow and proving that better EMG amplitude processing leads to better torque predictions during dynamic experimental tasks (forcevarying contractions).
1.2. THESIS CONTRIBUTION
The goal of this project was to demonstrate that the usage of high fidelity processing techniques (inclusion of whitening filters and multiple channels) for EMG amplitudes leads to improvements in the accuracy of estimating torque. To achieve this main
objective, it was necessary to develop a model to relate EMG amplitude to torque and compare the model performance, as the EMG amplitude processors were varied among four different types. These four types of processors were obtained using the combination of multiple channel recordings with the addition of adaptive whitening. The four
processors created were: singlechannelunwhitened, singlechannelwhitened, multiple3
channelunwhitened. along with the recommendations for improvements will serve as starting point for future research. Decimation solved some model nonconvergence problems encountered during prior research [Bouchard. there are several important deliverables from the completion of this project work. Further accomplishment was to determine the decimation rate required for the data prior to applying them to the system identification algorithms. focusing on the adaptive whitening filters as a new step that has revolutionized the existing processing methods. 2001]. CHAPTER 2 provides background information about the EMG signal starting from its recording to amplitude processing techniques. In conclusion. The first is the model used to relate extension/flexion EMG amplitudes to torque about the elbow.3. 1. and multiplechannelwhitened. The second is the data preprocessing routine (decimation) required to achieve the maximum performance from the model. 4 . The chapter ends with a brief review of the literature on EMGtorque modeling techniques. There is also a review of some of the most common system identification models. describing most of the algebraic steps involved into obtaining a linear least squares error solution. CHAPTER 2 is the model design development chapter. It includes all physiological concepts and thoughts that were poured into quantifying the EMG to torque relationship. THESIS CONTENT The content of this paper is presented in a logical and chronological. Following the background. reaching into the linear (ARX) model used in this project. the thorough documentation of the results and the steps achieving them. order as appropriate in order to explain the process involved in completing the project. The model then is solved. Finally.
During training. the document ends with the APPENDICES: that includes additional information on the experimental data and some additional plots that were not crucial to the results. Model validation requires utilizing a distinct dataset to estimate the output using the optimal coefficients. The chapter includes general observations and hypotheses derived through experimental data interpretation to validate the observations. EMG amplitude estimation. The subsequent CHAPTER 5 describes the results obtained after following the tests explained in the previous methodology chapter. but support their interpretations. Finally. This chapter summarizes the main contribution of this research and it lays out some suggestions for future work. decimation. based on the conclusions drawn. a coefficient vector is fit to the input data based on the least squares error minimization. The system identification procedure involves two main steps. 5 . and truncation are part of a preprocessing routine used prior to system identification. training and validation. Interpretation and the study limitations are discussed in detail in the last chapter (CHAPTER 6).CHAPTER 4 explains the data collection method and the process of obtaining EMG amplitudes from the four different processors. The details of the traintest paradigm along with definitions of model performance quantifiers are also explained in this chapter.
1.1. Electrical Activity Generation Electrical activity in the muscles arises from the contraction of the muscle fibers.1. The relation of surface EMG to torque makes EMG an attractive alternative to direct muscle tension measurements.1. 2. the structure of which is shown in Figure 2. the reader is suggested to review Basmajian and De Luca (1985) for more details. The chapter starts with a brief introduction of the physiological raw EMG signal. the complexity of the EMG signal origin has been a barrier for developing a quantitative description of this relation. there is a review of techniques used to process the EMG amplitude. it continues with a brief description of the At the end. However. achievements in relating the surface EMG amplitude to the torque about a joint following a summary of modeling techniques. The EMG signal origin and character is necessary background to understand the difficulty of establishing a relationship between surface EMG and torque. 2. necessary in many physical assessments. Then. EMG SIGNAL FUNDAMENTALS The electromyogram (EMG) is the recording of the electrical activity produced within the muscle fibers. Each muscle fiber contains a bunch of 6 . PROJECT BACKGROUND The content of this chapter is intended to provide background information necessary to understand the subsequent sections that describe the specific thesis contribution. focusing on the random character of EMG.CHAPTER 2. The description in this section is brief and selective.
The detected waveform resulting from the depolarization of the wave propagating between the motoneuron and end plate is called the muscle fiber action potential (MAP). 1981] Each of the myofibrils is chemically activated by local neurons. because they are recorded using microelectrodes. 7 . which contribute to the force exerted within the muscles. activating the chains of sarcomeres (Figure 2. The charge motion generates an electromagnetic field that induces volume conduction. generating an electrical charge that moves up and down the myofibril. Figure 2. The myofibrils contain long chains of contractile units called sarcomeres.myofibrils (long chains of contractile units). which enables recording of an electrical signal both internally at the muscle and externally at the surface over it.1: Muscle Fibers Composition [Perry and Bekey.2). and can not be picked up by the noninvasive surface electrodes. MAPs are not commonly seen in the general EMG literature.
2. 1998].2: Generation of Electric Field in Muscle Fibers [Perry and Bekey.1. To simplify analysis and mathematical interpretation of EMG. 1981] The muscle fibers contract in groups that are controlled by the central nervous system via nerve fibers (axons) transmitting the signal to the ending neurons. while larger ones contain up to 2000 myofibrils. 2. The number of the muscle fibers contained in a MU varies with the size of the muscle within which a MU belongs. Lamb and Hobart. It is important to emphasize the similar structure of the muscles.Figure 2. regardless of the scaling on the size and the number of myofibrils [Perry and Bekey. Origin and Character of EMG Microelectrodes on the cell surface are not the only way to measure motor action potentials. Smaller muscles have MUs that contain 310 myofibrils. 1981. the action potential 8 . The living tissues act as volume conductors. the smallest controllable functional unit of muscle fibers is defined as a motor unit (MU). The motor unit consists of a single motoneuron. therefore a potential at the motoneuron source is spread away via ion movements throughout the entire unit volume. its neuromuscular junction. Applying the same principles of conduction described above. and the muscle fibers that it excites.
Figure 2.3). 1985]. because the geometry of the conducting volume changes. The electrical potential surrounding the muscle fibers changes. Figure 2. [Basmajian and De Luca.3: Observed Motor Unit Action Potential. 1985].3 represents the motor unit action potential as the superposition of MUs generated by each of the myofibrils. the conduction times of the muscle fibers in a motor unit are different. Therefore. The spatiotemporal summation of the individual myofibril action potentials recorded by the electrode is called a motor unit action potential (MUAP). Each muscle fiber within the MU (on the left of the figure) contributes to the surface potential (on the right of the figure). 9 .propagates along the motoneuron to the endplate of the muscle fibers (Figure 2. MUAP [Basmajian and De Luca.
In particular. rather than individual MUAPs. 1977]. Typical surface EMG electrodes are used to record the myoelectric activity of the skeletal muscle as a whole. The duration of the MUAPs can vary from a few milliseconds to 14 ms. although under certain conditions surface electrodes can be used. The repetitive sequence of stimulations to the motor units results into a series of impulse responses referred to as the motor unit action potential train (MUAPT). the superposition of the MUAPTs is the physiological EMG signal and can be modeled as stochastic process (sum of independent random variables). the tissue acts as a low pass filter with a cutoff frequency proportional to the distance of the electrode to the signal source [Lindstrom and Magnusson. The MUAP is the response of the motor unit MU to a single motoneuron excitation. Each of the motor unit responses to the impulse train is independent from the sequence and the total series response has a random character. the pickup area of an electrode includes more than one motor unit. then the MUAP is considered the impulse response h(t). Therefore. the individual MUAP is recorded using fine wire electrodes. 10 . δ(t). 1992]. and their amplitudes vary from microvolt ranges to a maximum of 5 mV. Generally. If the stimulus is modeled as an impulse dirac function.The recorded MUAP is an attenuated version of the action potential generated in the muscle fibers because of the filtering effect that is due to the transmission line between the motoneuron and electrode. because muscle fibers of different motor units are mixed throughout the entire muscle [Lamb and Hobart. Usually.
represents the physiological EMG and it is not recordable or measured. The symbol mp(t. F) that is 11 . The detected EMG signal that is utilized in the research is the observed signal m(t.4. 1985] A schematic representation of the EMG generation is shown in Figure 2. myoelectric signal as a function of time (t) and the number of firings (F).4: EMG Signal Origin Block Diagram [Basmajian and De Luca.Figure 2. F).
2002]. Merletti. and position of the recording electrodes. the recorded EMG signal is dependent on the type. This random character of the EMG signal enables the later described approximation of EMG amplitude as the square root of the detected signal’s variance. and Enoka. This section also will follow the same rule. and delayed by 50100 msec. which is slower than the electrical response. In addition. 2004. 1992]. because there is not enough information to validate the assumptions.3. This mutual relation of EMG and mechanical activity to the MUAPs inspires the establishment of an EMGtorque relationship that will be discussed in detail in the upcoming chapters. Even though they all are important. Lamb and Hobart. and briefly describe some of the factors that directly effect the EMG signal interpretation and analysis when estimating torque. 2. The depolarization wave also causes chemical changes that result in a mechanical twitch.contaminated with electronic noise (almost white) and has lost some of the high frequency components due to the filtering effects at the electrodes [De Luca. Factors that Effect EMG Signal There are many factors identified in the research as having a great influence on EMG interpretation. geometry.1. Quantifying the factors that effect EMG signals is a complex task. 12 . 1993]. To conclude. it also is impossible to generalize the observations over all subjects and cases. a common practice among researchers has been to focus on the effects that have the most impact on the application for which the EMG signal is used [DeLuca. 1993. because EMG is the sum of a large number of MUAPs [Papoulis. 1981. Farina. considering the EMG signal as a time varying stochastic process gives the possibility to model it as a zeromean Gaussian distribution. Perry and Bekey. Considering the varieties in the structure of electrodes and living tissues.
The extrinsic factors are related to the electrode structure and its placement on the skin overlying the muscle. Such instances include the electrode configuration. and the orientation of detection surfaces relative to the muscle fibers. intermediate and deterministic factors. superposition. anatomical and biochemical character of EMG signals. The causative intrinsic factors include the number of active MUs at the time. On the other hand.e.e.). the intrinsic causative factors are related to the physiological. MUAP shape and duration. etc. crosstalk.) are the effects that are influenced by the causative factors and in consequence they influence the deterministic factors (i. location.De Luca (1993) categorizes the factors that effect EMG signal and force into three groups: causative. (2004) represents a summary of the known effects to EMG interpretation. conduction volume and velocity. because the loss of the high frequency components reduces the spectrum of the EMG signal. the blood flow. the pH level in the muscle fibers. number of MUs activated. etc. and geometry of the fibers. Crosstalk is defined as the interference pattern recorded from a distant muscle when the electrodes are intended to monitor another muscle. the issue of crosstalk is always present. The intermediate factors (i. MU firing rate. The causative factors are the basis of EMG signal and they are both intrinsic and extrinsic. Crosstalk is an issue that can be misleading when EMG is 13 .1 from Farina et al. but their knowledge and understanding help with the accuracy of EMG interpretation. The presence of subcutaneous fatty tissues becomes a significant factor. Besides the stability of the position of the electrodes and the stability of the MU firing rate. Table 2. These factors can not be controlled. The amount of the effect that the deterministic and the intermediate factors have on EMG is an applicationbased evaluation.
2004] Nonphysiological Anatomic Shape of the volume conductor Thickness of the subcutaneous tissue layers Distribution of the MUs territories in the muscle Size of the motor unit territories Distribution and the number of fibers in the MU territories Length of the fibers Spread of the endplates and tendon junction within MUs Spread of the innervations zones and tendon regions among MUs Presence of more than one pinnation angle Skin electrode contact (impedance or noise) Spatial filter for signal detection Interelectrode distance Electrode size and shape Inclination of the detection system relative to the fiber orientation Location of the electrodes over the muscle Muscle fiber shortening Shift of the muscle relative to the detection system Conductivities of the tissues Amount of the crosstalk from the nearby muscles Average muscle fiber conduction velocity Distribution of the MU conduction velocities Distribution of the conduction velocities within in MUs Shape of the intracellular action potential Number of recruited MUs Distribution of motor unit discharge rates Statistics and coefficient of variation for discharge rate MU synchronization Detection System Geometrical Physical Physiological Fiber membrane properties Motor unit properties 14 . it should be recognized while utilizing EMG to estimate muscle forces [Farina. and Enoka. Merletti. Therefore.1: Factors that Influence Surface EMG [Farina. Simulation and analyses have shown that the crosstalk can neither be measured nor eliminated with the existing technology. 2004]. Table 2.explained by the properties of volume conduction. Merletti and Enoka.
such as a full wave rectifier and a low pass filter made of simple passive components (resistors and capacitors). 1981]. According to Hof and Van Den Berg (1981). to detect the signal [Bigland and Lippold. 1997]. its statistical processing can be used for predictive purposes. the recorded EMG signal is described as the product of a zeromean stochastic process with the timevarying EMG intensity.. then it can be estimated by applying standard statistical techniques [Clancy and Hogan.2. 1952]. 1954]. 1 Moving Average Mean Absolute Value: MAVt = N Moving Average Root Mean Square: RMS t = i =t − N +1 ∑x 2 i t i (2. The state of art EMG amplitude processing includes six stages that will be discussed separately after a brief presentation of the complete process.1) 1 N i = t − N +1 ∑x t (2.2) 15 . This method eventually led to the use of the statistical moving average mean absolute value (MAV) and the moving average root mean square (RMS). The estimation of the EMG amplitude has been refined and improved since the early EMG amplitude developed from a simple rectifier and lowpass filtering [Imnan et al.2. Standard EMG Amplitude Estimation The most common technique of detection for EMG amplitude is the rectification process followed by a smoothing step. Therefore the intensity of the EMG signal (EMG amplitude) can be obtained by proper rectification and smoothing [Hof and Van Den Berg. The early researchers in the field studied and utilized nonlinear analog circuits.2.1. Since raw EMG is a stochastic process in nature. 2. SURFACE EMG AMPLITUDE ESTIMATION TECHNIQUES If the EMG amplitude is defined as the standard deviation of the raw EMG signal.
The detection for the MAV method is done by taking the absolute value of the terms. because of the lesser amount of time necessary for computations. the detection of the signal is achieved by squaring all the terms.2. The process of detection is followed by smoothing and relinearization. specifying and understanding the steps involved in the processing technique is extremely important. Advanced EMG Amplitude Estimation The EMG signal processing is a crucial factor in the way that EMG amplitude is interpreted and used in different applications. Adaptive whitening 3. Currently. there is no need to relinearize. The result is smoothed by taking the average of these terms. Therefore. An advanced EMG amplitude estimator consists of the following six stages (Figure 2. the computation time is not as problematic especially when processing is performed offline. However. Multiple Channel Combination and Gain Normalization 16 . The amplitude estimates found using RMS and MAV calculations exhibit very similar performance. EMG amplitude can also be computed in software using either one of the above formulae. The method of accomplishing the two last steps differs between RMS and MAV. The resulted squared terms are smoothed by taking their average and then relinearized by taking the square root of the mean. the MAV method has been initially used more than the RMS. t is the time at which this interval starts. and xi is the signal being smoothed in the timedomain. 2.where in both expressions N is the number of samples in each smoothing window of the moving average filter. In this case. Noise rejection filter 2. In the case of RMS.2. The estimator has evolved from the use of a simple rectification and a lowpass filter.5): 1.
. the demodulated samples are averaged (smoothed). Smoothing 6.6. ˆ Each of the surface EMG signal mk is transformed to the EMG amplitude s k after passing through all the stages of the processor. In the second stage. In the first stage. In stage four. motion artifact is attenuated with a highpass filter. the signal is 17 .4.5: Six Stages MultiChanWhit EMGamp Processor [Clancy et al. The adaptive whitening has demonstrated better performance for lowamplitude levels. the signal is whitened. Relinarization Figure 2. Stage three rectifies the signal and then raises it to a power to make it nonlinear. The output s (t ) is the estimated EMG amplitude (EMGamp). In stage five. inputs mk (k = 14) are the recorded signals from the surface ˆ electrodes placed on top of each of the muscle groups. The pictorial presentation of the signal transformation for each of the channels is given in Figure 2. Rectification and Demodulation 5. 2001] In the above figure.
6: SingleChanWhit process for EMGamp estimation [Clancy et al.. 2. are used to eliminate the noise from motion artifact. prior to RMS and MAV. Noise Rejection Filters High pass filters.3. Additional detail of these steps is given in the following sections. therefore. 2004] The smoothing step is omitted when the EMG amplitude obtained is used to estimate torque. d=1 for MAV and d=2 for RMS. The power density of motion artifact is mostly below 20 Hz.2.relinearized by raising it to the inverse of the power applied previously. a 18 . Figure 2. During the “Detect” and “Relinearize” stages.
The adaptive whitening removes the additive noise described in the physiological model of EMG created by Clancy and Farry (2000) presented in Figure 2. Adaptive Whitening The whitening step is recently included in EMG signal processing software algorithms. 1986]. considering that the rolloff of the real filters can coincide with the median frequency of the EMG signal.4.highpass filter with cutoff frequency between 1020 Hz is sufficient to reduce/eliminate these effects. 19 . The term whitening originates from the power of the white light spectrum spreading out uniformly over all frequencies. Doing so.2. especially during fatigue [Clancy. The advantage of using digital filters is the ease of implementing high order filters to achieve sharp rolloff and eliminate more of the noise power ensuring that the loss of useful information is minimal. Morin. The high pass filter can be analog incorporated into the hardware instrumentation and/or digital implemented in software. the statistical bandwidth increases therefore the approximation of standard deviation is more accurate [Bendat and Piersol.7. 2. Cutoff frequencies greater than 20 Hz can cause loss of EMG signal. 2002]. Whitening an EMG signal is the process of decorrelating the neighboring samples in the time domain. and Merletti. In some cases. analog filters are used in addition to digital filters to prevent saturation caused if the EMG signal is corrupted by large amplitude motion artifact.
2000] A more detailed description of the model and the math behind it can be found in the original source. Recalling the physiological description of EMG in the previous section.4.7: Model of EMG used for adaptive whitening filters [Clancy and Farry. This signal is passed through a shaping filter. Htime that creates the lowpass effect of the tissues and skin layers on real EMG signal while still maintaining unit standard deviation. mi. the signal wi is a zero mean Gaussian random process of unit variance that serves as a start for modeling the EMG. The output (ni) is then multiplied by the amplitude of EMG (si) resulting to the noisefree EMG (ri). Adaptive whitening is necessary. The signal vi is a zeromean random process representing additive electronic noise and random noise from the electrodeskin interface that is summed with ri to complete the model of the measured surface EMG. The shape of adaptive whitening filters is formed based on the power spectral density (PSD) of the noiseless signal and the additive noise. the shape of the original whitening filter is the inverse square root of the PSD taken from the true EMG signal. because it is observed that the noise exhibits a larger relative 20 .wi H time (e jw ) ni ri Σ vi m si Figure 2. Briefly. this model is consistent with the character of the raw EMG in Figure 2. Briefly. The adaptive whitening involves incorporating a noise attenuation stage that operates based on the relative power of the signal over the existing noise.
2000] The whitening process proposed by Clancy and Farry (2000) includes three stages used to improve amplitude estimation (Figure 2. but also a filtered version of the additive independent noise vi. hence. the EMG amplitude remains essentially constant during that period. This step is used to maintain the variance of the EMG signal throughout the complete whitening stage [Clancy and Farry. Figure 2.magnitude during low level contractions. making the adaptive whitening process quasistationary [Clancy and Bouchard. 2000].8: Adaptive whitening of EMGamp estimation [Clancy and Farry. where the relative EMG intensity is lower.8). The second stage optimally estimates the noisefree whitened signal m i by adaptively removing the noise through a Wiener filter. 21 . 2001]. The third stage applies an adaptive gain determined based on the transformations of EMG signal from the two previous stages. Without including the details (they can be found in the mentioned source) the first stage of this process whitens the noiseless EMG amplitude si. The time duration of the whitening filter is short.
Research has shown that using several electrodes for measuring the EMG from a muscle results in more accurate EMG amplitude estimation. Multiple Channel Combination and Gain Normalization This step involves the combination of EMG recordings obtained from several electrodes placed adjacent to each other. even though most of the time their difference does not affect the modeling process. BIOMECHANICAL SYSTEM MODELING TECHNIQUES System identification is a study of the dynamics and physical behavior of systems under external disturbances. 1980b]. The reason for the combination of multiple channels is that the SNR improves with the increase in the volume of muscles recorded. Specifically. 1980a]. 2. it is a set of standardized guides on building system mathematical models based on observations made on system reactions.3. shorted electrodes. is increased with the number of channels [Hogan and Mann.2. etc. This ensures equal contribution from each of the recordings. The external data that can be manipulated and measured by the user are referred to as inputs and others as disturbances.5. Since the gain and the distance from the muscle differ from electrode to electrode. There are also some disadvantages to the multiple electrode recording combination including that the chance of defects that may arise due to noise. SNR performance improvements of up to 91% have been observed using multiple channels. the combination of the recordings is followed by the gain normalization process. on the skin overlying the same muscle.2. and can be considered as decorrelation of the signal spatially. as compared to the results from a single channel processing [Hogan and Mann. The measured/observed response of the system is referred to as its 22 .
however little or no improvement is seen in doing so. Overview of Modeling Techniques Modeling of the complex relationship between muscular activity and torque has been approached in two different methods. this section also illustrates the forgoing theory of mathematical modeling with some EMGtorque examples found in the literature. The parameters are flexible and well adapted to the system itself. The black box type of modeling is referred to as system identification.output. the system is unknown and it is considered as a black box (Figure 2. rather than determining the structure of the system. 23 . The experimental studies have explored both linear and nonlinear models to achieve better accuracy. Although this modeling technique is more practical than the first one. The morphological modeling technique involves designing a model based on the physical characteristics of the system. Keeping in mind the ultimate goal of this research.1. a priori (morphological) and a posteriori (black box) type of modeling techniques [Westwick.3. the results require careful interpretation and validation with the physical concepts. The dependence of the recorded EMG signal and muscle tension on mutual physiological factors inspires ongoing research work to develop mathematical models relating EMG to torque. Detailed reference of the models and system identification techniques are found in Ljung (1999).9). 2. This section gives a brief description of the system identification. and it is used to obtain a relationship between inputs and outputs. The drawback of this method is the large number of parameters that result in a high level of complexity. Some researchers have even built complex models that describe the details of muscles. while most of the times. 1995]. Additionally. it requires a thorough understanding of the system structure.
Parametric System Identification The parametric model is a set of differential or difference equations that describe the operation of the system in terms of inputs and outputs. If this last step fails to achieve the error requirements. Even though these types of systems are limited.3. The types of models described in this study are linear time invariant. which involves performance error measures. The values of the parameters are numerically estimated to give the best agreement between the experimentally measured output and the model estimated output.9: System Identification Problem (blackbox type of modeling) The construction of a model via system identification commonly involves three steps.2. where error is defined as the difference between the measured and predicted outputs. the theory developed through them can be used to approximate real systems. The second stage is narrowing the model choice to several that fit the system physical capabilities. The matching criterion is usually the minimization of the squared error. The first step is input/output data collection. 24 . These equations also include a number of parameters that can be varied to alter the behavior of the model.Figure 2. than the steps are repeated until the desired results are obtained. which is mostly completed through prior experiments. the following sections describe basics of system identification standard models. The data collection process is explained in detail in another chapter. 2. The last step is model validation.
Although many are tempted to use a large number of parameters to describe the system. Figure 2. B ( z −1 ) z − d C ( z −1 ) u (k ) + e( k ) F ( z −1 ) D ( z −1 ) (2.3) A(z1) = 1 + a1 z 1 + … + anazna B(z1) = b1 z 1 + … + bnbznb C(z1) = 1 + c1 z 1 + … + cncznc D(z1) = 1 + d1 z 1 + … + dndznd F(z1) = 1 + f1 z 1 + … + fnfznf 25 . the number of parameters to identify should be small. 1999].Parametric system identification is basically a simplification of general standard equations for dynamic systems.10 shows a general block diagram of a dynamic system. e(k) C ( z −1 ) D ( z −1 ) u(k) B ( z −1 ) z − d F ( z −1 ) Σ 1 A( z −1 ) y(k) Figure 2.10: Generic Dynamic System Block Diagram (discrete time signals) The general equation (Ztransform) for the dynamic system is: A( z −1 ) y (k ) = where. The accuracy of coefficients estimation decreases with the number of the parameters to be estimated [Ljung.
D(z1) B(z1). B(z1) A(z ). G(z1) and H(z1) are the transformations of the inputs and disturbances. F(z1).4) where the input terms next to u(k) can be grouped to form the transfer function G(z1) and disturbance terms next to e(k) form H(z1). The values of these variables are determined in the process of the system identification.Output Error BJ – Box Jenkins 26 . C(z ) A(z1).4) derived from the state space equation of the system behavior. 1999 Chapter 4]. 2. C(z1) A(z1). nc. Table 2.Auto Regressive Moving Average ARARX . nd and nf. D(z1) 1 1 1 NAME OF THE MODEL FIR – Finite Impulse Response (na = 0) ARX – Auto Regressive with eXogenous input ARMAX .Auto Regressive Auto Regressive with eXogenous output ARARMAX – combination of ARARX with Moving Average OE .3 the term zd next to coefficient matrix [B] represents the time lag between input and output which means that some leading coefficients of [B] are zero when there is a delay in the system.The polynomials represent the components used to find the transfer functions (eq. Simplification of General Expression POLYNOMIALS USED B(z1) A(z1). In other words. nb.2: Common BlackBox Models. B(z1). B(z ). The order of the polynomials is described by na. F(z1) B(z1). In equation 2.Auto Regressive Moving Average with eXogenous output ARMA . C(z1). respectively to obtain the output [Ljung. The shift operator z1 is consistent with the ztransform and the negative power represents the right shift in sampletime. D(z1) A(z1). If both sides of the equation are divided by the feedback term A(z1). C(z1). to better match the behavior of the system. then: B ( z −1 ) z − d C ( z −1 ) u (k ) + e( k ) F ( z −1 ) A( z −1 ) D ( z −1 ) A( z −1 ) y (k ) = (2. B(z1).
The common use of single/multiple input and output systems has created a specific nomenclature for each of the cases. Multiple Outputs The system identification literature describes the solutions and techniques for the single input. System identification has no restriction on the number of inputs and outputs to the model.3. A relation between EMG and torque simplifies the situation. 2. SISO – Single Input. Multiple Outputs MIMO – Multiple Inputs.Simplifying the general equation 2. Single Output MISO – Multiple Inputs. Although many studies have made a great impact in the EMG field. there is no consensus on a standardized set of 27 . superposition enables the use of the techniques for any case.3. relates torque to the magnitude of electrical muscle activity (EMG signal). expensive. and they may also not be possible presently.2 summarizes the special case of a priori type of modeling techniques. because EMG is readily obtained by either surface or wire electrodes depending upon whether the muscle group or individual muscle measurements are needed [Perry and Bekey. 1981].4. single output models (SISO). however direct measurements are unnatural. Table 2. there are several types of standard models that can be developed. Single Output SIMO – Single Input. EMGTorque Relationship Modeling There are many applications that the tension exerted by the muscle group during the various activities is useful.3 or 2. however. The assumption of torque being related to the nervous excitation of the individual muscle or the muscle group. invasive.
and the form of the forceEMG relationship are vital components for accurately determining force levels. joint angle. including variations in electrode placement. 1997]. 1983]. Principal components have been used to minimize the effects of crosstalk. Several investigators have agreed that it is necessary to incorporate the control strategy for the muscles being investigated. recording configuration and limb position. perhaps due to variations in muscle composition. the progress in obtaining EMG amplitudes is not yet incorporated into the existing models. the overlapping affects of independent variables. In addition. The interaction of muscles during contractions must be accounted for during analyses. In general. but generalization may not be possible due to the large number of assumptions and originality of the situations examined [Hughes and Chaffin. and muscular coactivation [Solomonow et al. different procedures used to record and analyze EMG also need to be considered when determining the relationship between muscular forces and the EMG signal. Various approaches have utilized relatively simple models under controlled conditions to determine the torque produced by different muscles groups about different joints. muscle length. However. It is also determined that changes in recording procedures. 28 . including: the force generation rate. significantly alter the EMGtorque relationship of the biceps and triceps brachii [Woods and BiglandRitchie. predicting torque is difficult because so many factors can influence the resulting exertion. The muscle being investigated. 1990].models that relate a specific muscle (muscle group) to tension (torque). The development of generic prediction models has been less successful. procedures implemented.
They have also found that other muscles. 1969]. such as the biceps and triceps. Woods and BiglandRitchie (1983) investigated the degree of linearity in the torque to EMG relationship and found that linearity existed for muscles such as the adductor pollicis and soleus. Others have concluded that surface EMG. 1983]. Furthermore. behaved nonlinearly from 030% MVC (maximum voluntary contraction). after processing using rectification and integration. varies linearly with tension generated at a constant muscle length or during contractions with constant velocity [MilnerBrown and Stein. Characteristics of the muscle of interest may also influence the EMG to torque relationship. Moritani and DeVries (1978) determined that a linear relationship existed between the electrical muscle activity of the biceps brachii and the muscular tensions produced during exertions. the dependence on frequency coding (the frequency of the incoming action potentials) for force modulation in the muscles results in linearity while muscles such as the bicep brachii recruit throughout the total range of 29 . The main fiber type can also influence the linearity with slow twitch muscles behaving more linearly as compared to the nonlinear characteristics of fast twitch fibers [Zuniga and Simons. Muscles of uniform fiber composition exhibit a linear relationship while a random noneven composition of fibers behaves more nonlinearly [Woods and BiglandRitchie. 1975]. On the other side. In addition. and then linearly above this range.Studies of the relationship between surface EMG and force have found that there exist both linear and nonlinear relationships. the muscles display nonlinear behavior at lower torque levels due to selective recruitment of motor units at different distances from the electrodes.
1983]. Additionally. the relationship between integrated. or rectified EMG and muscle force depends on the physiological characteristics of the muscle. and velocity characteristics within muscles are nonlinearities that affect the overall relationship [Perry and Bekey. 1969]. but other factors must be also considered. Woods and BiglandRitchie (1983) have found that under isometric conditions. Moreover. In addition to muscle characteristics. If the muscle mechanics are known. fatigue. on the other side.force and behave nonlinearly. Recent advances in the research field have demonstrated that linear models can predict shoulder forces during isometric contractions [Laursen et al. type of measurement. whereas tension. the electrode arrangement. MilnerBrown and Stein (1975) concluded that there was a simple linear relationship between surface EMG and force within the first dorsal interosseus muscle of the hand. they can be incorporated into a Hilltype model that can be used to predict muscle forces [Dowling. synchronization. length. 1983]. The velocity of contractions and the tension produced can be related using Hill’s hyperbolic equation 30 . 1981].. 1997]. and tension nonlinearity also influence the behavior of EMG to torque relationship. with the discontinuity at approximately 30% of the maximum voluntary contraction [Woods and BiglandRitchie. However. Linear algebraic equations may not suffice when attempting to explain dynamic situations. and level of physical conditioning level may influence the apparent EMG to torque relationship [Zuniga and Simons. 1998]. smoothed. The degree of linearity is dependent on the muscle being investigated. Frequency coding has been shown to increase the linearity [Ray and Guha. Zuniga and Simons (1969) determined that there is a nonlinear relationship between averaged EMG potential and muscle tension. MilnerBrown and Stein (1975) suggest sampling bias.
Grant et al. but nonlinear models seem to capture more of the physiological behavior. Although clear progress has not been made toward development of generic models. some of the models developed for specific cases have made impact in the field. Sommerich et al. many muscles seem to exhibit a linear relationship between force and EMG. (1998) studied typing tasks in an attempt to determine a doseresponse relationship for general hand intensive tasks and create generic biomechanical assessments. (1993) used surface EMG and anatomical parameters to estimate isometric muscle forces about the wrist using an EMG coefficient method.” torque at the wrist could be estimated with coefficients of variation less than 10%. Several studies have examined muscle torques produced about the elbow. Armstrong et al. 1981]. Buchanan et al.[Perry and Bekey. Furthermore. All the above show that the degree of linearity depends on the muscle being investigated. including the lack of repeatability and restriction to “static isometric conditions. The usage of linear or nonlinear model depends on the focus of the research work and it is really a matter of perspective of the researchers. a model created by Wyss and Pollak (1984) approximated muscle forces about the elbow with 10% error. (1982) used rectified EMG signals of the forearm flexor muscles to predict the finger forces produced during tasks involving pinching. grasping and pressing. The EMGtorque 31 . Although there are limitations with this model. A multichannel surface EMG approach by Clancy and Hogan (1997) was used to develop a third order polynomial algebraic relation with an estimation error of approximately 3% to predict torques about the elbow. (1994) predicted grip force from EMG measures and ratings of perceived exertions. and reported that as much as 74% of the variation could be explained.
There is no consensus on the degree of linearity. Secondly. While the accomplishments have made an impact in the field. because the findings are influenced by many factors. most of the earlier (more than two decades ago) investigators assumed that the antagonist muscle can be safely neglected. The most relevant factor that should remain from this review is the necessity for a method to obtain accurate torque estimation.. First. The above experimental studies are not constrained only to static conditions. 1994. 1985]. In summary. 1989]. there are clear problems that still exist in some of these studies. McGill. The level of the details and the type (physiological or blackbox) on the various EMGtorque models is also relative to the focus of the study and it is driven by the main objective. improving the accuracy of 32 . since the mechanical activities of agonist and antagonist muscles are considered independent from each other. there is a substantial amount of work investigating surface EMG to torque models which confirms the importance of utilizing EMG as a physiologically powerful tool... such as the dynamic range. some researchers neglect the importance of it while some others go beyond and suggest calibrating to each subject separately [Hasan and Enoka. the level of the force contraction. even though calibration is a common practice nowadays. individual progress has been made establishing both linear and nonlinear relationships for quasiisotonic (slowly force varying) and even extending to fully dynamic conditions. and the size of the muscles.relationship of abdominal muscles required quadratic regression but still did not account for all of the variation around a linear regression line [Stokes et al. 1995]. 1992. Extensive work has been conducted on the lumbar musculature during static and dynamic situations with EMG based models being in the focus [Hughes et al. Nussbaum et al.
several important contributions in the literature.torque predictions. such as combination of multiple surface EMG recordings to improve the SNR and the adaptive whitening filters to improve the statistical bandwidth. 33 . whereas many of the above researchers examined fully dynamic tasks. have not yet been used to improve torque estimations. Although the same objective is intended. This present research is anticipated to further investigate the results for more profound knowledge and to reexamine the encountered model convergence problems. This prior research also modeled EMG amplitude to torque incorporating both agonist and antagonist muscles during tasks that involved force varying contraction. These simplifications were intended to allow for an assessment of the newly developed EMG amplitude processors. The results were positive demonstrating a clear improvement in torque prediction when advanced processing techniques were used to obtain EMG amplitudes. The experimental data used for the present research thesis are carried over from previous research. it avoided several of the earlier mentioned difficulties by examining constant posture efforts (similar to strict isometric conditions). However.
The concept of torque for the skeletal muscles is derived from the motion of the bones about a joint due to muscle contractions. SURFACE EMG TO TORQUE MODEL DESIGN In the previous chapter there was a brief review of the literature achievements on EMGtorque relationship. Using the principles of system identification.1. 34 . the EMGtorque relation can alleviate some issues such as measurements for mechanical verification. cocontraction. Since the EMGtension relationship for each individual muscle is not possible. but rather demonstrate the importance of incorporating the advances of EMG amplitude processing into each model.CHAPTER 3. the results presented in later chapters will display the improvements in EMGtorque model performance as a function of EMG amplitude method. Hence. therefore researchers create models that best fit their design application or that are derived from earlier experimental work (Hilltype model). there is not any generic model established yet. This chapter describes the design process of a linear EMGtorque model that will be used to compare four different types of processors. Since surface EMG signal measures the activity of the skeletal muscles. 3. and crosstalk. a mathematical relationship can be established between the EMG amplitude and net joint torque. The primary focus of this research thesis was not to find the best model. the model is standardized to a parametric type ARX (FIR) model. As mentioned. EMGTORQUE MODEL DESIGN This section is a summary of the theory involved to design the EMGtorque model.
For tasks that involve slow force variations. Physical Interpretation of the EMGTorque Model The level of the tension produced in the muscle is controlled through the recruitment of motor units and their firing rate adjustments. easily obtained.1. 1974]. which are not necessarily the ones under investigation. The tension developed by the muscle also depends on both the conduction velocity and the geometry of the muscle fibers. as the tension level varies from low to high. known as crosstalk and already discussed in Section 2. is one of the main factors affecting EMG signal interpretation.1.3. there are two fundamental issues with this model. Since both electrical and mechanical activities are mutually related through several mentioned physiological parameters. The surface EMG is a noninvasive. However. As mentioned. it is extremely difficult to measure EMG from only one muscle. measure of electrical activity in the skeletal muscle. Additionally. the low frequency motor units are the first to be activated while the ones with high minimum frequency are the last [Hannerz.3. surface mounted electrodes capture EMG generated from the muscles in a surrounding area. the dependence on the conduction velocity and the geometry dependent parameters also contribute to the nonlinearity. First. This condition. relating EMG amplitude to muscle tension would be ideal.1. In practice. there is no practical method for 1 Based on the EMG models created for isometric low force level contractions 35 . The motor recruitment is hypothetically1 done orderly based on the size of the muscle fibers. The assumption that muscular force depends only on the firings of the motor units (rate and number of units) makes the relation between EMG to force nonlinear in a sense that the number of MUs recruitment is higher for the higher contraction levels.
The prediction of the net torque requires the usage of both agonist and antagonist muscles. assuming that the agonist muscles are inhibited while the antagonist ones are contracted. Muscles that perform a desired action are known as agonist muscles. 1983. in the cases of 50% MVC it is necessary to acknowledge the contribution of both agonist and antagonist muscles. it is not as influential to the net torque estimates [Clancy. the model performance can be easily quantified by comparing the torque estimates to the actual torque about the joint that can be measured using a dynamometer. On the other hand. 2001]. Vredenbregt and Rau. Therefore. 1973. Although unlike in the case of EMG to individual tension relation. These two factors prevent the use of EMG amplitude to force models in terms of isolated muscle contribution as a method for evaluating performance of an EMG amplitude estimator [Clancy and Bouchard. whereas those that oppose the action are antagonist. the net torque is a result of the inhibition or agonist muscles [Lawrence and De Luca.accurately measuring the tension provided by an individual muscle. Zuniga and Simons. Hasan and Enoka (1985) have experimentally determined the existence of cocontraction in contraction levels exceeding 20% MVC. Another mechanical activity commonly related to EMG is the torque produced about a joint as muscular force is exerted. and Merletti. In addition. 36 . The problem of crosstalk is still present. Doing so. Some researchers have separated the contributions of agonist and antagonist muscles. 1969]. Morin. Woods and BigglandRitchie. 2002]. 1983.
EMG Flexion 1 EMG Flexion 2 EMG Flexion 3 EMG Flexion 4 Flexor EMG Amplitude Estimator F (n) Flexor EMG Amplitude Decimator F (k) Flexor EMG Amplitude to Torque Estimator TF + T EMG Extension 1 EMG Extension 2 EMG Extension 3 EMG Extension 4 Extensor EMG Amplitude Estimator E (n) Extensor EMG Amplitude Decimator E (k) Extensor EMG Amplitude to Torque Estimator Σ TE EMG amplitude System ID Estimated Torque Raw EMG Recordings Figure 3. After amplified. filtered.1: Raw Surface EMG to Torque Model [Clancy and Hogan. 1997] Four surface electrodes affixed on top of the muscles (biceps and triceps) record the EMG signals. The amplitude estimates are used as two inputs to a system identification algorithm to predict the net torque about the joint (T). and sampled they are applied to the EMG amplitude processor. The EMG amplitude estimations for flexion F(n) and extension E(n) are decimated to obtain F(k) and E(k) respectively. 37 .
The linear models are widely used because of the simplicity in their design associated with the linear least squares solution [Inman et al. 1994. Vredenbregt and Rau. Gottlieb and Agarwal (1977) related EMG to torque using the following transfer function: G (s) = F ( s) 1 1 =k ⋅ E RA ( s ) T1 s + 1 T2 s + 1 (3. 3. as mentioned in section 2. the model is designed based on their algebraic sum.3. 1952. The internal change in the muscles may be produced by processing the EMG signal [Perry and Bekey. The EMG amplitude processing (first stage from the left of Figure 3. 1986.1..2. Mathematical Modeling for EMGtorque Several studies. It is not clear whether nonlinear or linear models are the best choice. Thelen et al. rather than their independent contributions. al. Woods and BigglandRitchie. 1969]. it is assumed that the data are available and ready to be used in the EMGtorque model.Figure 3.1) and the data preprocessing (second stage in the same figure) stages will be discussed in detail in subsequent chapters.3. have determined that there exists a mathematical relationship between EMGtorque (linear or nonlinear). 1981]. Zuniga and Simons. whereas the researchers using nonlinear models argue that they better describe the EMG physiological nature [Solmonow et. 2001]. For now. 1973.1 shows a block diagram modified from the diagram from Clancy and Hogan (1997) and represents the model that is used to predict torque about the elbow joint from biceps and triceps muscle groups considering both agonist and antagonist muscles. Clancy et al. Without generalizing the model results.1) 38 . Even though the contributions of flexion and extension are attributed to agonist and antagonist muscles. 1983.
The model form is written as: T (k ) = − e1 E (k − 1) − e2 E (k − 2 ) − + f1F (k − 1) + f 2 F (k − 2 ) + − enb E (k − nb ) + f nb F (k − nb ) (3. linear.2.g. Even though it may require higher order to capture all the dynamics of a system. The conversion amongst them is straightforward if the sampling rate is known.where ERA is the EMG amplitude obtained from the averaged rectifier output [Gottlieb and Agarwal. Bouchard’s model includes the feedback matrix of the ARX model. The nb is the model order as defined in section 2. The proposed model is an FIR (zerosonly) model and the degrees of freedom depend on the order of the system. because it does not include any physiological parameters (e. In equation 3. FIR model [Ljung.3. 80–83]. where k is the decimated discretetime sample index] and extension [E(k)] EMG amplitudes were related to torque [T(k)] via the dynamic.2) Referring to Figure 3.1 that shows the complete EMGtorque modeling process and the above equation. the ei represent the extensor model coefficients. 1999. Another limitation is that the model can be used only for “isometric” tasks.2. but it is agreed that it is easier to work with difference equations that represent the system for discrete time signals (samples). a zerosonly model is mathematically possible and easier to implement. 1977]. This model is the Laplace transform of the differential equation with two degrees of freedom (time constants) and it assumes a continuous time domain system/signal. The model mathematical expression in 3. The model used for the purpose of this thesis is created by modifying the EMGtorque (ARX) model used by Stephan Bouchard (2001).2 does not 39 . which is equivalent to the poles of a system. joint angle). the fi are flexor model coefficients. the decimated flexion [F(k). pp.
. therefore only the final results are presented.2...E (k − nb) .change for the timevarying EMG amplitudes that are not decimated. where N is the number of samples.... .2.2 [Ljung..E (k + N − nb) F (k ) F (k + 1) ..2. . The inputs to this EMGtorque model φ(k) are written as: E (k ) E (k + 1) ϕ (k ) = ... therefore the equation is adapted for the MISO case. The system is dual input – single output.. . E (k + N − 1). (3....F (k + 1 − nb) . E (k + N ) E (k − 1).F (k − nb) 40 .. Analogous coefficient matrix is given by: θ = [e1 .5) The vector of the known input EMG amplitude samples (decimated) are represented in the φ(k) vector. The adjustable parameters in the equation are put in matrix form θ = [a1 a2 … ana b1 b2 … bnb] T..3) The FIR model used for EMGtorque is the special case of ARX.. In equation 3.+bnbu(tnnb) + e(t)...F (k + N − nb) . F (k ). Ljung writes the general model as: y(t) + a1y(t1)+… + anay(tna) = b1u(t1)+. 1999]. ... 3. This model is obtained using the standard linear difference equation of the ARX model that is found using the information given in Table 2. obtained by setting the parameter na=0.. is the FIR version of the ARX model..... MODEL SOLUTION The model represented in equation 3. If the measurements are repeated over time then the vector φ(k) becomes a matrix with N rows....4) (3... E (k ). In general N >> nb for the system identification to be possible. .. .e2 … enb f1 f2 … fnb] T (3....... F (k + N ) F (k − 1). the coefficients e1nb and f1nb represent Ljung’s b1nb coefficients..E (k + 1 − nb) .. F (k + N − 1)..
The output matrix T(k) for N samples is given by: T (k ) T (k + 1) T (k ) = . the minimum value of θˆ is found by computing the gradient of the difference between the measured output and the predicted output T(k)φ(k)θ. The gradient of the error matrix is expressed as: ∆E (θ ) = −T (k ) T ϕ (k ) − T (k ) T ϕ (k ) + 2T (k ) T ϕ (k )θ (3.7) The solution can not be obtained by taking the inverse of the data matrix. T (k ) = ϕ (k ) ⋅ θ T (3. because the inverse of the matrix exists only for square matrices (in this case the number of the unknowns exceeds the number of the existing linear equations). T (k + N ) (3. Therefore the coefficient matrix producing minimum error is equal to θˆ .8) Applying vector calculus concepts. Therefore there is not a unique solution..7 to obtain the coefficient matrix. The optimal value for θ is referred to as θˆ and is given as: ˆ θ = arg min ( T ( k ) − ϕ (k ) ⋅ θ ) θ = arg min[(T (k ) − ϕ (k ) ⋅ θ ) T (T (k ) − ϕ (k ) ⋅ θ )] θ = arg min[T (k ) T T (k ) − T (k ) T ϕ (k )θ − θ T ϕ (k ) T T (k ) + θ T ϕ (k ) T ϕ (k )θ ] θ (3. but rather a method referred to as linear squares error minimization of error can be used to obtain the “bestfit” coefficient matrix.. Hence. [φ(k)]1..9) 41 .6) The problem involves solving the equation 3. the minimum is denoted as the value at which the gradient of the matrix is zero.
an estimate of the torque about a joint can be calculated as the product of the EMG amplitude estimates from both extensor and flexor muscles and θˆ . In practice. then the resulting θˆ is used to estimate the torque produced by another independent “test” dataset.10) θˆ = [T ( k ) T ϕ ( k )] ⋅ [ϕ ( k ) T ϕ ( k )] −1 After calculating θˆ . yields to the minimization matrix in the least squared sense. 42 . − T ( k ) T ϕ ( k ) − T ( k ) T ϕ ( k ) + 2ϕ ( k ) T ϕ ( k ) ⋅ θˆ = 0 2ϕ ( k ) T ϕ ( k ) ⋅ θˆ = 2T ( k ) T ϕ ( k ) (3.Equating the expression to zero. and solving for θ. More on the testing procedure will be described in subsequent chapters. the coefficient matrix θˆ is computed using a “training” dataset. The difference between the estimated torque and the real torque in the test dataset is the error and it is used to quantify the model performance.
However. The entire preprocessing procedure is presented in a block diagram form and it is utilized as necessary during performance testing. The EMG amplitude estimates are also decimated to improve the model performance by eliminating erroneous spikes existent in estimated torque as will be in detail discussed later. Model validation requires utilizing a distinct dataset to estimate the output using the optimal coefficients. electrode placement. skin effects. EMG DATA COLLECTION Recording the EMG signal using surface electrodes faces challenges related to the signal power and external/internal noise.1. errors due to recording devices. As explained earlier. Then.1. since the experiment is not part of this thesis 43 . it continues by describing the process of obtaining EMG amplitudes from four different processors. System identification involves two main steps. The traintest along with model performance measures procedure is explained in detail in the present chapter. some of the internally generated noise found in EMG system can neither be eliminated nor reduced. During training. The description of the apparatus and experimental procedure is brief.3) can be reduced by taking the necessary precautions. DATA COLLECTION AND ANALYSIS METHODS This chapter explains the data collection method that followed some of the described literature suggestions. a coefficient vector is fit to the input data based on the least squares error minimization. 4. training and validation. and some other external error sources (2.CHAPTER 4. This data collection section explains some of the potential noise sources along with some suggestions for minimization and then it continues by describing the experiment conducted to collect the EMG data used in this project.
Noise Reduction Precautions One of the most important sources of error in EMG recording is the placement of the electrodes. The center position between the innervation on the top and myotendons on the bottom is preferred (see Figure 4.1: EMG Electrode Placement [De Luca.1). 2000]. and enables electrical coupling. More details can be found elsewhere [Clancy. Bouchard. Clancy and Farry. increases the chance of short circuits. 1999.1. 2001. 2002]. Figure 4. 2002] 44 .contribution. 4. the PSD of the recorded EMG signal depends on the location of the electrode in relation to the innervation zone (MUAP generation site) the myotendon zone and the lateral edge of the muscle [De Luca. In addition.1. The tight spacing of the electrodes in a multiple channel recording can produce correlated signals.
and Merletti 2002] Surface electrodes suffer from motion artifact due to displacement and deformation (stretch) of the underlying skin. or lipid solvents.The principle of EMG electrode functionality derives from a layer of charge created in the interface between the metal of the electrode and an electrolyte solution. Morin. To reduce its impedance. Both effects are minimized by cleansing the skin with solvents. rubbing alcohol. skin is prepared using conductingpaste. Morin. and Merletti 2002]. In addition. The charge layer creates a potential gradient that translates into voltage picked up by the electrode. The cumulative effect of the voltage created by the product of the skinelectrode interface impedance and the displacement current in addition to the magnetic coupling in some 45 . rubbing a conductive paste. which if exposed to a varying magnetic field (electric field) can produce alternating currents (150Hz). [Clancy. Motion artifacts can also be reduced through signal conditioning both online and offline. and affixing the electrodes carefully prior to recording EMG. thus reducing the power loss even more. The voltage is dependent on the type of the electrode material. The cables of the electrodes have an intrinsic capacitance. some of the power of the initial EMG signal is lost due to the effect of skinelectrode impedance. Since the typical power density of these types of motion artifacts is below 20 Hz. hence it is crucial to use electrodes of the same material to minimize the potential difference (especially in a multichannel application). The amplifier is also chosen to have input impedance at least 100 times more than expected electrodeskin impedance. they can be largely attenuated using highpass filter at that frequency either in the processing software or integrated in the hardware [Clancy.
and the power line signal may not be in phase [De Luca. Morin. A CMRR of 90 dB is generally advised in the literature. their stability. D. although electrostatic discharges spreading through measurements can be a harmful side effect [Clancy. Rancourt. 1999. the experiment description is summarized from other sources [Clancy. The second is preferred. Merletti. thus. 2001.cases is comparable to the magnitude of the real EMG.1. The reasons for not using amplifiers with 120dB at CMRR are the price. 2002. 2002]. and 50 Hz in Europe) can have power densities larger than EMG itself and are a commonly appearance in EMG recordings. O. Morin and Merletti. Bida.1. Clancy. E.1) and literature [De Luca. The subjects had signed consent for their participation and proper human studies Parts of this section are taken from the paper submitted to the Journal of Biomechanics (attached to Appendices) authored by Clancy. because the notch filters omit real signal in addition to the noise. 4. The data collection experiment was conducted previously.2.A. 2000]. The usage of active electrodes is a better solution. Apparatus and Experimental Procedure ii The experiment conducted to collect this set of EMG data was consistent with the suggestions mentioned in the previous section (4. 2002] for noise minimization. Bouchard.S. Shielding the cables can improve the magnetic environment. because the voltage buffer transforms high impedance at the electrode to low impedance at the output where the signal is fed to the cables [De Luca. 2002]. Clancy and Farry. The effect of the power lines can be reduced by notch filters at 60Hz or by differential amplifiers that have a high common mode rejection ratio (CMRR). and Merletti 2002]. ii 46 .. The interference of power line current and its harmonics (60 Hz in U. 1999. even though current technology can provide a CMRR of 120dB.
The position of the dynamometer was maintained throughout the entire experiment [Clancy. 1999].2) used to collect the EMG data was a Biodex exercise machine (Biodex Medical Systems.. The wrist was rigidly attached to the Biodex dynamometer with a cuff at the styloid process.4). The subject's right arm was positioned in the plane parallel to the floor. 2001] 47 . Shirley.permission was taken prior to the experiment. the wrist fully supinated and the elbow flexed 90o (Figure 4. Figure 4. NY).2: Biodex Exercise Machine for the Experiment [Bouchard. with the shoulder abducted 90o. Each of the subjects were seated in a lightly cushioned seat and secured using seatbelts. The apparatus (Figure 4. the forearm oriented in the parasaggital plane. Inc.
75 cm apart. The EMG active amplifying electrodes (Liberating Technologies Inc. The ground electrode was applied over the acromion process.The skin above the investigated muscles was cleaned with an alcohol wipe and a small amount of conducting paste was used to rub the subject’s arm. and low pass filtered (fourthorder filter at 2000 Hz). transversely across the arm. Amplification stage 48 . Figure 4. 2001] The output from each of the electrodeamplifier was electrically isolated. midway between the elbow and the midpoint of the upper arm.3: Surface EMG Electrodes and Acquisition Box [Bouchard. stainless steel. model MYO115. amplified. separated 15 mm centertocenter) of each electrodeamplifier (Figure 4. Each electrodeamplifier had a gain of 725. MA) were placed over each of the biceps and triceps muscles.3) were oriented along the muscle’s long axis. and a secondorder 10–2000 Hz bandpass filter. Holliston. Adjacent electrodeamplifier centers were spaced 1. centered on the muscle midline. The two contacts (4 mm diameter. a common mode rejection ratio of 90 dB at 60 Hz.
Fifteen healthy subjects (eight male. Mansfield.1: Subject Information (Code. they performed a 0% MVC (rest contraction) and separate flexion and extension 50% MVCs for five seconds. utilizing force feedback on a computer screen. Age. LB) 02 03 05 07 08 09 10 12 13 16 17 18 19 20 21 AGE 31 49 29 65 43 60 41 62 50 28 58 41 31 23 65 GENDER F M F M F M F F M M F M F M M Subjects initially performed two 2 second MVCs each in flexion and extension.contained a negative gain configuration standard opamp with a selectable gain (1 to 25). the averages of which were used as the subject’s MVCs for the experiment. 49 . MA). seven female.1±1.7% of the root mean square EMG at 50% maximum voluntary contraction (MVC). The EMG and dynamometer signals were sampled at 4096 Hz using a 16bit A/D converter (Computer Boards model CIODAS1600/16. Recordings with the two contacts of each electrodeamplifier shorted gave a measure of equipment noise. Next. aged 23–65 years) each completed one experiment. which averaged 2. These contractions were used to calibrate the advanced EMG amplitude processors [Clancy and Farry. Table 4. 2000]. and Gender) SUBJECT CODE (Exp.
Figure 4. as a biofeedback signal. The random pursuit profile followed a uniform random distribution with a bandwidth of either 0. A computer screen displayed either their elbow joint torque (the dynamometer signal) or the algebraic difference between realtime biceps and triceps EMG amplitude.25 Hz (slow tracking) or 1 Hz (fast tracking).4: Subject During Experiment [Bouchard. The computer produced a second “pursuit” display on the screen which moved randomly between 50% MVC extension and 50% MVC flexion. The EMG amplitude difference provided a biofeedback signal that was similar in character to the torque feedback.4. albeit with increased variance. 2001] 50 .The subjects then performed dynamic (constantposture. forcevarying) target tracking contractions positioned as in Figure 4.
Subjects completed 15 slow tracking trials (three sets of five) and 15 fast tracking trials (three sets of five). channels 2 and 9 are used for flexion and extension EMG amplitudes. an amplitude estimate was produced in six stages separately for the biceps and triceps muscle groups (Figure 2. 4. Natick. The design of the EMG amplitude processor is explained in detail in the background chapter.5). four different EMG amplitude (EMGamp) processors were compared. 51 . The electrode locations encoded in the channel number are given in Table 4. Before creating amplitude estimates from the raw EMG data (stored in forms of A/D channels for each of the electrode locations). EMG AMPLITUDE ESTIMATION METHOD For the present study. In order to determine the influence of amplitude processors on torque estimation. MA). each of 30 s duration. The subject’s arm was removed from the wrist cuff between all recording trials to allow 2–3 minutes of rest to avoid fatigue. This recording is assumed to capture the 60 Hz noise as well. respectively. while this section describes the experiment in a more practical perspective.2 and further description is found in the APPENDICES: (I). In the case of single channel processing. In each case. it was necessary to calibrate the noiserejection. and spatial uncorrelation. The A/D channels 14 are used to estimate biceps EMG amplitude and channels 811 are used to estimate triceps EMG amplitude in multiple channel processing. which in this case is the 0% MVC signal recorded while subjects were fully resting.2. all data analysis was performed offline using MatLab (The Mathworks. The calibration requires the additive noise signal. adaptive whitening.
the EMG data were first highpass filtered at 15 Hz using a noncausal.3).2: A/D Electrode Channels from the Experimental Data A/D Channel 1 2 3 4 8 9 10 11 16 Contents EMG: Biceps.2. most lateral Dynamometer For all processors (Table 4. After rectification. Table 4. since smoothing was incorporated within the subsequent preprocessing step. The raw EMG signal for each muscle group from one of the electrodes located centrally on the muscle was digitally highpass filtered and rectified. most lateral location EMG: Biceps. The detection was performed with an absolute value operation (MAV. Processor 3 is a fourchannel. lateral center EMG: Triceps. most medial location EMG: Triceps. the four EMG signals from a muscle group were normalized in magnitude and ensemble 52 . medial center EMG: Biceps.4 and it has been implemented in a standalone MATLAB toolbox [Clancy. unwhitened processor. most medial location EMG: Triceps. The adaptive whitening technique is explained briefly in section 2. effective 10th order. 2004]. FIR filter.Table 4. unwhitened processor. medial center EMG: Triceps. Processor 2 is a singlechannel. whitened processor obtained by adaptively whitening the same electrode channel as Processor 1 before the rectification stage.3: Four Processors Types (Processor 14) Single Channel Unwhitened (SCHUNWHIT) Single Channel Whitened (SCHWHIT) Multiple Channel Unwhitened (MCHUNWHIT) Multiple Channel Whitened (MCHWHIT) Processor type 1 is the singlechannel. d=1) and the smoothing stage was omitted. lateral center EMG: Biceps.
The performance of the model design was evaluated using conventional measures that will be discussed in detail in this section.3. Processor 4 is a fourchannel.1. whitened processor. 4. and torque (Figure 4. 386]. Data PreProcessing Some of the data manipulation to achieve the desired results involves normalization to % MVC and A/D offset subtraction for torque as well as decimation and truncation of both EMGamp(s).averaged. the data were further processed as necessary to improve the model performance [Ljung. 4. SYSTEM IDENTIFICATION PROCEDURE After estimating the amplitudes offline using the EMG toolbox. where the coefficients were computed using a train dataset and were validated using another dataset referred to as the test dataset. The torque estimates were then obtained using a train and test procedure. it was only necessary to normalize torque to %MVC according to either flexion or extension. The EMG amplitudes estimates were to % MVE (where 100% MVE EMG amplitude level corresponding to 100% MVC) in extension/flexion units.5). Each of the four channels from Processor 3 was adaptively whitened prior to rectification. 53 . hence. 1999]. The settings for optional properties for the amplitude estimation function can be found in APPENDICES: (II). 1999 pp. Normalization to maximal voluntary contraction % MVC is a common practice when trying to relate EMG to torque [Merletti.3.
The data truncation was used to eliminate the transient effect of the filters on the signals (a. has signal power over a much lower band of frequencies than raw EMG. Because the raw EMG data were sampled at 4096 Hz. the high sampling rate is unnecessary. The flexion (biceps) and extension (triceps) EMGamps are the inputs to a system identification model in which elbow joint torque is the output. The effect is seen as corruption only at the beginning if causal filters are used and in the case of noncausal filters both at the beginning and at 54 .5 is decimation. Therefore. Thus. the EMG amplitude estimates were also produced at 4096 Hz. the EMGamps are decimated.Raw EMG (4 channels) Flexion Amplitude Estimator (EMGamp) Decimation & Truncation Raw EMG (4 channels) Extension Amplitude Estimator (EMGamp) Decimation & Truncation A/D Offset Subtraction & Decimation System ID T(k) & Truncation Measured Torque Normalize to %MVC (flex. however. The decimation rate was found by evaluating various integervalued decimation rates from 1–900. / ext. The torque output.) Figure 4.5: Block Diagram of EMG Data Preprocessing for System ID Algorithm The next step common to both torque and EMG amplitudes (EMGamps) in Figure 4. In each case. then the backward time directions to achieve zerophase) prior to downsampling.ka. a “startup” transient). the data were also low pass filtered with a cutoff frequency equal to half the new sampling rate (8thorder Butterworth filter applied in the forward.
Within each set of three. 4.1) The optimal fit coefficients were determined using a linear least squares solution.25 Hz and 1 Hz) totaling to 3 × 5 × 2 = 30 sets of EMG data for each subject. Additionally. the torque about the elbow joint was measured using a dynamometer. In addition to four channels of EMG amplitudes for each of the muscle groups. A/D converter offset was subtracted from the torque signal to account for imbalance in the dynamometer output.2.2. The input data matrix φ(k) in section 3. For convenience the model mathematical expression is repeated here: T (k ) = − e1 E (k − 1) − e2 E (k − 2 ) − + f1F (k − 1) + f 2 F (k − 2 ) + − enb E (k − nb ) + f nb F (k − nb ) (4. Each of these five subsets was recorded at two different tracking speeds (0.the end. Data were collected from fifteen subjects using a multiple channel data acquisition system. The number of samples subtracted was based on the new decimated sampling rate. for each subject. As mentioned.3. three sets of data recordings were performed. it is necessary to subtract 500 ms on each side prior to system identification (ID) and 500 ms after it. data were recorded five times using various feedback mechanisms for controlling elbow torque. Adding up the effect based on the orders of filters in each of the amplitude processing stages. it is required to fit a set of coefficients to the input data matrix (EMGamps for flexion and extension) to estimate torque. Torque Estimation Procedure The EMG data used for this project were previously collected as described.2 was created using the toeplitz routine to increment the samples of amplitude estimation (flexion and extension) and the model order as in equation 4. The system solution involves matrix inversions that can be implemented 55 . Referring to the model solution in section 3.1.
If all possible combinations are used for the data combination.1 from a different test trial. since these data were corrupted by the startup transients of the various processing filters. The error signal is defined as the difference between the measured and the model estimated torques. Specifically.using QRfactorization (qr) or the pseudoinverse (pinv). Therefore. and then tested on the four remaining trials. The model coefficients are determined utilizing a traintest evaluation routine. a comprehensive crossvalidation is sufficient for the scope of this research. (The two speeds are evaluated separately) Within a set. one second of data from the beginning and end of each error signal was removed (trimmed). there would be a total of 30 × 29 combinations of training and test data sets. PseudoInverse is the process of inverting an overdetermined linear system (matrix has more rows than columns). coefficients were fit to the input data from a training trial and then used to compute the torque via equation 4. There were a total of 180 error signals available per type of processor (15 subjects × 3 sets per subject × 1 training trial per set × 4 test trials per training trial). only the results from pseudoinverse techniques will be discussed. optimal coefficients were trained to one trial. The advantage of using the QR factorization method is computation simplicity due to elimination of redundancy [Kolman and Hill. 56 . QRfactorization expresses the matrix as the product of a real orthonormal or complex unitary matrix and an upper triangular matrix. the 15 trials at a given tracking speed were organized as three sets of five contractions. 2001]. As mentioned. Because of the large computation time. Since the results and the computation time for both methods were the same.
5): • EMG amplitude estimator type: singlechannelunwhitened. Throughout this thesis results of model performance will be evaluated using two main time domain error measures: the percent mean absolute error (%MAE) as computed for each trial % MAE = 100 ⋅ mean( Error (k ) ) . Model Performance Measures The resulting EMGtorque error signal from 180 combinations for each processor type was investigated in several ways. when more complicated models are necessary. multiplechannelwhitened • • • Integervalued decimation rate subjectively chosen from 1 to 900 Matrix inversion method either pinv. The error mathematical expression used is: Error (k ) = ∑ (Tˆ (k ) − T (k )) N k =1 (4. or qr factorization EMGtorque model order from 1 to 60 The performance of the model implemented for this research matches the performance obtained using the builtin ARX parametric model function in MatLab System ID toolbox.3. All errors were normalized to twice the torque at 50% flexion MVC. the following parameters were varied in the computations performed using MatLab (R13. multiplechannelunwhitened.In summary. (4.3. denoted %MVCF. Version 6. 4.2) ˆ where N is the sample duration of the truncated estimated torque T (k ) and measured torque T(k). The toolbox is suggested to be utilized in the future. singlechannelwhitened. The measured torque was truncated to equate the number of samples with the estimated torque.3) 57 .
and the percent variance accounted for (%VAF). Kearney.6: System Identification Procedure [Created based on Ljung. 1999] 58 . START Experimental Model Design System ID Data PreProcessing Model Structure Parameter Estimation Physiological Interpretation & Performance Measures: %VAF. 1994] as: N 2 ∑ [Error (k )] %VAF = 100 ⋅ 1 − k =1 N ∑ T 2 (k ) k =1 . (4. PSD of ERROR Model Validation NO ? YES END Does model meet criteria? Figure 4. 50% overlap). %MAE. Crago. The PSD was estimated using Welch periodograms (Hamming window.4) The power spectral density (PSD) of each error sequence was another method used to evaluate the error frequency distribution and identify the character of the error. 1024point FFT. defined by Kearney [Kirsch.
59 .6). the model performance results (%VAF and %MAE) are interpreted in the physiological sense of the real system behavior.The final stage of the system identification (System ID) procedure is the model validation. then there is need for model redesign (Figure 4. During this stage. If the expectations are not satisfied.
During prior EMGtorque work. The spikes occurred infrequently (~15% to 20% of combinations) but their magnitude caused the overall %VAF (and MAE) to be unrealistic.. PROJECT RESULTS There were several important results derived after completing the described test procedure. The necessity for decimation prior to system identification was determined because some of the prediction torque sequences exhibited a few large “spikes.. 000 % MVCF when there was no decimation (c.1. Decimation is used if the system is oversampled and if it contains high frequency noise components.f. 2001]. Then. 60 .CHAPTER 5. The time domain torque plots are on the left of the picture.” The observed errors were referred to as spikes because they had much larger amplitude compared to the test torque and lasted only a few samples (see Figure 5. This chapter starts with the decision on the final decimation rate as a solution to model nonconvergence problems that were observed during prior research. 5. the trials exhibiting the described error were considered to be “nonconvergent” [Clancy et al. time sample about 2*10^4 when d=1). Figure 5.1). it continues with a graphical description of the model performance where the outcomes are contrasted for the four types of processors. DECIMATION The process of decimation includes lowpass filtering and downsampling.1 shows a typical example of observed torque spikes. The amplitude of torques spiked to 20.
On the left side of the figure.1: Changes of Predicted Torque while Increasing Decimation Rate The figure above shows the transformations of the torque predictions as a function of the sampling rate.Subject 07. When the decimation rate is d=1 the torque amplitude exceeds 104 and the coefficients gain exceed 100 dB outside the system band (~ 410 Hz).5 0 1000 2000 3000 4000 5000 CfGain (db) 100 0 d = 20 and VAF = 3431732.755 0 50 100 0 200 400 600 800 Time (samples) 1000 0 5 10 15 Frequency (Hz) 20 Figure 5.5 0 0.5 1 0. Trial 47.5 0 0.7499 0 50 d = 60 and VAF = 87. the gains of the coefficients computed during the system identification (extension) are plotted for each of the decimation rates.8592 100 50 0 20 40 60 80 100 CfGain (db) T and That d = 60 and VAF = 87.755 d = 100 and VAF = 87.6018 2 0 2 4 6 8 10 12 x 10 4 d = 1 and VAF = 134277469. On the right side of the figure. 60. order nb=15 2 T and That x 10 4 Torque %MVCF (Time Domain) 200 CfGain (db) Real Estimate 100 0 EXTENSION Coeffcients Frequency Response 0 d = 1 and VAF = 134277469.7499 100 50 0 5 10 15 20 25 30 35 0 500 1000 1500 CfGain (db) T and That d = 100 and VAF = 87. As the 61 . The purpose of this figure is the graphical presentation of the observed spikes. 20. the torques both predicted (red dashed line) and measured (blue solid line) are plotted versus sample time for four decimation rates (top to bottom: 1. and 100).8592 2000 1 0.6018 100 200 0 500 1000 1500 2000 1000 T and That 0 1000 d = 20 and VAF = 3431732.
2 shows the average torque power accumulation rate for fast tracking bandwidth.5 Frequency (Hz) 3 3.755).5 1 1. Further investigation demonstrated that the errors were related to sampling rate.5 5 Figure 5. and the left side gain plot shows the expected LPF like performance with a max gain of 0 dB.5 2 2.9% of the power within about 4 Hz.sampling rate is increased to d = 100 the model performance improves (%VAF = 87.5 4 4. but as seen in this figure that rate was not enough.2: Signal Power Accumulation (average PSD torque) vs. Figure 5. even at fast tracking bandwidth. a decimation rate of 20 was utilized. During previous research. the spikes disappear. where the spikes were initially observed. The raw EMG sampling rate of 4096 Hz was more 1000 times the natural bandwidth of the EMGtorque system that contained 99. Frequency 62 . Fast Tracking: Average Accumulation PSD of Torque 100 90 % Accumulated of Torque Power 80 70 60 50 40 30 0 0.
If the test trials were contaminated with noise extended to larger frequencies. Hence. which is consistent with the rate recommended by Ljung (1999) and captured all of the signal power. As seen from the %VAF values plot in Figure 5. Although the occurrence was infrequent because most of the torque power was contained within 4 Hz. the decimation rate of 100 is used to obtain the consequent results.96 Hz) extinguished all observed spikes. even a small amount of noise power at frequencies above 4 Hz in a test trial caused a noise spike in the predicted torque.3 the performance of the EMGtorque saturates for rates above 100 and eventually decreases as the decimation rate exceeds 400. the system produces models with unrealistically high gain at frequencies above 4 Hz.1). 1999]. A factor of decimation of 100 (effective sampling rate of 40. The low cutoff frequencies eventually remove signal power within the natural bandwidth. the spikes reduced both in occurrence rate and magnitude. The solution to this problem was to decimate the EMG amplitude signals prior to performing the system identification. Concluding. Progressively lowering the effective sampling rate. This eventual drop in the performance is due to the decrease of the cutoff frequency of the lowpass filter included in the decimation process. This phenomena encountered with oversampling is also described as common in the system identification literature [Ljung. they were an obstacle to further improvements. 63 .The occurrence of the noise spikes was associated with the cases that the system was calibrated with torque trials (training sessions to determine fit coefficients) that contained no power beyond 4 Hz (99% chance). The high gains beyond the system band are shown for a typical example on the right side of the picture (Figure 5.96 Hz is approximately 10 times the highest signal frequency. The optimal sampling rate of 40.
The method of estimation is explained earlier hence this section presents the drawn results.00E+01 100 200 300 400 500 600 700 800 900 VAF2 VAF3 VAF4 4. COMPARISON OF EMG AMPLITUDE PROCESSORS Applying decimation with a factor of 100 and the additional preprocessing steps (truncation and A/D offset subtraction) described. the EMG amplitudes obtained from four different processors were used to estimate torque.00E+01 4.00E+01 VAF1 VAF 0.00E+01 6.00E+01 8.3: Decimation Rate Evaluation Plot 5.00E+01 1.00E+02 8.00E+00 0 2.2.All Vaf(s) 1.00E+02 decimation LEGEND Subject: Train Trial: Test Trial % VAF1 02 25 28 % VAF2 07 45 47 % VAF3 20 65 69 % VAF4 04 60 61 Figure 5. 64 .00E+01 2.00E+01 6.
4. % MAE in eq. were averaged across the 180 combinations for each processor. The plot shows the results obtained for the fast tracking speed using PINV to invert matrices. for each of the four EMG amplitude processors.4: Raw EMG (flexion & extension) and Torques Figure 5.Flexion Torque Amplitude %MVC 0 Negative Torque .4 shows the raw EMG signal and the predicted torque from two different processors in a typical example.5 shows mean and median values of %VAF and MEA. The values obtained for performance evaluating expressions. Unwhitened EMG Predicted Torque.Extension 50 5 10 15 20 25 Time (sec) Figure 5. The predicted torque in both cases captures most of the dynamics exhibited in the actual torque.4. 4. as a function of the system identification model order. whereas the slow tracking speed and the use QRfactorization performance results are shown in APPENDICES: (III) 65 . Figure 5. Whitened EMG Time (sec) Real Torque and Predicted Torque Normalized in %MVC Positive Torque .3 and % VAF in eq.Extension (Triceps M/C) 400 E Flexion (Biceps L/C) 400 Raw EMG in %MVE 200 Raw EMG in %MVE 5 10 15 20 25 F 200 0 0 200 200 400 400 5 10 15 20 25 Time (sec) 50 Real Torque: Measured Predicted Torque.
Median of %VAF. nb System Identification Order. for all EMGtorque processors the plots show a progressive increase in performance as model order is increased up to about 10–15th order. and standard deviation) for these data are shown in the following tables. nb Mean of %VAF. Fast Tracking Speed 12 90 11 Mean of MAE. median. The distribution characteristics (mean. nb System Identification Order.5: Median (left) and Mean (right) of % VAF and % MAE for fast tracking Assuming that better EMGtorque performance is indicated by higher %VAF and lower % MAE. Fast Tracking Speed SCHUNWHIT SCHWHIT MCHUNWHIT MCHWHIT 80 MAE (Percent) 30 10 9 8 7 6 5 4 0 5 10 15 %VAF 70 60 50 0 5 10 15 SCHUNWHIT SCHWHIT MCHUNWHIT MCHWHIT 20 25 20 25 30 System Identification Order. nb Figure 5. Fast Tracking Speed 12 90 11 Median of MAE. The improvement seems to stop passing the 20th order and it is expected to eventually decay as the model order starts fitting into the noise. Fast Tracking Speed SCHUNWHIT SCHWHIT MCHUNWHIT MCHWHIT 80 MAE (Percent) 30 10 9 8 7 6 5 4 0 5 10 15 %VAF 70 60 50 0 5 10 15 SCHUNWHIT SCHWHIT MCHUNWHIT MCHWHIT 20 25 20 25 30 System Identification Order. 66 .
Table 5.1: Distribution Info of % VAF Values for Each Processor (Fast Tracking)
Processor 14 SCHUNWHIT % VAF MEDIAN MEAN STDEV MEDIAN MEAN STDEV MEDIAN MEAN STDEV MEDIAN MEAN STDEV n=1 51.32 49.17 43.03 55.49 54.58 33.73 59.98 56.91 31.42 61.99 61.31 26.27 n=2 57.56 53.03 44.17 62.65 59.89 32.40 65.23 61.98 29.99 69.61 67.92 22.71 n=3 63.86 56.48 45.71 69.05 63.47 32.19 69.99 66.06 29.36 75.72 71.97 21.21 n=4 69.55 59.65 47.40 73.41 66.66 32.13 74.64 69.81 28.93 79.60 75.50 20.02 n=5 73.64 62.60 49.31 77.57 69.33 32.19 78.70 73.15 28.75 82.88 78.33 19.18 n = 10 83.29 68.50 54.54 84.80 74.73 32.66 88.31 80.35 28.62 89.45 84.09 17.84 n = 15 84.70 69.07 56.99 85.51 75.37 33.03 89.33 81.39 29.16 89.98 84.83 17.86 n = 20 84.94 69.07 57.41 85.79 75.52 32.95 89.35 81.47 29.49 90.14 84.97 17.92
MCHUNWHIT
SCHWHIT
MCHWHIT
Table 5.2: Distribution Info of % MAE Values for Each Processor (Fast Tracking)
Processor 14 SCHUNWHIT % MAE MEDIAN MEAN STDEV MEDIAN MEAN STDEV MEDIAN MEAN STDEV MEDIAN MEAN STDEV n=1 10.64 11.21 3.97 10.25 10.51 3.66 9.98 10.15 3.27 9.49 9.53 3.01 n=2 9.90 10.73 3.99 9.41 9.89 3.59 9.40 9.55 3.13 8.62 8.74 2.71 n=3 9.42 10.27 4.09 8.74 10.51 3.63 8.68 9.01 3.09 7.88 8.20 2.63 n=4 8.77 9.82 4.20 8.03 8.99 3.68 8.02 8.49 3.27 7.11 7.69 2.59 n=5 8.16 9.36 4.35 7.56 8.59 3.75 7.36 7.98 3.27 6.56 7.25 2.58 n = 10 6.50 8.25 3.97 6.34 7.61 4.01 5.65 6.63 3.35 5.31 6.12 2.71 n = 15 6.24 8.04 5.03 6.10 7.45 4.08 5.43 6.37 3.47 5.12 5.94 2.76 n = 20 6.20 8.01 5.10 6.08 7.42 4.10 5.40 6.33 3.52 5.02 5.90 2.77
MCHUNWHIT
SCHWHIT
MCHWHIT
As seen, the performance of the processors follows a ranking order with the whitened multiplechannel processor providing the best performance, followed by multiplechannel unwhitened, then singlechannel whitened, then singlechannel unwhitened. It is also noticed that the median results are higher than the mean, suggesting that there are still a few influential large errors that drag the mean downward. Finally the PSD of the error is graphed to gain an insight on the origin of the error. The plots show that 8090% of the error power was within the first 0.5 Hz (shown in the figure below).
67
PSD of Total Error Cumulated vs. Frequency for each Processor Type 100
90
80
% PSD of Error Cumulated
70
60
50
40
30
SCHUNWHIT SCHWHIT MCHUNWHIT MCHWHIT
20
10
0
0.5
1
1.5
2
2.5 3 Frequency in Hz
3.5
4
4.5
5
Figure 5.6: PSD of Error Accumulation Rate
Error Spectral Distribution over Total Error 0.06 SCHNWHIT MCHWHIT 0.05 Average PSD of Error %MVC F
0.04
0.03
0.02
0.01
0 0.5
0
0.5 1 Frequency (Hz)
1.5
2
Figure 5.7: The Average PSD of Error as Estimated from Welsh Periodogram
68
Even though Figure 5.6 suggests that almost all error is accumulated at low frequencies, the windowing effect and the spectral leakage in the Welsh periodogram do not allow further resolution (Figure 5.7). Therefore, it is hard to determine the origin of the remaining error. After a closer study of the trials with the largest error, it was identified that the error was due to DC shift on the estimated torque as compared to measured one. Moreover, some trials matched were almost perfectly, except for the existence of a positive/negative offset (Figure 5.8).
DCPROBLEMS: Subject 16, Trial 64, order nb=15 EXTENSION Coefficients Frequency Response Time Domain Torque in % MVCF 40 Real Estimate 20 CfGain (db) 0 d = 1 and VAF = 532.72 20 20 CfGain (db) 0 500 1000 1500 2000
1 T and That 0 1 2 1 T and That 0 1
0
2
4
6
8
10
12 x 10
4
0 d = 20 and VAF = 752.2972 20 20 d = 60 and VAF = 876.7484 0 20 40 60 80 100
d = 20 and VAF = 752.2972 2 1 0 1 d = 60 and VAF = 876.7484 2 1 d = 100 and VAF = 911.0364 0 1 2 0 500 1000 1500 0 1000 2000 3000 4000 5000
CfGain (db)
T and That
0
20 40 CfGain (db)
0
5
10
15
20
25
30
35
T and That
d = 100 and VAF = 911.0364 20 0 20
0
200
400
600 800 Time (samples)
1000
0
5
10 15 Frequency, Hz
20
Figure 5.8: High DC Offset Error on Estimated Torque
69
20. The next largest error in torque estimation seems to be due to a DC offset. Because most of the error power is contained at such low frequencies and the observation of DC shift suggested the same. In conclusion. 70 . The results for the slow tracking speed data were similar and are shown in the appendices. it can be hypothesized that there is a large amount of error remaining at approximately DC. The typical model performance was shown previously in Figure 5. The performance of the system indicated by % VAF is getting negatively larger and the DC shift remains as the only apparent error as the AC portion of the estimated torque starts looking exactly like the measured one.4. the optimum EMGtorque model (15th order) produced an average error of 6% MVCF with a %VAF of 90% when the EMG amplitude estimates were obtained from a multiple channel whitened processor. 60.The figure shows the changes of the estimated torque as the decimation rate is increased (1. More investigation of the error plots is required to ensure the hypothesis is correct. 100).
The results presented are achieved considering several assumptions on physiological characteristics of the EMGtorque system. The results derived were consistent with the expectations showing that both whitening and multiplechannel combination of the EMG lead to reduced EMGtorque prediction errors. Even lower errors are obtained when the two techniques are used in combination..1. 6.1. the discussion and conclusions. D. The first part contains the interpretation of the previously anticipated results and then in the second part conclusions are derived from them. DISCUSSION OF RESULTS The objective of this research was to demonstrate that incorporation of the recent advances in EMG amplitude processors into EMGtorque estimation model produce lower torque prediction errors. Another important result was identifying and resolving the problem of oversampling.1. O. various analyses were conducted to demonstrate that improvement in EMG amplitude processing reduces the estimation error in the torque prediction models. 6. In addition. Advances to EMGtorque Estimation Keeping in mind the main objective. Bida.A. iii 71 . This problem arises when EMG data (typical bandwidth from 20–500 Hz) and torque data (typical bandwidth ≈ 510 Hz) are simultaneously sampled at the highest required rate.CHAPTER 6. this chapter includes the study limitations and the suggestions for future work. DISCUSSION AND CONCLUSIONS iii This last chapter contains two main sections. to preserve information in the Parts of this chapter are taken from the paper submitted to the Journal of Biomechanics (attached to Appendices) authored by Clancy. Thus. E. Rancourt.
experimental design consisted of constantposture nonfatiguing contractions about the elbow. Increasing the number of muscles in the model would not necessarily improve the EMG amplitude estimates and would increase the complexity of system identification. the combination of multiple channels improves the EMG amplitude estimation. In addition.EMG data. as recommended by Ljung (1999). such as extensor (flexor likewise) torque was a result of extension (flexion likewise) only. 6. Another important assumption to consider is that each muscle group contributed only to a certain torque component. the Several assumptions were necessary to complete this project. there are only two inputs to the model (EMGamp flexion and EMGamp extension) as opposed to several EMG amplitudes. This issue was observed and solved by decimating the data to a rate of 40. Consequently. Otherwise. the resultant transfer functions produce unrealistic gains at frequencies above the typical band of torque data. representative of each individual muscle recording. Most practical contractions are more fully dynamic (posture varying). 72 .1. If the so overasmpled torque data are used to calibrate EMGtorque models. with only one agonist and one antagonist muscle group. the torque data are oversampled. Study Limitations and Future Suggestions First. the mechanical model for the elbow treated the joint as a simple hinge. This assumption enables the possibility of obtaining one single EMG amplitude estimate using electrodes that are placed anywhere on the skin overlying that muscle group.2. the crossterms of EMG amplitude had to be incorporated in the model requiring corresponding torque components for mechanical validation.96 Hz that correspond to approximately ten times the highest torque frequency. Second.
pp. 1998]. In the case of abrupt changes in muscle activation from 0%MVC to 50% MVC. corresponding to a standard parametric model in the literature (FIR type of ARX model). Although the EMGtorque model for this study included both “AC and DC characteristics” of the system. Another advantage is that nonlinear models can potentially capture additional subtle behavior in an EMGtorque relationship such as the electromechanical delay between action potential activation and muscle fiber contraction. it is highly recommended in the literature to separate them as part of the data preprocessing routine [Ljung 1999. the EMD is a dynamic parameter. The investigation of the trials displaying large errors demonstrated that the operating point (DC bias) contaminated in the inputs (EMGamp) and output (torque) data influenced the dynamics of the system. and Harron. its inclusion in the model can be potentially neglected [Vint. during slow varying force tasks (25% MVC to 50% MVC) the electromechanical delay is constant (60 ms). The reason for selecting a standard linear model was the level of simplicity in the solution and the information available on the system identification techniques. Even though many system identification paradigms remove the DC component of each signal prior to identification 73 . therefore. resulting in a DC shift in the estimated torque. hence it is dependent on the number of motor units activated and the fatiguing effects. 2001]. McLean. 458460].The EMGtorque model was identified using a zerosonly system design. However. Electromechanical delay (EMD) is defined as the temporal delay (26131 ms) that exists between the onset of muscular activity and the generation of force [Strojnik and Komi. Nonlinear models can be more accurate. for example including hysteresis to capture any systematic differences in the EMGtorque relation between concentric and eccentric contractions.
Although tracking targets were at low frequency.2. The skin was properly prepared to reduce the impedance prior to placing the electrodes. the distances between the electrodes. Advances in EMG amplitude estimation were applied to the EMGtorque problem for constantposture. Therefore. 6. active electrodes rejected the cable motion artefact. and their placement on the skin. Further studies are suggested.so that only the system dynamics (AC portion) are identified. precautions were taken when recording the data in order to avoid hardware problems such as electrode failures. the tracking target will need to produce a deterministic trajectory so that ballistic movements (which are inherently faster) can be utilized. SUMMARY AND CONCLUSIONS This research focused on demonstrating that advances in EMG amplitude processors result in EMGtorque model performance improvements. Finally. 1 Hz is about as fast as subjects were able to track. Faster speeds would require ballistic force trajectories. since a complete system response requires separate torque data for model validation and should not have influenced the relative comparison of EMG amplitude processors The selection of contraction bandwidth that sufficiently excites the system is one of the main issues that need to be addressed in the future. 74 . digital high pass filters were utilized to eliminate noise due to physiological motion artefact. a complete output response (torque) is found by adding the estimated AC portion to a separate estimate of the output DC value. nonfatiguing. and the lowpass filters along with decimation attenuated other interferences such as UV lights and power harmonics. the contraction bandwidth was limited to a 1 Hz tracking target. If higher frequencies are required in the future.
EMGtorque errors with a fourchannel. 75 . Certainly. By comparison. whitened processor produced an average error of 6% MVCF (%VAF of 90%) at the fast tracking speed. the study was limited in several manners in order to be able isolate the effect of EMG amplitude without the complexity of less restrictive EMGtorque models. As such. Concluding. the primary interest was the influence of different EMG amplitude processors on EMGtorque prediction performance. using an 8th order. Butterworth filter. Using 15thorder and higher linear FIR models. Results from 15 subjects showed that EMG whitening and multiplechannel combination both reduce EMGtorque errors and their combination provides an additive benefit. The EMG amplitude sampling rate was reduced to 40. The dynamic relationship between EMG amplitude and joint torque was formulated as a standard linear least squares problem. Accomplishing the objective. the singlechannel. the issue of nonconvergent trials was isolated and resolved by decimating and low pass filtering the data prior to system identification.96 Hz and both EMG amplitude and torque were lowpass filtered at the Nyquist rate. The expectation is that the benefits shown here of improved EMG amplitude processors would transfer to many other EMGtorque modeling problems. unwhitened (conventional) processor produced an average error of 8% MVCF (%VAF of 68%). zerophase. it now seems justified to progressively release these restrictions and validate these benefits in more general applications in future work.9% of the system power. The power spectral density analysis showed that the chosen cutoff frequency of the filter preserved 99.forcevarying contractions about the elbow.
J. W. Basmajian and C. 6. Accessed January 30. Laval Univ. 1993] T. “Relation dynamique entre les signaux electromyogrphiques et le couple au coude lors de contarctions a angles constants.C. 1990. Z.G. 76 . [Agarwal and Gottlieb. 2004] E.Y. 1990] “LeastSquares Identification of the dynamic relation electromyogram and joint moment. MA 1991.166–170... Eng. 20.04. Mag. Cambridge. [Bouchard. May 1993. 1954. “Estimation of muscle forces about the wrist joint during isometric tasks using an EMG coefficient method. pp. “Electromyogram amplitude estimation with adaptive smoothing window length. Goldstein. Quebec. 2001] S. J. 26. [Bendat and Piersol.” IEEE Trans. Random Data: Analysis and Measurement Procedures. J. Biomed.edu/~ted/emg_tool. pp.” Amer. Industrial Hyg. 2001.REFERENCES [Armstrong et al. John Wiley.A.L. (Lond)..Dec.J. P. vol. pp. 1991] E...” J Biomech Eng. Armstrong. O. Clancy. 1986. EMG amplitude estimation toolbox: User’s guide. vol. 12756. “Stochastic modeling of the relationship between the surface electromyogram and muscle torque.. 1999] E. . Rymer. B. vol. Rancourt. Physiol. D. vol. S.wpi. Gottlieb. Bouchard. S. 1985] J. Eng. S. vol.A.. De Luca. Clancy. “Estimation and application of EMG amplitude during dynamic dontractions. N. 2001] E. Clancy. Foulke. 2. Bio. A. Bouchard.J Lippold. Nov. [Buchanan et al. [Bobet and Norman. no. [Clancy.“Compliance of the human ankle joint. 2005. 1985.S. velocity and integrated electrical activity in human muscles.A. Dept. vol.” J. no. no. [Basmajian and DeLuca.S. Dewald. “Investigation of cumulative trauma disorders in a poultry processing plant. 46.” J. [Clancy. Joseph. Canada 2001.. 1986] J. no. Biomech. Muscles Alive: Their Functions Revealed by Electromyography. Moniz. 1982. M.A. A. Piersol. 1954] B. A. Buchanan. 1977] G. 123. 23. Med. 214224. Bendat.” J.. Williams & Wilkins Company. Biomech. Assoc. Bigland. J.V. Mech. 547560.htm.. Sci. J. 1977. no. 45. Eng. 12. Dept. 99. pp. [Clancy.” M.C. Alpha version 0.” IEEE Eng. vol. G.6. [Clancy et al. 1999. 43. S. Available online at: http://ece. pp. Comp.” PhD thesis. Clancy. [Bigland and Lippold. “The relation between force. 1982] T. thesis. Elect.. 103116. A. Agarwal. MIT.
24. 44.A Clancy. 1997] J. Electromyog. Clancy. G. 1997. Roberto Merletti and Philip A.” J. G. Filligoi.. and Merletti 2002] E. Farry. vol. Hannerz. Merletti.” IEEE Trans. Eng. New Jersey. [DeLuca. vol. Morin. [Gottlieb et al. 77 . Biomed.. Agarwal. Song. [Clancy and Hogan. 1997] G. 1997 [Hannerz. 30. pp. vol. Hong. vol. no. vol. De Luca. “Relating agonistantagonist electromyograms to joint torque during isometric. “Dynamic relationship between isometric muscle tension and the electromyogram in man. [Clancy and Farry. Gottlieb and G. Scand. pp. [Clancy. June 2000. E. Habes.L. June 2000. 1971. Clancy. Copyrighted 2002 [Dowling. Enoka. 374– 384. Merletti. 203–211..A. PutzAnderson. Farina. 8. Parker.. R. A. John Wiley & Sons. D.” IEEE Trans. 8296. Neurophys.. “Adaptive whitening of the electromyogram to improve amplitude estimation. vol.A. Biomed. “The use of electromyography for the noninvasive prediction of muscle forces. Hoboken. vol. nonfatiguing contractions. 1971] G. 1995. and Enoka. 1995] E. 2004. no.[Clancy et al. Aug. noisereduction and amplitude estimation issues in surface electromyography. “The extraction of neural strategies from the surface EMG. “Multiple site electromyograph amplitude estimation. Inc. 78. 12. 1974. [Farina...A. pp. quasiisotonic.L.” Acta Physiol. Kines. 2. Merletti. Farina. 6. 47. Grant. 3139.M. J.. and Noninvasive Applications.. “Directional control of planar human arm movement. 2004. Clancy and N. “Sampling. 2004] E. 345 – 351. pp 1–16. 96 pp. D. Eng.A. 2002] C. “Surface electromyography: detection and recording” DelSys Inc. Hogan. 2004] D. J. “Discharge properties of motor units in relation to recruitment order in voluntary contraction.J. Corcos. D. no. vol. vol. 29852998. 1997] E. pp.L.Hogan.. 1974] J. Appl. Industrial Ergonomics. pp.” in Electromyography: Physiology. eds. N. Gottlieb. R. V. Clancy and K. 133−168. Eng. Q.C.A. Almeida. December 2002.” Sports Med. Biomed. 1994] K. “Single Channel Techniques for Information Extraction from the Surface EMG. [Clancy and Hogan. vol. 1994.” J. 14861495.” Int. D.” J. 42. Physiol.J Dowling. Physiol. R. pp. Engineering. [Gottlieb and Agarwal. Morin. [Grant et al. pp... 13.” IEEE Trans. “Psychophysical and EMG correlates of force exertion in manual work.” J.. 2000] E.L. Appl. 91.
” IEEE Trans. “Myoelectric signal versus force relationship in different human muscles. Biomechanics vol. Laursen. vol. 639642. Kearney.” Exp. Eng. no. (CH 2). no. [Laursen et al.. November 1994.F. vol.B. [Kirsch. vol.327339. [Hof and Van Den Berg. 1992] R. B. Wright. Int. March 2002.. Jw. Hill. 747–758. U. vol. [Hughes and Chaffin. Hogan and R. Biomed Eng. “EMG to force processing I: An electrical analogue of the Hill muscle model. [Lawrence and DeLuca. R.” IEEE Trans. Biomech..W. D. J. vol.L. July 1980. 1998] B. BME27. 1983. [Hogan and Mann.[Hasan and Enoka. [Inman et al. 7th ed. 1998. 2001. 1983] J. R. H. G. pp. “Relation of human electromyogram to muscular tension. pp.W. Feinstein. Kirsch. and Crago. P. Hasan. 1952. Jensen. 54. Kearney. New Jersey. Hogan and R. 1980b] N. 59. Kolman.H. [Hogan and Mann. 187–194. vol. Mann. 396410.. Department of Health and Human Service (CDC) – NIOSH. Van den Berg.. Proc. Brain Res. M. Conf. Neurophys. “Myoelectric signal processing: optimal estimation applied to electromyographyPart I: Derivation of the optimal myoprocessor. Prentice Hall. “Dynamic stiffness measurements and control movement. 1980a] N. 382395. Chaffin. 1997] R. Crago. 428429. Sjogaard. Hobart. 4.” J.” J.” IEEE Trans. Hof. 16531659. 44. B. Inman. “Isometric torque angle relationship and movement related activity of human elbow flexors: Implications for the equilibriumpoint hypothesis.” IEEE 16th Ann. pp.E. 1997. Hughes. Mann. 14. Nemeth. D..” J.M Saunders. G. [Lamb and Hobart. pp. BME27. “A model predicting individual shoulder muscle forces based on relationship between electromyographic and 3D external forces in static position. pp. and Clin. Ralston.” Electroencephol. pp.. 1985. [Kolman and Hill.Lawrence. vol. “Introductory Linear Algebra with Applications”. pp. “Selected topics in surface electromyography for use in the occupational settings: exepert perspectives”. “Using principalcomponents regression to stabilize EMGmuscle force parameter estimates of torso muscles.J De Luca. Appl. 1994] R. 78 . B. pp.W. D. July 1980. 11. Enoka.T.C. “Myoelectric signal processing: optimal estimation applied to electromyographyPart I: Derivation of the optimal myoprocessor. on Biomed. 1952] V. vol. 441450.J.1. 31. Biomed Eng. E. Lamb. pp. 2001] B. C. 7..E. 731739.. 1981.S. 1985] Z. R. E. R. Physiol. 1981] A.
[MilnerBrown and Stein. Solomonow.. MilnerBrown. 30.” J. vol. H. and motor units control strategy. 25. B. [Nussbaum. and muscular force. 1999] R. S. in Biomed. Eng. and Stochastic Processes. System Identification: Theory for the User. R. J. vol.r Clin. Med.. 1995. NJ: PrenticeHall. 246. 1975] H.” Amer. pp. “Relationship between the surface e. D’Ambrosia. McGill. 6. 57. J. [Merletti. Papoulis. Stein. gravity and recruitment. Guzzi.1992.. B. no. 1999] L. Random Variables. pp 653662. “A backpropagation neural network model of lumbar muscle recruitment during moderate static exertions.C. Electromyography Kinesiology. Shoji. Biomech.[Lindstrom and Magnusson. 4th edition. 122. Merletti. R. [Perry and Bekey. vol. 1977] J.m. 1983] G. J. 1990] M. “The relation between the surface electromyogram and muscular force. no. vol. 65. Ray. T. Baratta. Solomonow. G. Ljung. Apr. 223–244. S. 1981] J. “The EMGforce relationships of skeletal muscle.” Med. Comput. B. no. 28.M. Magnusson. Upper Saddle River. Neurophysiol. pp. H. Chaffin. 79 . [McGill.” Proceedings of the IEEE. 5. [Solmonow et al. “Interpretation of myoelectric power spectra: A model and its application. 141152.” Electromyog. 2002] A. and Martin. 1977. R. pp. 549569. A.” CRC Critical Rev.” J. pp. R. 3. 1986] M. 1990. no. vol. 1999.. R. 263277.K. Sept. “EMGforce model of the elbow’s antagonistic muscle pair: The effect of joint position. D’Ambrosia. Lindstrom. Nussbaum.” Am. pp.” J. [Ljung. 395414. Biol. Shoji. “Standards for Reporting EMG Data. dependence on contraction rate. “EMGforce relationships in skeletal muscle. McGraw Hill. Perry. Physic. 10151024. 65. 9. Eng. A. Chaffin. Physiol. Biomech. Probability. Med. 1983 [Solmonow et al. [Papoulis. “A myoelectrically based dynamic threedimensional model to predict loads on lumbar spine tissues during lateral bending. Martin. DeVries. vol... 1978] T. 2002.. Baratta. A Bekey. [Ray and Guha. 4. Phys.” J. 1975... 21. 1986. Moritani. R. Guha. pp. pp.g. 579586. no. D.. 1999. 1992] S. December 1981. H. “Reexamination of the relationship between the surface integrated electromyogram (IEMG) and force of isometric contraction. vol. pp. vol. [Moritani and DeVries. 1995] M. 1978.
62. 36.J. Methods for the Identification of MultipleInput Nonlinear Systems.B. Issue 1. 80 .. 85761. Vint.. McGill University.” Clinic. 84. AshtonMiller. 344350. M. Vredenbregt. Appl. 1998. http://www. Woods. G. M. V.D.. 6. 1994] D. pp. 2003] F. Strojnik. pp. vol. Schultz. 1973. in Sports & Exercise. and Haugh. J. “Neuromuscular fatigue after maximal stretchshortening cycle exercises. S. 33. Haugh. [Stokes. D. 1995] D.” Journal of Biomechanics. L. [Wyss and Pollak. 1. A.A. Stokes.J ValeroCuevas. Phys.. “Electromechanical delay in isometric actions initiated from nonresting levels. V. McLean. vol. Komi..F. [Strojnik and Komi. 2003.[Sommerich et al. [Thelen. S. 607622. Towles. J. “Adjustment of muscle mechanics model parameters to simulate dynamic contractions in older adults. 1.” J. [Westwick. B. Moffroid. 1989. pp. S. 13.. 13. [Thelen et al. McLean. Eng.” Eng. vol.. vol. vol.. [Vint. Dept. 2001] P. 2733.D. G. Physiol. EMG Clin. Sommerich. no.. pp. 1998] V.A. F. Westwick.T.enel.” Am.ucalgary. Pollak “Surface electromyogram/muscle force: a muscle model based on RMG peaks. [ValeroCuevas et al. Vol. no. Med. Johanson. issue 6 pp. vol. 1989] I. BiglandRitchie. Biomech. vol. “Linear and nonlinear surface EMG/force relationships in human muscles: An anatomical/functional argument for the existence of both.M. 2001. 1998] C. Rau. Quebec. Biomech. 1983] J. 2003. Wyss. M. Eng. pp. 978983. P. vol. pp. “Towards a realistic biomechanical model of the thumb: the choice of kinematic description may be more critical than the solution method or the variability/uncertainty of musculoskeletal parameters. J. Fassois.P.G. “Identification of dynamic myoelectric signaltoforce models during isometric lumbar muscle contractions. 287299.G. 1983. 1984] J. Harron. 10191030.” J. 2003] D. Med. 27. Canada.” Spine. 1994. 261271. pp. 907–919. “EMG to torque relationship in rectus abdominis muscle. S. 14. Montreal. Rush. 1973] J. “A method for developing biomechanical profiles of handintensive tasks.ca/People/Westwick/Publications/ [Woods and BiglandRitchie. pp.P. Thelen.E. 1984. and Harron. Neurophys. & Sci.” New Develop. W. pp. vol. “Surface EMG in relation to force.” J Biomech.. muscle length and endurance. Rush. Parnianpour. EE Biomed. Marras. [Vredenbregt and Rau. Moffroid. A. 7077. 125. no. 8. 1995.” Med. M. Thelen. 1998.
D. November 1969.[Zuniga and Simons. Med..G. “Nonlinear relationship between averaged electromyogram potential and muscle tension in normal subjects. 81 . Phys. and Rehab. Zuniga. 1969] E. Simons.N. 613620.” Arch.
Multiple Channel Unwhitened EMG.1: Trial ID Name Codes TRIAL ID (TT) 20 (train) 21 (test) 22 (test) 23 (test) 24 (test) 25 (train) 26 (test) 27 (test) 28 (test) 29 (test) 40 (train) 41 (test) 42 (test) 43 (test) 44 (test) 45 (train) 46 (test) 47 (test) 48 (test) 49 (test) 60 (train) 61 (test) 62 (test) Feedback Mechanism EMG.25 0. Single Channel Whitened EMG. 18. Table 0.25 0. Single Channel Whitened Target Bandwidth (Hz) 0.25 0. Multiple Channel Whitened Dynamometer EMG. Single Channel Whitened EMG.25 1 1 1 1 1 0.25 0. Single Channel Unwhitened EMG. Single Channel Whitened EMG. AND FIGURES I. 6069). PLOTS. 16. Multiple Channel Unwhitened EMG. 0710.25 0. LBXXXX EXPERIMENT DATA FILE DESCRIPTION The file name code is LBSSTT where SS stands identifies the subject (0205. Multiple Channel Whitened Dynamometer EMG.25 1 1 1 1 1 0. Multiple Channel Unwhitened EMG. 4049.25 0. Single Channel Unwhitened EMG. 20. Multiple Channel Whitened Dynamometer EMG. The different trials within a set were obtained using different feedback mechanisms and bandwidths for the tracking signal. Single Channel Unwhitened EMG.25 0. Single Channel Whitened EMG.APPENDICES: ADDITIONAL INFORMATION. There are 15 subjects and 3 sets of 10 trials per subject.25 0. 13. Single Channel Unwhitened EMG. Multiple Channel Whitened Dynamometer EMG. Multiple Channel Unwhitened EMG. Single Channel Unwhitened EMG. One out of four trials was used to train the EMGtorque model the other four to test it.25 0. 21) and TT identifies the experiment trial (2029. 12.25 0. 17. Multiple Channel Unwhitened EMG.25 82 .
63 (test) 64 (test) 65 (train) 66 (test) 67 (test) 68 (test) 69 (test) EMG. most lateral Dynamometer 83 .2: A/D Channel Name Codes A/D Channel 1 2 3 4 8 9 10 11 16 Contents EMG: Biceps. lateral center EMG: Biceps. Single Channel Unwhitened EMG. lateral center EMG: Triceps. most medial location EMG: Triceps. Table 0.25 0. most medial location EMG: Triceps. medial center EMG: Triceps. most lateral location EMG: Biceps. The table lists only the channel used for this study.25 1 1 1 1 1 Each of the files contained 16 channels from the DAQ describing the electrode positions as follows. Multiple Channel Unwhitened EMG. Multiple Channel Whitened Dynamometer EMG. medial center EMG: Biceps. Multiple Channel Whitened Dynamometer 0. Single Channel Whitened EMG.
SampFreq [. 84 . sCal. . EMGinfo) EMGinfo – obtained from calibration EMGin – input EMG channel (CH 2 for flexion and CH 9 for extension) or Multiple Channels (CH 1:4 for flexion and CH 8:11 for extension) Calibration for EMG amplitude estimation as in EMG toolbox User Manual. NoiseMat. 'PropertyName'. OPTIONAL PROPERTIES FOR AMPLITUDE ESTIMATION ALGORITHM EMG amplitude estimation as taken from the EMG toolbox User Manual. HpassOrder Order: Default value is set by e_h_pass() and Order = 5. Properties as set of this project: sCal: 0. This signal is specifically recorded for calibration process. rest Trial 15 (take corresponsing channel) For example: NoiseMat for flexion is trial 15 (CH 2 for singlechannel and CH 1:4 for multiplechannel) and for extension is still trial 15 but different channels. Syntax : EMGamp = e_amp(EMGin..Noisy signal Matrix (Input Channel). Whitening Filter Settings WhiteFlag 'Flag': Depends. HpassFlag 'Flag': Default value is 'Filter'.. PropertyValue. Trial 10 is used for flexion and trial 12 for extension.II.5 SampFreq: 4096 Hz High Pass Filter Settings Causality 'Flag': Default value is 'Noncausal'. HpassWn: Default value is set to correspond to 15 Hertz.] SandNmat . Syntax: EMGinfo = e_cal(SandNmat. The channels are the same as the input channels shown for EMGin. NoiseMat – Noise recorded per trial. SCHWHIT and MCHWHIT is ‘ON’ else ‘OFF’.
WhiteSmFilt = 'MAV': WhiteSmFixWin Window = 1024 = 250ms.edu/~ted/emg0_04/front. White_sSpec sSpec: Default is set by e_cal_wh(). SmoothWn: 0 < Wn < 1= > Wn = 20/2048 which is 20Hz SmoothOrder: Order = 8 For example EMGinfo for Extension in the case of Single Channel Processor EMGinfoExt = e_cal(SandNmatE. WhiteMaxGain MMaxGain: Default is set by e_cal_wh(). WhiteOrder Order: Default is set by e_cal_wh().wpi.WhiteEdges EdgeString: Default is set by e_whiten(). Multiple Channel Combination Setting UncorrFlag 'Flag': Default is 'GainOnly'. Demodulation Settings DemodFlag 'Flag': Default value is 'On'. NmatE. WhiteSum SumString: Default is set by e_whiten(). 0. WhiteSmFilt FiltOption: Not Used. 'WhiteFlag'. 'WhiteSmFixWin'. For more information on the settings and the EMG toolbox functionalities refer to the User Manual created and maintained by Clancy (2004) found in the website: http://ece. WhiteNfft NNfft: Default is set by e_cal_wh(). 4096. The code looks exactly the same for all the other cases. 'MAV'.html 85 .5. 'WhiteSmFilt'. DemodM M: Default d = 1 (MAV) Smoothing Filter Setting SmoothFilt FiltOption: ‘butter’ SmoothEdges EdgeString: Default value is set by e_smooth(). if multiple channels. Wflag. 1024).
86 .02 0 0 0 0.04 0.5 1 1.5 1 Frequency 1.02 0.1 is used to support the argument that the estimation error for all processors behaves similarly in frequency domain.06 0.04 MCHUNWHIT SCHUNWHIT 0.7 shows only two of the processors (schwhit. Figure 5.06 0.5 SCHWHIT 0.5 2 Frequency MCHWHIT 0.01 0 0 1 1. mchwhit).5 Frequency 2 Figure 0.III. EXTRA FIGURES AND PLOTS Average of Error PSD (%MVCF) 0.04 0.05 0.02 0.04 0.1: Estimation Error PSD (Welsh Periodogram) for all 4 Processors The power density spectra (PSD) of estimation error are approximated using Welsh periodogram for all four processors.5 2 0 0 0.03 0. Figure 0.02 0.5 Frequency 2 0 0.08 0.5 1 1.06 0.
2: System Performance (% VAF & MAE) using QR Factorization (Fast Tracking) The system performance versus system ID order using QR factorization to compute the inverse of the matrices is given to support the argument that the pseudoinverse and the QR factorization give equal results.04 0 20 SI order. nb 60 0. nb 40 60 Figure 0.12 SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT Median of MAE. 87 . Figure 0. fast tracking speed 100 90 0.06 0. fast tracking speed 0.5 and the results seem to match perfectly.1 MAE 80 %VAF 70 60 50 40 0 20 SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 40 SI order.08 0.08 0. fast tracking speed 100 90 0.06 0.QR Method Median of %VAF. nb Mean of MAE. fast tracking speed 60 Mean of %VAF.2 is contrasted with Figure 5. nb 60 0.12 SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 0.04 0 20 40 SI order.1 %VAF 80 MAE 70 60 50 40 0 20 SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 40 SI order.
06 SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 0 20 SI order. 88 . Regardless. nb Mean of %VAF. nb 40 60 Mean of MAE.05 0 20 40 SI order.Median of %VAF.04 0 20 SI order. nb 40 60 0. slow tracking speed 100 90 80 70 60 50 SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT MAE 0.5 and the processor performance ranking order seems to follow the same pattern. slow tracking speed 60 0.09 MAE SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 100 90 80 %VAF 70 60 50 0.11 0.08 0. The model gives better results for slow target tracking tasks which is consistent with the statement that the subjects have trouble tracking ballistic random trajectories. slow tracking speed 0.06 0 20 40 SI order.3 is contrasted with Figure 5.07 0.08 0. slow tracking speed 0. the positive influence of EMG amplitude processor on the EMGtorque model performance is proven for both cases (fast and slow bandwidth signal target).12 Median of MAE.3: System Performance (% VAF & MAE) using PseudoInverse (Slow Tracking) Figure 0.1 0. nb 60 Figure 0.1 %VAF SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 0.
04 0 20 40 SI order. fast tracking speed 60 100 90 80 %VAF 70 60 50 SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 0 20 40 SI order.12 Median of MAE. fast tracking speed 60 0. 89 .AC Model ONLY Median of %VAF. This is another observation toward solving the remaining problems and leading again to the assumption that the inclusion of DC in the models is erroneous.3 and Figure 5.1 MAE 0.08 0. nb Mean of %VAF.06 0.1 %VAF SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 0 20 40 SI order. The model gives even better results when only the AC portion of EMG recordings and measured torque are used in the system. fast tracking speed SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 0. fast tracking speed 100 90 80 70 60 50 MAE 0. nb Mean of MAE. nb 60 0.5.4 is contrasted with Figure 0.06 0. nb 60 Figure 0.08 0.12 SCHUNWHIT MCHUNWHIT SCHWHIT MCHWHIT 0. The processor performance ranking order seems to follow yet the same pattern.04 0 20 40 SI order.4: System Performance (% VAF & MAE) using AC part of EMG Amplitudes (Fast Tracking + PINV) Figure 0.
Notice that as the order of the system passes the value ten little or no improvement is seen within 5Hz (bandwidth of the system). Higher orders seem to add shape to frequencies outside the system bandwidth. 90 . This observation is consistent with the system saturation after 15th order seen in many of the plots shown above. Response Order = 2 Order = 30 Order = 25 Order = 20 Order = 15 Order = 10 Order = 5 5 10 15 20 5 10 15 20 5 10 15 20 5 10 15 20 5 10 15 20 5 10 15 20 5 10 Frequency (hz) 15 20 Figure 0.5: Coefficients Frequency Response for a typical EMGTorque model (Slow Tracking) The most important factor observed in this figure is the shape change as the number of zeros (model order) is increased. Response 3 2 1 0 5 10 15 20 30 2 1 0 5 10 15 20 30 2 1 0 5 10 15 20 30 2 1 0 5 10 15 20 30 2 1 0 5 10 15 20 30 2 1 0 5 10 15 20 30 2 1 0 5 10 15 20 0 Frequency (hz) Order = 2 Order = 30 Order = 25 Order = 20 Order = 15 Order = 10 Order = 5 Flexion Coefficient Freq.3 2 1 0 30 2 1 0 30 2 1 0 30 2 1 0 30 2 1 0 30 2 1 0 30 2 1 0 0 Subject 03 Trial 40 Test 42 Extension Coefficient Freq.
Response Flexion Coefficient Freq.3 2 1 0 30 2 1 0 30 2 1 0 30 2 1 0 30 2 1 0 30 2 1 0 30 2 1 0 0 Subject 03 Trial 45 Test 47 Extension Coefficient Freq.5 for fast target tracking tasks. In this plot it is also observed that during low order the fast later zeros do not leave time for the previous ones to create the dip effect. 91 . This is observed on the highpass look alike response for second and fifth orders. Response 3 2 1 0 5 10 15 20 5 10 15 30 2 1 0 5 10 15 20 5 10 15 30 2 1 0 5 10 15 20 5 10 15 30 2 1 0 5 10 15 20 5 10 15 30 2 1 0 5 10 15 20 5 10 15 30 2 1 0 5 10 15 20 5 10 15 30 2 1 0 5 10 15 20 0 5 10 15 Frequency (hz) Frequency (hz) Order = 2 Order = 30 Order = 25 Order = 20 Order = 15 Order = 10 Order = 5 Order = 2 20 Order = 30 Order = 25 Order = 20 Order = 15 Order = 10 Order = 5 20 20 20 20 20 20 Figure 0.6: Coefficients Frequency Response for a typical EMGTorque model (Fast Tracking) This plot is equivalent with Figure 0. Similarly. as order increases more shape is added to the response.
MA. Department of Biomedical Engineering.edu 92 . Worcester Polytechnic Institute. Clancy (Corresponding Author) Department of Electrical and Computer Engineering Worcester Polytechnic Institute 100 Institute Road. Rancourtc a Department of Electrical and Computer Engineering.. O. 01609. USA c Department of Mechanical Engineering. Worcester. University of Sherbrooke. (508) 8315778. MA. MA 01609 USA Tel. PAPER SUBMITTED TO THE JOURNAL OF BIOMECHANICS Influence of advanced electromyogram (EMG) amplitude processors on EMGtotorque estimation during constantposture. 100 Institute Rd. (508) 8315491.IV. Email: ted@wpi.*. Bidab. Worcester Polytechnic Institute. forcevarying contractions by E. Worcester. Sherbrooke. USA b Department of Electrical and Computer Engineering.A. Fax. PQ. Canada Edward (Ted) A. Worcester. Clancy a. D.
EMGtorque model.96 Hz. such as rectification followed by low pass filtering. Keywords: EMG.Abstract Numerous studies have investigated the relationship between surface EMG and torque exerted about a joint. nonfatiguing. EMG amplitude. Using a 15thorder linear FIR model. we compared the performance of EMGamptorque estimators with and without these advanced EMGamp processors. In this study. Torque. Fifteen subjects produced constantposture. forcevarying contractions about the elbow while torque and biceps/triceps EMG were recorded. EMGtorque errors with a fourchannel. These studies have used conventional EMG amplitude (EMGamp) processing. unwhitened (conventional) processor produced an average error of 8% of maximum voluntary contraction (variance accounted for of 68%). This problem occurs when the torque data are sampled at the same rate as the EMG data. the study describes the occurrence of spurious peaks in estimated torque when the torque model is created from data with a sampling rate well above the bandwidth of the torque. Recently. the equivalent singlechannel. By comparison. EMGamp was related to torque using a linear FIR model. advanced EMGamp processors that incorporate signal whitening and multiplechannel combination have been shown to significantly improve EMGamp processing. The problem is corrected by decimating the EMGamp prior to relating it to joint torque. in our case to an effective sampling rate of 40. whitened processor averaged 6% of maximum voluntary contraction (or 90% of variance accounted for). In addition. to preprocess the EMG before relating it to torque. Optimal sampling rate 93 . Both whitening and multiplechannel combination reduced EMGtorque errors and their combination provided an additive benefit.
This action might not be possible to undo. Are you sure you want to continue?
Use one of your book credits to continue reading from where you left off, or restart the preview.