You are on page 1of 7

Quantitative Relationship Between Feature Extraction

of sEMG and Upper Limb Elbow Joint Angle


Triwiyanto1,3, Oyas Wahyunggoro1, Hanung Adi Nugroho1, Herianto2
1
Department of Electrical Engineering & Information Technology, Universitas Gadjah Mada, Yogyakarta, Indonesia;
2
Department of Mechanical and Industrial Engineering, Universitas Gadjah Mada, Yogyakarta, Indonesia;
3
Department of Electromedical Engineering, Health Polytechnic of Surabaya, Ministry of Health, Indonesia
Email: triwiyanto123@gmail.com

Feature extraction is an important part in the classifier (VAR), waveform length (WL), zero crossing (ZC), and sign
systems. In this study, feature extraction was used to extract the slope change (SSC) are commonly used by researchers to
information of the surface electromyography (sEMG) and to extract the feature of sEMG.
predict upper limb elbow joint angle. To predict the upper limb
elbow joint angle, we explored the EMG signal characteristics on MAV is one of the feature extraction that is often used for
biceps, triceps lateral head and triceps long head. Time domain signal EMG extraction. Kiguchi [1], Andreasen [2] and
of feature extraction is still the best feature extraction to get the Loconsole [3] used MAV as input classifier in exoskeleton
information on signal in a real time processing. Feature development. RMS is often used as feature extraction for
extraction, RMS, MAV, IEMG, ZC, VAR, and SSC are exoskeleton development. Some researchers [4], [5], [6]–[12]
commonly used by researchers to extract feature in sEMG. The had developed exoskeleton using this feature to extract EMG
quantification of the relationship between feature extraction and signal. WL is used to measure the complexity of the EMG
elbow joint angle was measured using the root mean square error signal and is an accumulation of EMG signal on windows
(RMSE) and Pearson Correlation Coefficient (CC). In this segment. Some researchers studied exoskeleton [13], [7], and
research, we found that the feature extraction ZC was the best [14] using a WL as feature extraction. Ding [15] used the
feature extraction in time domain to predict the elbow joint featured WL to determine the movement of the elbow joint.
angle, with normalized RMSE 0.187o and CC 0.907. With these Liu [16] used this feature for the development of
findings, it can facilitate the researcher in classifier step to build
multifunctional prostheses. SSC is a feature that can be used to
exoskeleton based EMG control.
get frequency information. SSC can indicate a number of the
Keywords: EMG, feature extraction, time domain, RMS, MAV, slope which changes the sign. The threshold is used to reduce
IEMG, ZC, VAR, SSC. the noise that comes with the EMG signal. Some studies on
exoskeleton [17], [12] used this feature to extract EMG signal.
I. INTRODUCTION ZC is a feature that can be used to get information about the
frequency of the signal without using a transformation process.
sEMG is a bioelectric signal that is generated when the
Zero crossing is calculated based on the number of the signal
muscles do a contraction. EMG signal has frequency
crosses the zero point. At the time of calculation of ZC, a
characteristic of 20 to 500 Hz and amplitude of 0 to 10 mV.
threshold requires to limit a minimum amplitude. Chan [12]
Over the years, the researchers have used the EMG signal to
used ZC feature as one of the featured input on the fuzzy rule,
control an external device, for example an exoskeleton and
in the development of prostheses.
prosthetic devices. EMG signal contains many informations
describing the human skeletal movement. This information can The previous researchers did not explain the validation of
be gotten using feature extraction. the feature extraction for angle prediction. In this paper, we
will represent the quantitative relationship between feature
Various feature extraction has been used by researchers to
extraction of EMG and elbow joint angle. In this research, we
analyze the EMG signal. Feature extraction can be done in time
used the statistic parameter of RMSE and Pearson Correlation
domain, frequency domain and time-frequency domain. Time
to give the quantitative value.
domain analysis is still the best analysis for real time
application. In this research, we used time domain feature II. METHOD
extraction to get the information of the upper limb elbow joint
angle. A good prediction in elbow joint angle is very important A. Experimental Protocol
in the development of exoskeleton devices. Upper limb elbow In this research, a male subject, 23 years old and has no
joint angle is the object for this research. Good accuracy in muscle problem was chosen. Before getting the data, the
joint angle prediction is very important to determine the subject was trained to move his hand for elbow flexion and
effectiveness of the exoskeleton. extension according to a digital metronome with 10 BPM, 15
The elbow joint angle prediction depends on how well we BPM and 20 BPM period. Metronome is a windows
define the feature extraction. Feature extraction of root mean application that can be used to synchronize the movement of
square (RMS), mean absolute value (MAV), variance of signal the elbow. TempoPerfect version 4.08 was used in this
research. Exoskeleton aluminum frame was installed to guide
the elbow flexion and extension movement in the range of 0 to EMG was processed using a home made bioamplifier that
145o. The elbow flexion motion is the movement of the elbow has been calibrated using a function generator. This
joint from 0o to 145o and the elbow extension motion is the bioamplifier used main component instrumentation amplifier
movement of the elbow joint from 145o to 0o. AD620 [18]. The bioamplifier module consisted of a
preamplifier, a bandpass filter, a notch filter 50 Hz and a
summing amplifier. Microcontroller ARM STM32F429 [19]
was used to record three channel EMG data, with analog to
digital (A/D) and USART function to communicate with the
computer. Three channel EMG was used with sampling
frequency 1000 Hz. In this experiment, the EMG signal and
elbow joint sensor were collected in range of 10, 15 and 20
BPM.
C. Digital Signal Processing
ARM STM32F429 In this research, we tested five feature extraction of RMS,
Biceps MAV, VAR, WL, SSC and ZC and two parameter statistic
EMG CH0 Internal descriptive of RMSE and CC. Time domain analysis and
T.Lateral Amplifier CH1 ADC 12 Bit statistic descriptive were calculated using Delphi
USART
AD620 CH2 Programming Version 7.0. Data acquisition was conducted
T.Long 3 Ch. CH3 using microcontroller STM32F429 and Coocox IDE Version
3.18. This five features were often used by some previous
researchers. Processing sequence of feature extraction was
Potensio shown in Fig. 3.
computer

Fig. 1. Microcontroller collects data from three channels EMG (biceps,


triceps lateral head, and triceps long head) and one analog from
potensiometer. Computer receives data from microcontroller via USART
communication.

B. Data Acquisition
The contact between the electrode and skin must be clean,
no dust, sweat, and oil, because it can influence data
acquisition of EMG signal. Therefore, before the electrode
installment, it must be doing the cleaning process on the skin
surface using alcohol. Ag (AgCl) electrode was used in this
research because it has a conductive pre-gelled, that could
muffle the noise artifact when the elbow moved. Electrode was
used to acquire surface EMG on biceps, triceps lateral head,
and triceps long head, each used two electrodes and one neutral
electrode.

Fig. 3. Digital signal processing of EMG feature extraction.

1) Mean absolute value (MAV) [20] or some references are


mentioned as average rectified value (ARV). It is defined
as:

= ∑ | | (1)

Where is the i-th EMG signal and N is number of sample in


each segment.

2) Root Mean Square (RMS) is formulated as amplitude


Gaussian random process. This feature extraction can be
Fig. 2. Electrode disposable Ag(AgCl) collects EMG signal from biceps,
triceps lateral head, triceps long head, and ground.
related to standard deviation, which is expressed as:
Where is the EMG feature extraction value, is the
= ∑ (2) maximum of EMG feature extraction value, is the
minimum feature extraction value and is
Where is the i-th EMG signal and N is number of sample in normalization of feature extraction value.
each segment.
D. Validation of The Feature Exraction
3) SSC (Sign Slope Change) is defined as number of slope
that changes every cycle of signals. This feature Validation is required to see the truth of the EMG-angle
modeling results. Two parameters which are often used to test
extraction can be used to count the frequency of signal,
the validation are root mean square error (RMSE) and
which is expressed as:
Pearson’s correlation coefficient ( ). The higher value of the
= ∑ [ [( − ) ( − )]] (3) correlation and a smaller value of RMSE the better the
performance of the model. The equation can be calculated as
1, ≥ ℎ ℎ follow [21]:
= ∑( )
0, ℎ = (8)
∑ ∑

Where is the i-th EMG signal, N is number of sample in (∑ )(∑ )



each segment and threshold is the amplitude level limitation. = (9)
(∑ ) (∑ )
(∑ )(∑ )
4) ZC (Zero Crossing ) is defined as number of time that the
amplitudo of signal crosses the zero base line, which is
Where, x is measured value and y is estimated value of
expressed as:
exoskeleton angle, n is number of samples.
= ∑ ( × )∩| − |≥ ℎ ℎ The CC can be in the range of -1 to 1. The value close to 1
(4) or -1 indicates a high correlation. RMSE can be any range of
1, ≥ ℎ ℎ measurement. The value close to 0 indicates a low error in the
= prediction. In this paper, RMSE and CC were used to calculate
0, ℎ
the accuracy between feature extraction and real angle of the
elbow joint and find how does the correlation between feature
Where is the i-th EMG signal, N is number of sample in extraction and real angle of the elbow joint.
each segment and threshold is the amplitude level limitation. III. RESULT
5) WL (Wave form Length) is summed of signal length for A. Signal EMG
time segment, which is expressed as: In this research, EMG signal was collected from biceps,
triceps lateral head, and triceps long head muscles. In this
= ∑ | − | (5) process, the data acquisition of the EMG signal and the elbow
joint angle were acquired at the same time. Fig. 4 showed the
Where is the i-th EMG signal and N is number of sample in three EMG signals for 10 BPM period. EMG signal was
each segment. recorded at the beginning of zero angle in the extension
position and the recording process was carried out till at the
6) VAR (Variance) is mean value of the square of deviation zero position again.
of the signal, which is expressed as:

= ∑ (6)

Where is the i-th EMG signal and N is number of sample in


each segment.
After the feature extraction process, then we got various
range of EMG feature value. In order to get the same value for
all values of extraction feature, it needs a normalization
process. The process of normalization is done by specifying
the minimum and maximum values for each feature extraction.

= (7)

Fig. 4. Amplitudo EMG signal form biceps, triceps lateral head, triceps long
head and real angle.
B. Feature Extraction of EMG signal

(a) (b)

(c ) (d)

(e) (f)
Fig. 5. Normalized feature extraction for various muscle: biceps, triceps lateral head, triceps long head and real angle of (a) RMS, (b) VAR, (c) MAV, (d) WL, (e)
ZC and (f) SSC for motion of BPM 15 period.

RMS, VAR, MAV, WL, ZC and SSC were the time domain 15 BPM period. From the Fig.5, we can see that triceps lateral
feature we selected. EMG signal was collected from biceps, head (green color) and triceps long head (blue color) did not
triceps lateral head and triceps long head. The Normalization follow the movement of the elbow joint angle. Visually, we can
was carried out for EMG feature extraction and real angle, in see that the EMG signal from the biceps muscle in some
the range of 0 to 1. This normalization was used to simplify the feature extraction followed the movement of the elbow joint
comparison process between the result of the feature extraction angle. In the next pages, we discussed only the relationship
and the real angle. Fig.5 was normalized feature extraction of between feature extraction and angle for biceps muscle. To get
the EMG for various muscles: biceps, triceps lateral head, the value of the relationship, we used descriptive statistic by
triceps long head and real angle. Fig. 5 was the example of using CC and RMSE only for biceps muscles. CC and RMSE
feature extraction for flexion and extension movement in the for various feature extraction were shown in Fig. 6 and Fig. 7.
This box plot diagram was the feature extraction of the EMG 0.4
signal from 10 BPM period. From those box plot diagram, we
can assume that the feature extraction from ZC, SSC and WL
will be taken into consideration for the next discussion. This 0.3
box plot diagram was created using Microsoft Excell 2013 and

RMSE
Add-In Real-Statistic from http://www.real-statistics.com. 0.2
This box plot diagram can show the distribution of the CC and
RMSE of the feature extraction. Box plot gives five
information of the distribution of the data (minimum, first 0.1
quartile, median, third quartile, and maximum).
0
1 10 BPM 15 BPM 20 BPM

0.8 Fig. 9. RMSE of ZC feature extraction (biceps) for various motion period,
10, 15 and 20 BPM.
CC

0.6
Fig. 8 to Fig.13 were the box plot diagram used to see the
distribution of the RMSE and CC of the feature extraction of
0.4 ZC, SSC and WL. In this box plot diagram, we tested each
selected feature extraction with various periods (10, 15 and 20
0.2 BPM). This test was performed to show the consistence of the
RMS ZC SSC VAR MAV WL CC and RMSE of the feature extraction This box plot diagram
can describe the percentile, mean, median, minimum and
Fig. 6. The CC from various feature extraction of EMG signal (biceps) maximum of the CC and RMSE.
for 10 BPM period.
1
0.6
0.95
0.5
0.9
0.4
CC

0.85
RMSE

0.3
0.8
0.2
0.75
0.1
0.7
0 10 BPM 15 BPM 20 BPM
RMS ZC SSC VAR MAV WL
Fig. 10. The CC of SSC feature extraction (biceps) for various motion
Fig. 7. RMSE for various feature extraction of EMG signal (biceps) for period, 10, 15 and 20 BPM.
10 BPM period.
0.35
1 0.3
0.25
0.95
0.2
RMSE

0.15
CC

0.9
0.1
0.85 0.05
0
0.8 10 BPM 15 BPM 20 BPM
10 BPM 15 BPM 20 BPM
Fig. 11. RMSE of SSC feature extraction (biceps) for various motion
period, 10, 15 and 20 BPM.
Fig. 8. The CC of the ZC feature extraction (biceps) for various motion
period, 10, 15 and 20 BPM.
1 IV. DISCUSSION
0.95 This paper is a part of our research in upper limb assistive
exoskeleton based EMG. The feature extraction selection is our
0.9 pre-analysis step that it will determine the accuracy of the
0.85 prediction. In this research, we seek the relationship between
the time domain feature extraction and elbow joint angle. The
CC

0.8
accuracy will be validated using RMSE and CC.
0.75
0.7 The first finding in this research, the biceps was the main
muscle that has the highest CC, 0.9055 and the lowest RMSE
0.65 0.187o as it is shown in Fig. 5 (e) and (f). In this Fig.5, the SSC
0.6 and ZC feature extraction can follow the angle position. The
10 BPM 15 BPM 20 BPM second finding in this research, SSC, ZC and WL were the best
feature extraction in correlation with elbow joint angle as
Fig. 12. The CC of WL feature extraction (biceps) for various motion described in box plot diagram Fig. 6 and Fig.7. for 10 BPM
period, 10, 15 and 20 BPM. period if compared with the other feature extraction, RMS,
MAV and VAR.
Repeatability of the feature extraction ZC, SSC and WL
0.4 was tested using ANOVA repeat measure with varying period,
0.35 10 BPM, 15 BPM and 20 BPM. This descriptive statistic test
0.3 is to see the consistency of the feature extraction. The alpha
for this ANOVA test was chosen at confidence value 0.05.
0.25
The p value of the CC for ZC, SSC and WL was 0.33192,
RMSE

0.2 0.746492, and 0.199033 consecutively. This value indicates


0.15 that there is no significant difference between the means of the
0.1 three methods. The p value of RMSE for ZC, SSC and WL
was 0.705347, 0.33383, and 0.732644 consecutively. This
0.05
value indicates that there is no significant difference between
0 the means of the three methods.
10 BPM 15 BPM 20 BPM According to the result of this research, Oskoei [22]
recommended for using MAV, WL, SSC and ZC because of
Fig. 13. RMSE of WL feature extraction (biceps) for various motion its relatively high accuracy, compared to other features.
period, 10, 15 and 20 BPM.
Kiguchi [23] have tested many time domain feature
extractions, and found out that waveform length (WL) was the
Table I. Mean RMSE and CC of ZC feature extraction for various BPM most fit feature extraction to describe the elbow joint angle.
The findings of this research can be used as a reference for
Parameter 10 BPM 15 BPM 20 BPM
other researchers doing the same in this research area. High
RMSE 0.1870 o
0.2266 o
0.2374 o accuracy in elbow joint prediction is very important in
±0.0107 ±0.0068 ±0.0061 exoskeleton based EMG development because it determines
CC 0.8788 0.9055 0.8997
±0.0069 ±0.0056 ±0.0064
the effectivity the exoskeleton. Some problems needed to be
solved in this research are the high variance in RMSE for 10
BPM period if compared to other period of 15 BPM and 20
Table II. Mean RMSE and CC of SSC feature extraction for various BPM BPM as shown in Fig. 8 and 9. In the future research, it will
Parameter 10 BPM 15 BPM 20 BPM need a compensate algorithm to minimize the difference
RMSE 0.2538 o
0.2313 o
0.2266 o
between period.
±0.0077 ±0.0065 ±0.0068
CC 0.8541 0.8797 0.8721 V. CONCLUSION
±0.0116 ±0.0071 ±0.0095 The objective of this research is to quantify the relationship
between feature extraction and elbow joint angle. Six time-
Table III. Mean RMSE and CC of WL feature extraction for various BPM domain of feature extraction are tested. The results show that
Parameter 10 BPM 15 BPM 20 BPM
ZC, WL and SSC are the best feature when compared with
other feature extraction. In this research, The CC of ZC
o o
RMSE 0.2216 0.2862 0.2374 o feature extraction at 15 BPM period has highest accuracy at
±0.0090 ±0.0150 ±0.0107
CC 0.8387 0.8393 0.8493
0.9055. The RMSE of ZC feature extraction at 10 BPM period
±0.0141 ±0.0067 ±0.0097 has highest accuracy at 1.870o.
Biceps has a high relationship which is correlated to elbow
angle joint position. While other muscles, triceps lateral head
and triceps long head have not shown significant relationship.
REFERENCES [12] F. H. Y. Chan, Y. S. Yang, F. K. Lam, Y. T. Zhang, and
[1] K. Kiguchi, T. Tanaka, and T. Fukuda, “Neuro-fuzzy P. a. Parker, “Fuzzy EMG classification for prosthesis
control of a robotic exoskeleton with EMG signals,” control,” IEEE Trans. Rehabil. Eng., vol. 8, no. 3, pp.
IEEE Trans. Fuzzy Syst., vol. 12, no. 4, pp. 481–490, 305–311, 2000.
2004. [13] J. Rosen, M. Brand, M. B. Fuchs, and M. Arcan, “A
[2] D. S. Andreasen, S. K. Allen, and D. a. Backus, myosignal-based powered exoskeleton system,” IEEE
“Exoskeleton with EMG based active assistance for Trans. Syst. Man, Cybern. Part ASystems Humans., vol.
rehabilitation,” Proc. 2005 IEEE 9th Int. Conf. Rehabil. 31, no. 3, pp. 210–222, 2001.
Robot., vol. 2005, pp. 333–336, 2005. [14] X. Guo, P. Yang, L. Chen, X. Wang, and L. Li, “Study
[3] C. Loconsole, S. Dettori, A. Frisoli, C. A. Avizzano, and of the control mechanism of robot-prosthesis based-on
M. Bergamasco, “An EMG-based approach for on-line the EMG processed,” Proc. World Congr. Intell. Control
predicted torque control in robotic-assisted Autom., vol. 2, pp. 9490–9493, 2006.
rehabilitation,” 2014 IEEE Haptics Symp., pp. 181–186, [15] Q. Ding, X. Zhao, A. Xiong, and J. Han, “A Novel
2014. Motion Estimate Method of Human Joint with EMG-
[4] S. Wang, Y. Gao, J. Zhao, T. Yang, and Y. Zhu, Driven Model,” in 2011 5th International Conference on
“Prediction of sEMG-based tremor joint angle using the Bioinformatics and Biomedical Engineering, 2011, no. 1,
RBF neural network,” 2012 IEEE Int. Conf. pp. 1–5.
Mechatronics Autom. ICMA 2012, pp. 2103–2108, 2012. [16] X. Liu, R. Zhou, L. Yang, and G. Li, “Performance of
[5] K. Kiguchi and Y. Hayashi, “Estimation of joint torque various EMG features in identifying arm movements for
for a myoelectric arm by genetic programming based on control of multifunctional prostheses,” Proc. - 2009
EMG signals,” World Autom. Congr. (WAC), 2012, pp. IEEE Youth Conf. Information, Comput. Telecommun.
1–4, 2012. YC-ICT2009, pp. 287–290, 2009.
[6] Z. Tang, K. Zhang, S. Sun, Z. Gao, L. Zhang, and Z. [17] M. R. Ahsan, M. I. Ibrahimy, and O. O. Khalifa, “EMG
Yang, “An upper-limb power-assist exoskeleton using motion pattern classification through design and
proportional myoelectric control.,” Sensors (Basel)., vol. optimization of Neural Network,” 2012 Int. Conf.
14, no. 4, pp. 6677–94, 2014. Biomed. Eng., no. February, pp. 175–179, 2012.
[7] Z. O. Khokhar, Z. G. Xiao, and C. Menon, “Surface [18] A. Devices, “AD620 Low Cost, Low Power
EMG pattern recognition for real-time control of a wrist Instrumentation Amplifier,” Online] www.
exoskeleton.,” Biomed. Eng. Online, vol. 9, p. 41, 2010. datasheetcatalog. com.[Online], pp. 1–20, 1999.
[8] R. a R. C. Gopura and K. Kiguchi, “Electromyography [19] St, “Datasheet STM32F427xx STM32F429xx,” no.
(EMG)-signal based fuzzy-neuro control of a 3 degrees April, 2014.
of freedom (3DOF) exoskeleton robot for human upper- [20] A. Phinyomark, C. Limsakul, and P. Phukpattaranont,
limb motion assist,” J. Natl. Sci. Found. Sri Lanka, vol. “A Novel Feature Extraction for Robust EMG Pattern
37, no. 4, pp. 241–248, 2009. Recognition,” vol. 1, no. 1, pp. 71–80, 2009.
[9] T. D. Lalitharatne, K. Teramoto, Y. Hayashi, T. [21] F. Mobasser and K. Hashtrudi-Zaad, “Hand force
Nanayakkara, and K. Kiguchi, “Evaluation of Fuzzy- estimation using electromyography signals,” Proc. -
Neuro Modifiers for Compensation of the Effects of IEEE Int. Conf. Robot. Autom., vol. 2005, no. April, pp.
Muscle Fatigue on EMG-Based Control to be Used in 2631–2636, 2005.
Upper-Limb Power-Assist Exoskeletons,” J. Adv. Mech. [22] M. a. Oskoei and H. H. H. Hu, “Support Vector
Des. Syst. Manuf., vol. 7, no. 4, pp. 736–751, 2013. Machine-Based Classification Scheme for Myoelectric
[10] P. W. Persons, “Upper-Limb Exoskeletons for Control Applied to Upper Limb,” IEEE Trans. Biomed.
Physically Weak Persons,” Robotics, no. August, 2007. Eng., vol. 55, no. 8, pp. 1956–1965, 2008.
[11] K. Kiguchi and Q. Quan, “Muscle-model-oriented EMG- [23] H. He and K. Kiguchi, “A study on EMG-based control
based control of an upper-limb power-assist exoskeleton of exoskeleton robots for human lower-limb motion
with a neuro-fuzzy modifier,” IEEE Int. Conf. Fuzzy assist,” Proc. IEEE/EMBS Reg. 8 Int. Conf. Inf. Technol.
Syst., pp. 1179–1184, 2008. Appl. Biomed. ITAB, vol. 00, pp. 292–295, 2008.

You might also like