35th Annual International Conference of the IEEE EMBS

Osaka, Japan, 3 - 7 July, 2013

Exercise Muscle Fatigue Detection System Implementation via

Wireless Surface Electromyography and Empirical Mode
Decomposition
Kang-Ming Chang, Shing-Hong Liu, Jia-Jung Wang, and Da-Chuan Cheng

AbstractSurface electromyography (sEMG) is an important

measurement for monitoring exercise and fitness. A wireless
Bluetooth transmission sEMG measurement system with a
sampling frequency of 2 KHz is developed. Traditional muscle
fatigue is detected from the median frequency of the sEMG
power spectrum. The regression slope of the linear regression of
median frequency is an important muscle fatigue index. As
fatigue increases, the power spectrum of the sEMG shifts toward
lower frequencies. The goal of this study is to evaluate the
sensitivity of empirical mode decomposition (EMD) quantifying
the electrical manifestations of the local muscle fatigue during
exercising in health people. We also compared this method with
the raw data and discrete wavelet transform (DWT). Five male
and five female volunteers participated. Each subject was asked
to run on a multifunctional pedaled elliptical trainer for about 30
minutes, twice a week, and there were a total of six recording
times for each subject with a wireless EMG recording system.
The results show that sensitivity of the highest frequency
component of EMD is better than the highest frequency
component of DWT, and raw data.

I. INTRODUCTION
Muscle fatigue is thought of as a loss of required or
expected force and has been an attractive research issue for a
long time. The nature of muscle fatigue and its relation to
muscle activity have been studied[1]. Spectral parameters
such as the mean frequency (MNF) and the median frequency
(MF) derived from the sEMG power spectrum are widely used
to detect static and dynamic muscle contractions [2]. The
Fourier transform is one of methods used to obtain the power
spectrum of a signal. However, within the analysis window,
the signal must be stationary or exhibit a periodic frequency;
otherwise, the resulting spectrum will make little sense.
Dimitrova et al. proposed new spectral indices of muscle
fatigue (FInsmk) that perform better than the traditional MNF
and MF[3]. Wavelet-based spectra and derived spectrum
features have been used to compare the traditional power
spectrumderived MNF and MF performance for fatigue
quantification.

*Shing-Hong Liu is with the Department of Computer Science and

Information Engineering, Chaoyang University of Technology, Taichung,
41354, Taiwan, ROC (phone: 866-423323000; e-mail: shliu@cyut.edu.tw).
Kang-Ming Chang. is with Asia University, Taichung, 41354 Taiwan,
ROC.
Da-Chuan Cheng is with the Department of Biomedical Imaging and
Radiological Science, China Medical University, Taichung, Taiwan, ROC
Jia-Jung Wang is with the Department of Biomedical Engineering, I-Shou
University, Kaoshiung, Taiwan, ROC.

978-1-4577-0216-7/13/$26.00 2013 IEEE

Recently, a novel nonstationary and nonlinear signal

processing technique has been proposed, known as empirical
mode decomposition (EMD). EMD was introduced by Huang,
and it has been widely used for nonlinear signal analysis[4].
The principle of EMD is based on a decomposition derived
from the data, and EMD is useful in the analysis of nonlinear
and nonstationary time series signals. With an iterative
decomposition of signals, EMD separates the full signal into
ordered elements with frequencies ranging from high to low in
each intrinsic mode function (IMF) level. The filter bank-like
property of EMD has been widely applied in many fields, such
as to the sound analysis of an infant crying to assess a
newborns pain [5] and classification of ship-radiated
underwater sound [6]. Weather-related issues are the main
application for EMD scholars [7]. Another major application
of EMD is biomedical signal analysis [8-9]. The decomposed
IMFs were further extracted with the power or entropy
approach to analyze the nonstationary biosignals for noise
reduction and for feature extraction[10]. The main topics of
concern for EEGthe detection of epileptic seizure [11] and
evoked potential extraction [12] have been investigated by
EMD with impressive results. Heart rate signal analysis is
another major application of EMD. Reconstructions of
selected IMFs of heart beat intervals were used for noise
filtering [13], feature extraction for discrimination from local
anesthesia [14], fetal heart rate monitoring [15], and
ventricular fibrillation detection [16]. Modulation of
respiratory sinus arrhythmia between respiratory and heart
beat signal is also achieving promising results. EMD had been
applied to extract the MNF of sEMG as a muscle fatigue index
[17]. Srhoj et al. have extracted the MF from selected IMFs of
sEMG recorded over quadriceps muscles during cyclic
dynamic contractions [18]. Their results showed that
HHT-derived spectral and linear regression parameters were
consistent and more reliable than those obtained with the
short-time Fourier transform and the wavelet transform.
To reduce the nonstationary problem of the long EMG
segment, this study investigates the EMD performance for
muscle fatigue spectrum estimation and compares it with
discrete wavelet transform (DWT) and EMD. The MFs were
used as fatigue indices during dynamic contractions. There
were 10 volunteers who joined this experiment; they ran in a
multifunctional pedaled elliptical trainer. A self-designed
wireless device was used to record the sEMG signal of the
vastus lateralis in the left leg of each volunteer. Each subject
performed six experiments in three weeks. Furthermore, the
comparison of the different decomposition methods revealed

1001

that the IMF 1 component of EMD was best for evaluating

muscle fatigue.

C. EMD algorithm
The EMD algorithm used in this study comprised the
following steps [4]:

II. METHODS

Step 1: Extrema (maxima and minima) of the signal, x(t), are

identified.

A. sEMG recording and subjects

A wireless sEMG recording device developed by the
authors was worn on the left lateral waist of the subject to
measure the sEMG signal. The gain of the device is 1000, and
the bandwidth is 30 Hz to 1000 Hz to avoid the aliasing
problem. This device is based on the microcontroller
MSP430-F5438 as the core structure, which is a 12-bit
analog-to-digital converter with a sampling rate of 2000 Hz.
The digital EMG signal is transferred by a Bluetooth chip to a
remote server. A Visual Basic-based interface system is used
to display and store the digital EMG data in real time [19].
There were ten volunteers involved (5 male and 5 female),
with ages ranging from 19 to 27 years. Subjects were required
to run in a multifunctional pedaled elliptical trainer (Johnson
E8000). We measured the vastus lateralis of the left leg. The
surface electrodes used for the EMG recording were Ag/AgCl
with a 10 mm diameter on self-adhesive supports. The bipolar
electrodes were placed over the midline of the muscle belly
between the motor point and the myotendinous junction, and
the inter-electrode distance was 5 cm. The electrode
arrangement ensured negligible crosstalk between adjacent
muscles. The positions of the electrodes for each subject were
recorded, and the electrodes were placed at the same position
in the subsequent experiments.
B. Experimental procedure for evaluating muscle fatigue
The muscle fatigue experiment is based on the
following procedures.
Step 1: The subjects are required to wear the wireless sEMG
device. Alcohol is used to clean the surface, and electrolytic
gel is smeared on the electrodes to decrease the contact
impedance. Athletic tape is used to fix the electrodes and so
avoid movement of the electrodes. Before data collection, a
consent form was signed by each subject.
Step 2: There are three load levels in the multifunctional
pedaled elliptical trainer, L2, L4 and L6, with L2 being light
and L6 being heavy. The speed range of L2 is 55-60 steps per
minute (SPM) for males and 50-55 SPM for females. The
speed range of L4 is 60-70 SPM for males and 55-65 SPM for
females. The subjects are required to run at their maximum
speed until exhaustion for L6, which has a faster speed range
than L4. A ten minute session is required for both L2 and L4.
The average duration for L6 was also approximately 10
minutes. In the pre-experiment, the subjects tested the speed
range and chose the most appropriate speeds for the L2 and L4
levels, separately, and ran at their maximum speed for the L6
level. These speeds were recorded for every subject. In the
experiments, the subjects ran at their self-selected speeds
during the experimental procedure.
Step 3: Each subject was recorded twice a week at the same
time, and there were a total of six recording times for each
subject.

Step 2: Upper and lower envelope of the extreme point is

developed.
Step 3: Mean function of the upper and lower envelope, m(t).
Step 4: Difference signal d(t)=x(t)-m(t).
Step 5: If d(t) becomes a zero-mean process, then the iteration
stops, and d(t) is a first IMF (IMF1), called c1(t); otherwise, go
to step 1 and replace x(t) with d(t).
Step 6: Residue signal r(t)=x(t) c1(t).
Step 7: Replace x(t) with r(t) and repeat the procedure from
steps 1 to 6 to obtain the second IMF (IMF 2), called c2(t). To
obtain cn(t), continue steps 1 to 6 after n iterations. The
process is stopped when the final residual signal r(t) is
obtained as a monotonic function.
Now, the original signal can be represented as:

xt

c (t ) r (t ) .
i 1

(1)

Often, we can regard r(t) as cn+1(t).

D. Discrete wavelet analysis
Assuming the raw sEMG signal is x[n], the DWT
decomposition involves the following filtering process:
,
Aj n
k A j 1k h2n k

(2)

,
D j n
k Aj 1k g2n k

(3)

where A0[n]= x[n] and Aj[n] and Dj[n] indicate the coarse and
detailed sequences, respectively, after the jth decomposition.
The variable h[n] represents the half-band low-pass filter, and
g[n] represents the half-band high-pass filter. The original
signal is decomposed from the high-frequency component to
the low-frequency component as a combination of Aj[n] and
Dj[n]. For example, if the decomposition level is 5 (j=5), then
the original signal can be represented as:
x[n]=D1[n]+D2[n]+D3[n]+D4[n]+D5[n]+A5[n].

(4)

E. Muscle signal processing

The recorded sEMG is divided into segments, and a Fast
Fourier Transform is performed. Each segments MF is
extracted. The MF is defined as the frequency at which the
accumulated spectrum energy is half of the total spectrum
energy, as shown in equation (5):

MF

p f df

MF

p f df

1
2

p f df

(5)

The sEMG segment window size is 30 seconds, and the step

size is 15 seconds. There is one MF for each sEMG segment.
A further linear regression analysis is applied to the MF series

1002

during the three stages of the muscle fatigue examinations.

The linear regression equation is defined as:

x
n

i 1

x
n

i 1

x yi y

y
2

i 1

(7)

L4

In this study, the SIGMAPLOT software package was

used to conduct the data analysis. Descriptive statistics were
applied to subjects personal information and muscle fatigue
parameters (regression slope, A, and correlation coefficient,
R). The data were represented as the mean (standard
deviation). Statistical testing of the muscle fatigue parameters
obtained from the raw data and the different decomposition
methods was performed using t-tests. The significance level
for the p value was set at 0.05.

L6
All

Coefficient

MF (Hz)

Raw

-0.012

0.859

269.6 (5.8)

DWT,
D1

-0.030

0.894

665.7 (14.5)

D2

-0.001

0.365

344.6 (1.4)

D3

-0.001

0.218

184.6 (1.3)

EMD

-0.049

0.865

474.0 (24.5)

-0.025

0.874

269.7 (12.4)

IMF3

-0.010

0.813

167.7 (5.1)

RAW
(n=60)
-0.0164*
(0.0144)
-0.0125*
(0.0100)
-0.0193*
(0.0154)
-0.0197*
(0.0128)

DWT
(n=60)
-0.0232*
(0.0176)
-0.0209*
(0.0157)
-0.0276*
(0.0264)
-0.0296*
(0.0153)

EMD
(n=60)
-0.0456*
(0.0356)
-0.0362*
(0.0302)
-0.0501
(0.0503)
-0.0594*
(0.0364)

SD is standard deviation, ALL represents the entire experiment, P<0.05*;

In Table 2, the MF slope of the raw data is also

significantly different from the highest-frequency components
of the other two decomposed methods. The ranges of the
absolute MF slope value are also EMD> DWT> raw EMG
within the three loading levels and the complete experiment.

III. RESULTS

Slope Hz/s

IMF2

Level

F. Statistics

TABLE I. TYPICAL REGRESSION RESULTS FOR THE RAW DATA

AND THE DECOMPOSED SIGNALS OF THE DWT AND EMD,

MF (Hz)

TABLE 2. STATISTICAL RESULTS OF THE MF SLOPE OF THE RAW

DATA AND THE DECOMPOSITION METHODS WITHIN THE THREE
LEVELS AND FOR THE ENTIRE EXPERIMENT. THE DATA ARE
REPRESENTED AS THE MEAN (SD).

L2

DWT and EMD were used to decompose the sEMG

signal. The lower IMF and the lower wavelet detail function
both correspond to higher-frequency components. Table 1
shows the analysis results of one experiment for the raw data
and the decomposed signals of the other two methods for the
entire 30 minute experiment. The MF slope of the raw sEMG
is -0.012 Hz/s. For the DWT decomposition, the absolute MF
slope of the first detail component is larger than the rest of the
decomposition (D1, slope = - 0.03 Hz/s). This is also true for
EMD: the absolute MF slope of the first IMF is significantly
larger than that of the other IMFs (IMF1 of EMD, slope = 0.049 Hz/s. In the following analysis, only the D1 component
of the DWT and the IMF1 component of the EMD were
chosen for further MF estimation and regression analysis. The
absolute MF slope of the raw data and the high-frequency
components of the three methods is EMD (0.049 Hz/s) >
DWT (0.030 Hz/s) > raw EMG (0.012 Hz/s). Figure 1 shows
the MF distributions of the raw data, the D1 component of the
DWT, and the IMF1 components of the EMD during a
complete experiment.

Coefficient

IMF1

y Ax b ,

(6)
where y is estimated as the MF, x is the time interval, A is the
regression slope and b is the bias. The greater the muscle
fatigue, the smaller the slope [20]. We also used the
correlation coefficient (R) to represent the stability of sEMG
in terms of muscle fatigue. It is well known that the MF shifts
toward lower frequencies as a muscle fatigues. Parameter R
and A were used as indexes of the muscle fatigue.

Slope Hz/s

IV. DISCUSSIONS AND CONCLUSION

In Table 1, the absolute MF slope of the
highest-frequency component of the different decomposition
methods is significantly larger than the rest of the
decomposition. The results show that the intrinsic information
about muscle fatigue could be embedded in the
high-frequency portion of the sEMG. Therefore, in this study,
we only used the highest-frequency component of the different
decomposition methods to evaluate the muscle fatigue.
From the results, EMD has been proven to
quantify the electrical manifestations of muscle fatigue at the
local muscle being better than the DWT and raw data. The
reason could be that EMD suits the nonlinear signal
decomposition of the intrinsic mode function. Although EMD
acted as a filter-bank, there was no strict bandwidth restriction
with the IMF. The frequency range of each IMF level is
adaptive, depending on the raw signal content. The DWT
decomposition is based on the successive filtering of the
symmetric half-band high-pass and low-pass filters. The
frequency range of the more detailed component is nearly
twice that of the adjacent less detailed component. The EMD
approach can extract major high-frequency components in the
first IMF level with better adaptation than wavelet transforms.
Although the MF of the D1 component of the DWT is larger
than the IMF 1 component of the EMD, the absolute MF slope
of the IMF 1 component of the EMD is larger than the D1
component of the DWT with the time course of fatigue.
Finally, we used our designed wireless device to record the
sEMG and quantify the electrical manifestations of muscle
fatigue at the local muscle. We found that the intrinsic fatigue

1003

information of the sEMG could be embedded in the

high-frequency component. Two decomposed methods, DWT
and EMD, were used to extract this component. The
preliminary results revealed the potential of EMD for sEMG
signal processing.

[2]
[3]
[4]

RAW
290

[5]

285

280

[6]

Hz

275

270

[7]
265

260

[8]

255
0

200

400

600

800

1000

1200

1400

(a) Time (sec)

[9]

DWT D1
710

[10]

700
690

[11]

Hz

680
670
660

[12]
650
640

[13]

630
0

200

400

600

800

1000

1200

1400

(b) Time(sec)

[14]

EMD IMF1
540

520

[15]

Hz

500

[16]
480

460

[17]

440

[18]

420
0

200

400

600

800

1000

1200

1400

(c) Time(sec)

Figure 1. The distribution of MF during a complete experiment, (a) Raw data,

(b) D1 of DWT, (c) IMF1 of EMD.

[19]

Acknowledge

[20]

This research is in part funded by the National Science

Council (Taiwan) with the grant number NSC
100-2221-E-324-013-MY2, NSC 101-2221-E-214 -001
REFERENCES
[1]

S. L. Morris and G. T. Allison, "Effects of abdominal muscle fatigue on

anticipatory postural adjustments associated with arm raising," Gait
Posture, vol. 24, pp. 342-8, Nov 2006.

1004

M. Gonzalez-Izal, et al., "EMG spectral indices and muscle power

fatigue during dynamic contractions," J Electromyogr Kinesiol, vol.
20, pp. 233-40, Apr 2010.
G. V. Dimitrov, et al., "Muscle fatigue during dynamic contractions
assessed by new spectral indices," Med Sci Sports Exerc, vol. 38, pp.
1971-9, Nov 2006.
N. E. Huang, et al., "The empirical mode decomposition and the Hilbert
spectrum for nonlinear and non-stationary time series analysis.,"
Physical and Engineering Sciences, vol. 454, p. 93, 1998.
B. Mijovic, et al., "Assessment of Pain Expression in Infant Cry Signals
Using Empirical Mode Decomposition," Methods of Information in
Medicine, vol. 49, pp. 448-452, 2010.
F. Bao, et al., "Ship classification using nonlinear features of radiated
sound: An approach based on empirical mode decomposition," Journal
of the Acoustical Society of America, vol. 128, pp. 206-214, Jul 2010.
A. C. Yang, et al., "Temporal Associations between Weather and
Headache: Analysis by Empirical Mode Decomposition," Plos One,
vol. 6, Jan 31 2011.
M. Hassan, et al., "Combination of Canonical Correlation Analysis and
Empirical Mode Decomposition Applied to Denoising the Labor
Electrohysterogram," IEEE Transactions on Biomedical Engineering,
vol. 58, pp. 2441-2447, Sep 2011.
A. Karagiannis and P. Constantinou, "Noise-assisted data processing
with empirical mode decomposition in biomedical signals," IEEE
Trans Inf Technol Biomed, vol. 15, pp. 11-8, Jan 2011.
M. Hu and H. Liang, "Adaptive multiscale entropy analysis of
multivariate neural data," IEEE Trans Biomed Eng, vol. 59, pp. 12-5,
Jan 2012.
R. B. Pachori and V. Bajaj, "Analysis of normal and epileptic seizure
EEG signals using empirical mode decomposition," Computer
Methods and Programs in Biomedicine, vol. 104, pp. 373-381, Dec
2011.
C. L. Yeh, et al., "Extraction of single-trial cortical beta oscillatory
activities in EEG signals using empirical mode decomposition,"
Biomedical Engineering Online, vol. 9, Jun 17 2010.
K. M. Chang and S. H. Liu, "Gaussian Noise Filtering from ECG by
Wiener Filter and Ensemble Empirical Mode Decomposition," Journal
of Signal Processing Systems for Signal Image and Video Technology,
vol. 64, pp. 249-264, Aug 2011.
K. Shafqat, et al., "Empirical mode decomposition analysis of HRV
data from patients undergoing local anaesthesia (brachial plexus
block)," Physiological Measurement, vol. 32, pp. 483-497, Apr 2011.
N. Krupa, et al., "Antepartum fetal heart rate feature extraction and
classification using empirical mode decomposition and support vector
machine," Biomedical Engineering Online, vol. 10, Jan 19 2011.
E. M. Abu Anas, et al., "Exploiting correlation of ECG with certain
EMD functions for discrimination of ventricular fibrillation,"
Computers in Biology and Medicine, vol. 41, pp. 110-114, Feb 2011.
H. B. Xie and Z. Z. Wang, "Mean frequency derived via Hilbert-Huang
transform with application to fatigue EMG signal analysis," Computer
Methods and Programs in Biomedicine, vol. 82, pp. 114-120, May
2006.
V. Srhoj-Egekher, et al., "The application of Hilbert-Huang transform
in the analysis of muscle fatigue during cyclic dynamic contractions,"
Medical & Biological Engineering & Computing, vol. 49, pp. 659-669,
Jun 2011.
K. M. Chang, et al., "A Wireless sEMG Recording System and Its
Application to Muscle Fatigue Detection," Sensors, vol. 12, pp.
489-499, Jan 2012.
H. Xie and Z. Wang, "Mean frequency derived via Hilbert-Huang
transform with application to fatigue EMG signal analysis," Comput
Methods Programs Biomed, vol. 82, pp. 114-20, May 2006.