dynamic muscle fatigue detection using self-organizing maps

www.elsevier.com/locate/asoc

Applied Soft Computing 5 (2005) 391–398

Dynamic muscle fatigue detection using self-organizing maps

Dimitrios Moshoua, Ivo Hostensa, George Papaioannoub,1,*, Herman Ramona

aDepartment of Agro-Engineering and Economics, Laboratory for Agro-Machinery and Processing,

Katholieke Universiteit Leuven, Kasteelpark Arenberg 30, 3001 Heverlee, BelgiumbDepartment of Biomedical Engineering, The Catholic University of America,

620 Michigan Avenue N.E., Washington, DC 20064, USA

Received 15 June 2002; received in revised form 5 April 2004; accepted 10 September 2004

Abstract

Wavelets are used for the processing of signals that are non-stationary and time varying. The electromyogram (EMG)

contains transient signals related to muscle activity. Wavelet coefficients are proposed as features for identifying muscle fatigue.

By observing the approximation coefficients it is shown that their amplitude follows closely the muscle fatigue development.

The proposed method for detecting fatigue is automated by using neural networks. The self-organizing map (SOM) has been

used to visualize the variation of the approximation wavelet coefficients and aid the detection of muscle fatigue. The results show

that a 2D SOM separates EMG signatures from fresh and fatigued muscles, thus providing a visualization of the onset of fatigue

over time. The map is able to detect if muscles have recovered temporarily. The system is adaptable to different subjects and

conditions since the techniques used are not subject or workload regime specific.

# 2004 Published by Elsevier B.V.

Keywords: Wavelets; Self-organizing maps; Neural networks; Biomedical signal processing; EMG; Comfort; Muscle fatigue; Transient

signal; Visualization

1. Introduction

Electromyograms (EMG) signals have typically

many transient components, which can be isolated and

classified according to their physiological signifi-

* Corresponding author. Tel.: +32 16 321 922;

fax: +32 16 328 590.

E-mail addresses: [email protected]

(D. Moshou), [email protected] (I. Hostens),

[email protected] (G. Papaioannou),

[email protected] (H. Ramon).1 Tel.: +1 202 319 5891.

1568-4946/$ – see front matter # 2004 Published by Elsevier B.V.

doi:10.1016/j.asoc.2004.09.001

cance. The EMG signal contains transient signals

related to muscle activity. An example of such a signal

is shown in Fig. 1. These signals contain information

on which muscle groups have been activated.

Another property that can be extracted from

analyzing the transient myoelectric signal (MES) is

the presence of fatigue. Muscle fatigue can be defined

as reduction in the force generating capacity of a

muscle. This change in mechanical performance

capacity reflects in EMG changes. In the presented

work car driving is used as an example of a situation

where muscle fatigue can occur. The identification of

D. Moshou et al. / Applied Soft Computing 5 (2005) 391–398392

Fig. 1. An example of a surface EMG. The recording shows the contraction of a trapezoid muscle.

specific groups of muscles that show signs of fatigue

during the execution of certain driving tasks can be

associated with vibration transmission to the body of

the driver. In order to be able to assess the effect of the

vibrations that are transmitted to the human body the

myoelectric signals are recorded and analyzed.

However, a number of problems are associated with

the recording and the analysis of the transient

myoelectric signal. A main problem is due to the

noisy character of the signal. The noisy character is

due to the fact that several muscle activations that

occur simultaneously are recorded from the electro-

de(s) that are attached to the skin. What is actually

being recorded is the superposition of several muscle

activations filtered through different transfer paths of

the surrounding tissues and the skin itself.

The most efficient way to use the transient MES is by

extracting from the signal a number of useful features.

Such features should be closely related with the

condition of the muscle and the type of the muscle

group that is activated. An example of this is the

identification of the type of movement of the arm from

recorded triceps and biceps signals. Certain combina-

tions of time-based features that have been extracted

from recorded MES can be used as input to a neural

network classifier in order to accurately identify

movement types in the case of amputees. Such an

application has been reported in [1]. However, time

domain features are not so robust against signal

amplitude variations. A number of more advanced

features have been investigated in [2]. These features

were based on the use of different types of wavelet

transforms and specifically the wavelet packet trans-

form. This transform was used in combination with

neural networks to identify the type of movement of the

arm. It was shown that in combination with principal

components analysis (PCA), a very accurate classifier

could be constructed based on wavelet features.

Basmajian and De Luca [3] have shown how the

effects of superposition and tissue filtering join to

produce a single motor unit action potential (MUAP)

detected by electrodes. In additional work both Kupa

et al. [4] and Solomonow et al. [5] have reported

overall spectral shifts in the surface EMG. These shifts

are attributed to the type of muscle fibers activated and

may therefore be used for characterization of motor

unit recruitment and muscle composition. Investiga-

tion of these shifts have been limited to changes in the

median frequency of the power spectrum derived

using windowed FFT. These methods, however, try to

capture time varying spectral shifts that are due to

changes in the underlying irregular discrete wave-

forms, using continuous regular sine waves.

Wavelet analysis allows investigation of these

changes using irregular discrete ‘‘little waves’’. These

are functions whose shape and duration are much

more similar to an actual MUAP. By scaling and

translating these ‘‘little waves’’ the resulting decom-

position may produce information about the recruit-

ment of the motor units of different type.

Another interesting property that can be extracted

from analyzing the surface electromyogram (SEMG)

is the presence of fatigue. Successful analysis of

transient SEMG signals for fatigue analysis requires

suitable spectral estimation techniques. The use of

short-time Fourier transform avoids the question of

stationary signals by defining the time-interval (local

stationary signal) to be used in the computation. There

are however restrictions in the use of Fourier transform

due to the time frequency resolution. It is found that

the minimum allowable window width is approxi-

mately 250 ms for a typical SEMG signal [6–8].

A number of studies mention already the good

performance of wavelets in transient SEMG proces-

sing for fatigue detection, reaction time detection or

pattern recognition [9–11].

D. Moshou et al. / Applied Soft Computing 5 (2005) 391–398 393

In the presented work car driving is used as an

example of a situation where muscle fatigue can occur.

The identification of specific groups of muscles that

show signs of fatigue during the execution of certain

tasks can be associated with the vibration transmission

from the seat and the steering wheel. In order to be

able to assess the effect of the vibrations that are

transmitted to the human body the SEMG signals are

recorded and analyzed. However, it is clear that the

transmitted vibrations are not directly measured from

the SEMG signal. The SEMG signal records the

collective response of activated muscle groups.

In the presented work, flexible intelligent classi-

fication methods based on wavelet preprocessing of

SEMG signals and self-organizing maps have been

used to identify car driver fatigue. The effectiveness

and performance of the proposed hybrid technique is

proven against currently used techniques that are not

able to detect fatigue under dynamic conditions.

The proposed technique is using SEMG signal

recording to detect the early onset of muscle fatigue.

Analogous methods exist for recording and inter-

pretation of EEG and ECG towards monitoring of

cardiac and brain pathological conditions. The

proposed method for muscle fatigue detection shows

potential for applications in ergometrics, fitness

therapy, myopathy or for early warning systems that

prevent the onset of muscle fatigue.

In the current work wavelets are used for the

processing of signals that are non-stationary and time

varying. The EMG is such a signal. Wavelet

coefficients are proposed as features for identifying

muscle fatigue. The proposed method for detecting

muscle fatigue is automated by using neural networks.

The self-organizing map (SOM) neural network is

proposed as a visualization and detection tool. The

SOM is combined with wavelet feature preprocessing.

More specific, it is used to visualize the variation of the

approximation wavelet coefficients and aid the

detection of muscle fatigue. The map is able to detect

if muscles have recovered temporarily.

2. Experimental procedure

Five healthy subjects (four male and one female;

four right-handed and one left-handed) gave their

written informed consent to participate in this study.

Their mean (S.D.) age, weight and height were 25.3

years (2.0), 73.6 kg (7.1) and 179 cm (7), respectively.

The local ethical committee approved the experi-

mental design.

Subjects performed the same repeated movement

of turning the steering wheel during driving a car. This

type of movement was just an example of dynamic

movement (non-constant force) that was used for the

experiments. A rest period of 30 min was allowed

between the two consecutive measurement sessions

for each subject.

The surface EMG signals were obtained from the

trapezoid muscle at the dominant side of the body. To

reduce the electrical impedance between the skin and

the electrode, pre-gelled bipolar surface EMG

electrodes (Ag–Ag chloride, 10 mm diameter,

Nikomed, Denmark) were used. After shaving and

cleaning the skin with ether, the electrodes were

placed on the muscles according to standard proce-

dures resulting from the SENIAM studies [12,13]. The

distance between the pairs of surface electrodes was

20 mm. A bipolar EMG device (ME3000p8, Mega

Electronics Ltd., Finland) was used to continuously

register the electrical activity. The signals were pre-

amplified (analog differential amplifiers, preamplifier

gain 375, CMMR 110 dB) and sampled at 1000 Hz.

3. Wavelets

The Fourier transform converts a signal into a

continuous series of sine waves, each of which is of

constant frequency and amplitude and of infinite

duration. In contrast, most real-world signals (such as

music or images) have a finite duration and abrupt

changes in frequency.

A wavelet is a waveform that is bounded in both

frequency and duration. Wavelet transforms convert a

signal into a series of wavelets. In theory, signals

processed by the wavelet transform can be stored more

efficiently than ones processed by Fourier transform.

Wavelets can also be constructed with rough edges, to

better approximate real-world signals.

The basic idea underlying wavelet analysis consists

of expressing a signal as a linear combination of a

particular set of functions (wavelet transform, WT),

obtained by shifting and dilating one single function

called a mother wavelet. Several different mother


Fig. 2. The tiling of the time–frequency plane in the case of short-time Fourier transform (STFT) and discrete wavelet transform (DWT),

respectively. In the case of the STFT, time–frequency localization is fixed. Contrary to that, in the case of the DWT, the width of the time window

can be adapted with frequency. This adaptation provides better time localization in higher frequencies.

wavelets have been studied in [14,15]. The decom-

position of a signal into a basis of wavelet functions

implies the computation of the inner products between

the signal and the basis functions, leading to a set of

coefficients called wavelet coefficients. The signal can

consequently be reconstructed as a linear combination

of the basis functions weighted by the wavelet

coefficients. In order to obtain an accurate reconstruc-

tion of the signal, a sufficient number of coefficients

have to be computed.

The main characteristic of wavelets is the time–

frequency localization. In effect, time localization

means that most of the energy of the wavelet is

restricted to a finite time interval. Frequency

localization means that the Fourier transform is band

limited. Time and frequency energy concentrations are

restricted by the Heisenberg uncertainty principle.

This principle has a particularly important interpreta-

tion in quantum mechanics as an uncertainty with

respect to the position and momentum of a free

particle. A function f whose energy is well localized

in time and Fourier transform has energy concentrated

in a small frequency neighbourhood is restricted by

the Heisenberg uncertainty principle which constrains

the temporal and frequency variance:

s2t s

2v�

1

4(1)

This means that decreasing the deviation in frequency

(increasing the resolution) must result in an increase in

the deviation in time (decrease in resolution) and vice

versa. This is the fundamental weakness of the Fourier

transform. The boundary of the Heisenberg uncertainty

principle is reached if a Gaussian window is used [16].

The advantage of time–frequency localization is

that contrary to the short-time Fourier transforms, a

wavelet analysis varies the time–frequency aspect

ratio, producing good frequency localization at low

frequencies (long time windows), and good time

localization at high frequencies (short-time win-

dows). This produces segmentation, or tiling of the

time–frequency plane that is appropriate for most

physical signals, especially those of a transient nature.

This means that for a signal with rapid changes in

frequency at high frequencies and slow changes in

frequency at low frequencies, the wavelet transform

will give a better time–frequency representation of

the signal than the Fourier transform. The difference

between the short-time Fourier transform and the

wavelet transform is illustrated graphically in Fig. 2.

It must be said that in general no time–frequency

regions but rather time–scale regions are defined. The

time–scale expression has an equivalent time–

frequency expression, since wavelets which are well

localized around a non-zero frequency v0 at scale

s = 1 (i.e., the mother wavelet) have an inversely

proportional relationship between scale and fre-

quency.

4. Application of neural networks in EMG

analysis

4.1. Self-organizing maps

The self-organizing map (SOM) introduced in [17],

is a neural network (NN) that maps input signals (x)

from a high-dimensional space to grids of arbitrary


dimension, but 1D and 2D are more in use since the

visualization of SOM with high-dimensional grids is

problematic

4.2. Introduction to current techniques

The SEMG recorded during a voluntary muscle

contraction may be modeled as the spatio-temporal

superposition of the action potential trains of the

recruited motor units. Hence, modifications of the

motor unit action potential shape and of the motor unit

firing rate or muscle fiber propagation velocity

(MFPV) occurring during a muscle contraction cause

a variation of the power spectrum of the SEMG.

Several studies have been performed aiming at

discovering signs of fatigue and/or causes of fatigue.

Opinions agree that muscle fiber propagation velocity

(MFPV) decreases with fatigue and that EMG power

spectrum shifts during fatigue, mainly owing to a

slow-down of MFPV [18]. The main cause would be

increased lactate levels in fatigued muscle [19].

EMG signals from five subjects were collected

from trapezoid muscles under realistic dynamic

operation (varying force). Muscle fatigue can be

defined as reduction in the force generating capacity of

a muscle. This change in mechanical performance

capacity reflects in EMG changes. Typical changes

include increase in EMG amplitude and shift of

spectrum towards lower values.

Currently, the detection of muscle fatigue is based

on the joint frequency and amplitude analysis of the

Fig. 3. A graphical explanation of JASA (joint analysis of spectrum

and amplitude). The symbol E�A denotes the derivative of electrical

activity (EA) while denotes M�F the derivative of the median

frequency (MF). The subspace of fatigue is defined by an increase

in EA and a decrease in MF.

EMG [13]. The relevant technique is called JASA:

joint analysis of (EMG) spectrum (median fre-

quency) and amplitude (electric activity). JASA is a

necessary condition for comparing situations with

equal muscle force production. An illustration of how

fatigue is detected using JASA is shown in Fig. 3.

Such an analysis however is only valid under the

limiting assumption that only static muscle contrac-

tions occur. Under this limiting condition, which is

the working assumption of JASA, fatigue detection is

only possible when under constant force, the

amplitude of the EMG signal increases and the

median frequency decreases with time. In practice

only dynamic contractions occur so it is impossible to

detect fatigue using time–frequency analysis. All the

five subjects that took part in the experiments were

experiencing muscle fatigue but the presence of

fatigue could not be detected using JASA. An

indicative example is shown in Fig. 4. From Fig. 4

it is evident that after calculating regression lines, the

amplitude decreases while the mean and median

frequencies show no significant trend but a slight

tendency to increase. These effects are in fact the

opposite of what it would be expected under static

conditions. This result is expected since the main

assumption of JASA is that the force that is applied

from the subject is kept constant.

4.3. Wavelets and SOM for fatigue detection

For detecting muscle fatigue under dynamic

conditions more advanced techniques based on

wavelets and neural networks are proposed in the

current paper. Conventional techniques of frequency

and amplitude analysis do not work in this case. The

signal amplitude for all subjects is falling near the end

of the test showing that the subjects are putting less

force. This indicates that while fatigue is clearly

present according to the subjects’ own experience it

cannot be detected due to the condition of constant

force being not applicable.

Wavelet based decomposition is used to isolate

coordinated trapezoid muscle activity from five

volunteer subjects related to the same repeated

movement of turning the steering wheel during

driving a car. The type of wavelet used was the

Daubechies-5 wavelet [14] and the signal was

decomposed into 10 levels. By observing the trend


Fig. 4. An example is shown where application of JASA on a subject that was experiencing fatigue does not lead to positive results. The term EA

denotes electrical activity, MNF stands for mean frequency and MDF stands for median frequency. It can be observed that EA is decreasing while

MNF and MDF show no significant trend. Similar results were obtained from all subjects.

of the approximation coefficients in Fig. 5 it is

evident that the amplitude of the approximation

coefficients follows closely the fatigue development

of the test subject. In all subjects the filtered

coefficients decrease when the subject rests for half

an hour and then rise again but from a higher starting

level. The proposed method for detecting fatigue by

observing the behaviour of the approximation wavelet

coefficients can be automated by using neural

networks.

A SOM with 34 neurons arranged in a 17 � 2

configuration has been used to visualize the variation

of the approximation wavelet coefficients and aid the

detection of fatigue. Similar on-line neural network

training approaches in medical diagnosis have been

proposed in [20]. The SOM has been used in robotics

for the unsupervised recognition of environments by

performing clustering of action sequences [21]. The

creation of the input dataset for the SOM has been

performed by creating vectors of two elements with a

sliding window running over the approximation

coefficients. Therefore, two consecutive values of the

approximation coefficients have been used as input

vector to the SOM. The activated neurons at a certain

time instant are associated with the trend of the

approximation coefficients. Due to the topology

preserving capability of the learning algorithm that is

imposed through the Gaussian lateral excitation

profile, the activated neurons in the case of presence

of fatigue will tend to be in a clearly defined region of

the map. Because of the particular shape of the

proposed SOM configuration the fatigue associated

regions will tend to be even more clearly defined. The

trajectory plots in Fig. 5d provide a visualization

of the clustering tendency of the activations since

they connect subsequently activated neurons. The

observed clusters which are indicated by the

ellipsoids show that the 2D SOM separates EMG

signatures appearing at periods associated with

fatigue from periods where the muscles are fresh,

i.e. at the beginning and after the 30 min resting

period. This approach is similar to the one referred in

[22] where a SOM has been used to diagnose faulty

conditions during motor operation by mapping them

onto corresponding output clusters. The map is able

to detect the resting period in the middle of the test

period in which the muscles have recovered

temporarily. The procedure is repeatable for different

subjects. However, the same trained SOM cannot be

used from subject to subject.


Fig. 5. (a) Original EMG signal from the trapezoid muscle sampled at 1000 Hz. (b) The Daubechies-5 wavelet decomposition vector containing

all the 10 detail levels and the approximation level. (c) The approximation coefficients shown in detail. A trend is evident. (d) The 2D SOM has

34 neurons arranged in a 17 � 2 configuration. The subplots show the trajectory of SOM activations during different stages of the EMG signal

evolution. The short activity trajectories indicate the temporal sequence of activations of the SOM neurons for a quarter of the total signal

duration. The trajectory plots show that a 2D SOM separates EMG signatures appearing at periods associated with fatigue from periods where the

muscles are fresh, i.e. at the beginning and after the 30 min resting period. The rightmost sub-figure shows the SOM structure at the end of

learning. The codebook vectors are plotted.

5. Conclusions

Wavelet coefficients can be used as features for

identifying muscle fatigue. Flexible intelligent classi-

fication methods based on self-organizing maps have

beenused to identify car driver fatigue. The effective-

ness and performance of the proposed technique is

proven against currently used techniques that are not

able to detect fatigue under dynamic conditions. The

proposed techniques can be used to assess the presence

of fatigue and act against it at an early stage, possibly by

adjusting comfort related parameters that can affect the

presence of fatigue. Such a fatigue detection system can

form the core of a comfort assessment system aiming

at redesign of a driver seating system or on-line

adaptation of the parameters of such a seating system.

The latter can be envisaged as a predictive maintenance

system related to driving comfort. The system is

adaptable to different subjects and conditions since the

techniques used are not subject or workload regime

specific as indicated by the experiments.

References

[1] B.S. Hudgins, A new approach to multifunction myoelectric

control, Ph.D. Dissertation, University of New Brunswick, 1991.


[2] K. Engelhart, Signal representation for classification of the

transient myoelectric signal, Ph.D. Dissertation, University of

New Brunswick, 1998.

[3] J.V. Basmajian, C.V. De Luca, Muscles Alive, Williams &

Wilkins, Baltimore, MD, 1985.

[4] E.J. Kupa, S.H. Roy, S.C. Kandarian, C.V. De Luca, Effects of

muscle fiber type and size on EMG median frequency and

conduction velocity, J. Appl. Physiol. 79 (1) (1995) 23–32.

[5] M. Solomonow, C. Baten, J. Smit, R. Baratta, H. Hermens,

D’Ambrosia, H. Shoji, Electromyogram power spectra fre-

quencies associated with motor unit recruitment strategies, J.

Appl. Physiol. 68 (3) (1990) 1177–1185.

[6] G. Balestra, R. Knaflitz, R. Merletti, Stationarity of voluntary

and electrically elicited surface myoelectric signals, Electro-

physiol. Kinesiol. (1988) 275–278.

[7] R. Merletti, M. Knaflitz, C.J. De Luca, Electrically evoked

myoelectric signals, Crit. Rev. Biomed. Eng. 19 (4) (1992)

293–340.

[8] J.R. Potvin, L.R. Bent, Avalidation of techniques using surface

EMG signals from dynamic contractions to quantify muscle

fatigue during repetitive tasks, J. Electromyogr. Kinesiol. 7 (2)

(1997) 131–139.

[9] S. Karlsson, B. Gerdle, Mean frequency and signal amplitude

of the surface EMG of the quadriceps muscles increase with

increasing torque—a study using the continuous wavelet trans-

form, J. Electromyogr. Kinesiol. 11 (2001) 131–140.

[10] M. Knaflitz, P. Bonato, Time–frequency methods applied to

muscle fatigue assessment during dynamic contractions, J.

Electromyogr. Kinesiol. 9 (1999) 337–350.

[11] M.H. Pope, A. Aleksiev, N.D. Panagiotacopulos, J.S. Lee, D.G.

Wilder, K. Friesen, W. Stielau, V.K. Goel, Evaluation of low

back muscle surface EMG signals using wavelets, Clin. Bio-

mech. 15 (2000) 567–573.

[12] H.J. Hermens, F. Freriks, C. Disselhorst-Klug, G. Rau, Devel-

opment of recommendations for EMG sensors and sensor

placement procedures, J. Electromyogr. Kinesiol. 10 (2000)

361–374.

[13] A. Luttmann, M. Jager, K. Witscher, M. Vorgerd, M. Tegenth-

off, Fatigue-induced EMG changes during low-level

isometric contractions, in: H.J. Hermens, G. Rau, C. Dissel-

horst-Klug, B. Freriks (Eds.), in: Proceedings of the Third

General SENIAM (Surface EMG for Non Invasive Assessment

of Muscles) Workshop on Surface Electromyography: Appli-

cation Areas and Parameters, Aachen, Germany, May 1998,

Roessingh Research and Development, Enschede, 1998 , pp.

141–150.

[14] Daubechies, Orthonormal bases of compactly supported wave-

lets, Commun. Pure Appl. Math. 41 (1988) 909–996.

[15] Y. Meyer, Orthonormal wavelets, in: J.M. Combes, A. Gross-

man, P. Tchamitchian (Eds.), Wavelets Time–Frequency Meth-

ods and Phase–Space, Springer-Verlag, Berlin, 1989, pp. 21–

37.

[16] M.R. Rao, A.S. Bopardikar, Wavelet Transforms—Introduc-

tion to Theory and Applications, Addison-Wesley, 1998.

[17] T. Kohonen, Self-organized formation of topographically cor-

rect feature maps, Biol. Cybernet. 43 (1982) 59–69.

[18] N.K. Vøllestadt, Measurements of human muscle fatigue, J.

Neuro. Sci. Meth. 74 (1997) 219–227.

[19] J.R. Doud, J.M. Walsh, Muscle fatigue and muscle length

interaction, effect on the EMG frequency components, Elec-

tromyogr. Clin. Neurophysiol. 35 (1995) 331–339.

[20] G.D. Magoulas, V.P. Plagianakos, M.N. Vrahatis, Neural

network-based colonoscopic diagnosis using on-line learning

and differential evolution, Appl. Soft Comput. 4 (2004) 369–

379.

[21] S. Yamada, Recognizing environments from action sequences

using self-organizing maps, Appl. Soft Comput. 4 (2004) 35–

47.

[22] X.Z. Gao, S.J. Ovaska, Soft computing methods in motor fault

diagnosis, Appl. Soft Comput. 1 (2001) 73–81.

dynamic muscle fatigue detection using self-organizing maps

Documents