Fuzzy Systems and Neural Networks Methods to Identify Hand
and Finger Movements Using Surface EMG signals
1123ABDELHAFID ZEGHBIB; FRANK PALIS; GEORGI TSENOV; NICOLA SHOYLEV; 2VALERI MLADENOV 1 Otto-von-Guericke-Universität Magdeburg, Institut für Elektrische Energiesysteme
39016 Magdeburg, GERMANY
2 Department of Theoretical Electrical Engineering, Technical University of Sofia,
8 Kl. Ohridski St., BG-1000 Sofia, BULGARIA 3 Department of Electrical Engineering, University of Chemical Technology and Metallurgy,
8 Kl. Ohridski St., BG-1756 Sofia, BULGARIA
Abstract: - With help of exploitation of myoelectric signals, amputee persons can have a chance to improve their life with myoelectric prosthesis which are able to function with the amputee’s muscle movements. The
myoelectric signal (MES) is the electrical manifestation of muscular contraction. This signal recorded at the surface of the skin of the forearm has been exploited to provide the recognition of the movement of the hand and finger Movements of healthy subject. The objective of the paper is first to describe the identification procedure, based on EMG patterns of forearm activity using Fuzzy logic and Neural Networks methods. Second to show the advantage of using features in Time-frequency domain, in comparison to those in time domain. Suitable features in time-frequency domain give high classification rates. Third is to compare between different intelligent computational methods of identification, which are used in this work: Multi-Layer Perceptron (MLP), Radial Basis Function Networks (RBF) and Learning Vector Quantization network (LVQ) as supervised methods and fuzzy Subtractive Clustering (FSC) as unsupervised method.
Key-Words: - EMG signals, Hand and finger Movements Identification, Neural Networks, Fuzzy Logic, Subtractive clustering, Feature extraction, Short Time Fourier Transform.
unsupervised methods, like Self Organizing Map, 1 Introduction
Fuzzy Subtractive Clustering and Competitive MES classification is one of the most difficult
Layer. Some features in time domain and time-pattern recognition systems because there usually
frequency domain are extracted from raw EMG exist large variations in Electromyograph (EMG)
signal and used for identification of movements features. The EMG signal has been used as a tool to
with help of the above mentioned intelligent provide advanced man-machine interfaces ,
computational methods. In practice, determination rehabilitation of the handicapped people, functional
of relevant features is very difficult. The aim of this electrical stimulation devices (FES)  and control
paper is to distinguish between three finger commands for limb prostheses [3,4].
movements (thumb, index and middle) and hand The classification problem may be divided into
closing. three steps: signal presentation, feature extraction
and pattern recognition. It is shown in this paper that classification performance of hand and finger
2 EMG Signal Preprocessing movements depends significantly upon feature
Surface muscle activity signals cannot be analysed extraction, which is very important to improve
using classical methods, since they are non-considerably the accuracy of classification. Many
stationary and have complex time-frequency researches proposed several EMG features for
characteristics. EMG signals, fig 1, which are classification that showed good performance  
evolving in time in an unpredictable way (like a . These identification methods belong to two
speech signal or an EMG signal) require the notion categories, first are supervised methods, like Multi-
of frequency analysis for each local time. Although Layer Perceptron, Radial Basis Networks, and
frequency-domain representations such as the power Learning Vector Quantization network, Second are
spectrum of a signal often show useful information, subject’s arm is given in fig. 2, from the input
these representations don’t show how the frequency feature space, the classifier must be able to classify content of a signal evolves over time. Time-the three output classes exploiting the EMG signals Frequency Analysis can identify not only the measurements.
frequency content of a signal, but also how that For each channel the signal was acquired using a content evolves over time. single bipolar surface electrode pair. A differential
amplifier with an isolated input is used. The signal
was sampled at a rate of 4Khz using A/D board in
an IBM PC/AT compatible microcomputer; this
algorithm is developed with MATLAB 6 and is
performed in a PC-based off-line process. The
human subject was asked to produce a number of
continuous movements, 34 single contraction
periods are separated from the corresponding sets of
continuous movements. Each single contraction
period extracted from the raw signal by determined
threshold is analysed with Short time Fourier
Transform (STFT), which gives a measure of time and frequency information, fig 3 and fig 4, for small Fig. 1. Measured raw EMG signal from channel 2 segments of a signal. For the two channels we (extensor digitorum) muscle, and his absolute value. prepare some EMG training and test data, each class
has 17 training and 17 test patterns. The four classes There are a number of different methods available
labelled 1, 2, 3 and 4 have 68 train-samples and 68 for Time Frequency Analysis. Each type shows a
test-samples. different time-frequency representation. The Short
Time Fourier Transform (STFT), which is used in
this paper, is the simplest TFA method and the
easiest to compute.
Fig. 3. Thumb: 400 ms EMG Signal analyzed with STFT
Fig. 2, EMG training and test patterns recorded using
two pairs of electrodes in Max Planck Institute
laboratory in Magdeburg, Germany.
Four types of finger and hand movements to be
classified are selected: thumb, pointer, middle and
hand close. The placement of EMG surface
electrodes on muscle groups is important to have
more information about each movement. Two EMG
surface electrodes are placed on two muscle groups,
palnaris longus (channel_1) and extensor digitorum
(channel_2), the locations of electrodes on the Fig. 4. Contour presentation for the same above signal
5.1 Fuzzy system initialised with Subtractive 4 EMG Feature extraction
Clustering method The problem of classification is the partitioning of
The subtractive clustering algorithm was proposed the feature space, into regions (classes). Relevant
by Chiu (1994). It estimates the number of clusters features will lead to high and accurate classification
and the cluster centres in a set of data by an iterative rates. In the time domain, four features are extracted:
procedure. The clusters obtained, with iterative Mean absolute value (MAV), Variance (VAR),
optimisation-based clustering methods fuzzy c-Waveform length (WL), and Median Value (Med).
Nmeans (fcm), are used to initialise the fuzzy sets, for 1MAV，emg, (1) model identification method ANFIS. As initial ，kN，1kmodel the first order Takagi-Sugeno (T. S.) model N12is used. We test some values of cluster radius, VAR，emg, (2) ，kN？1 For these values we test which are: 0.3, 0.4 and 0.7.k，1
N？1our model during 20 epochs using the four-extracted
WL，emg？emg (3) ，k？1ktime features, see table 1 for results. k，1 Table 1: accuracy and number of correct instances for thWhere emg is the k sample data, which is N keach class with extracted time features.
samples in length. The following Table shows us Classification accuracy (test data) % the worse classification with these features cluster Thumb Pointer Middle HC average
In time-frequency domain using STFT (Short Time 0.3 52.94 23.52 0 5.88 20.58 Fourier Transform), the Hannaford’s moments of 0.4 58.82 17.64 0 17.64 23.52 first and second order are extracted as time-0.7 88.23 11.76 0 5.88 26.47 frequency features with the dominant frequency Cluster # of correct classified instances /17
radius value FDV, which present the frequency value of
Thumb Pointer Middle HC Total maximum amplitude obtained from spectral analysis.
0.3 9 4 0 1 14 The nth moment of the frequency distribution at
0.4 10 3 0 3 16 time t is defined as:
0.7 15 2 0 1 18 n;；;；Mt，！STFTt,k (4) ，nk k
In the same way we test, table 2, our features ！n: order, t: time, : frequency.
extracted with help of time –frequency analysis
5 Methods and results Table 2: accuracy and number of correct instances for There is a large set of neural networks and fuzzy each class with extracted time-freq features. logic methods in the literature addressing Classification accuracy (test data) % identification problems. Some of these methods are Cluster Thumb Pointer Middle HC average applied, using different features, to identify our four radius fingers and hand movements and to compare their 0.3 82.35 70.58 76.47 82.35 77.94 performances. These methods in case of supervised 0.4 88.23 64.70 76.47 82.35 77.94 learning employ optimisation techniques to 0.6 94.11 76.47 100 94.11 91.17 processes the inputs and compare their resulting 0.7 94.11 76.47 100 82.35 88.23 outputs against the desired outputs. Errors are then
calculated, causing the system to adjust the Cluster # of correct classified instances /17 parameters. In case of unsupervised learning, radius
training algorithms attempt to locate clusters in the Thumb Pointer Middle HC Total
0.3 14 12 13 14 53/68 input data, which approximate the distribution of
0.4 15 11 13 14 53/68 the data. More details about clustering data in
0.6 16 13 17 16 62/68 [Kaufman & Rousseeuw, 1990].
0.7 16 13 17 14 60/68
5.2 Neural Networks Methods
Various ANN based models were applied to identify We test our MLP Network with different number of these different four movement classes. Hykin neurons in hidden layer: 5, 10, 20 and 50 neurons publishes a comprehensive foundation for the study for four extracted time features, (IEMG, WL, VAR of Neural Networks . There are many functions and Med, see table 4. It’s obvious in table 4 that and variables to be determined for each neural there is no effect of the number of neurons in network model. The four neural networks methods hidden layer on the rate of classification. The are trained during 100 epochs. increasing in number of neurons in hidden layer
doesn’t enhance the accuracy.
In case of extracted time-frequency features, we 5.2.1 Multi Layer Perceptron (MLP)
This network is used in many different types of obtained the results, which are resumed in table 5. applications. This architecture has a large class of
Table 5: network types with many different topologies and
Classification accuracy (test data) % training methods. The number of neurons in the
# of Thumb Pointer Middle HC average only hidden layer is determined based on their
neurons performance in training process. For the one-neuron 5 100 52.94 94.11 94.11 85.29 output-layer we use log sigmoid transfer function 10 100 76.47 100 94.11 92.64 “logsig”, which gives an output in the range of 0 to 20 88.23 41.17 94.11 94.11 79.41 1. Our output range between 0 and 1 will be divided 50 94.11 64.70 70.58 94.11 80.88 in four ranges, since we have four classes to be identified, see Table 3. # of # of correct classified instances /17 neurons Table 3: Thumb Pointer Middle HC Total Classes Target Output Type of 5 17 9 16 16 58/68 output range Movement 10 17 13 17 16 63/68 Classe 1 0.125 0 – 0.25 Thumb 20 15 7 16 16 54/68 Classe 2 0.375 0.25 – 0.5 Pointer 50 16 11 12 16 55/68 Classe 3 0.625 0.5 – 0.75 Middle Classe 4 0.875 0.75 - 1 Hand close 5.2.2 Radial basis Functions (RBF) The classification results are summarized in table4. The RBF Network is a one hidden layer neural Network with several forms of radial basis Table 4: Rate of classification and correct classified activation functions, like Gaussian function. We use instances (extracted time features) for each class and
the method, which creates neurons one at a time. In average value with four different hidden layer neurons
each iteration the input vector is used to create a number.
Classification accuracy (test data) % new neuron. The error of the new network is
# of Thumb Pointer Middle HC average checked, and if it is not low enough the next neuron
neurons is added. This procedure is repeated until the error
5 17.64 35.29 17.64 64.70 33.82 goal is met, or the maximum number of neurons is
10 35.29 11.76 29.41 29.41 26.47 reached. The output layer is linear and the rate of
20 52.94 0 5.88 41.17 25 classification is determined by the spread of the 50 41.17 23.52 5.88 52.94 30.88 hidden unit. We give many values between 0.5 and 2.5 with a step of 0.2 to find the optimal value of # of # of correct classified instances /17 spread, which is in our application equal to 0.6 and neurons 0.7, see fig. 5. Thumb Pointer Middle HC Total
5 3 6 3 11 23/68
10 6 2 5 5 18/68
20 9 0 1 7 17/68
50 7 4 1 9 21/68
5.2.3 Learning Vector Quantization (LVQ)
Learning Vector Quantization [9, 10] networks can
classify, faster than other neural network techniques
like Back Propagation, any set of input vectors; not
only linearly separable sets of input vectors. Its
architecture resembles to that of unsupervised
competitive learning network, except that each
output is assigned to a target class and works in two
steps. First it uses an unsupervised data clustering
method to locate several clusters. Second it
optimises the cluster centres. The number of clusters can be specified a priori or determined via Fig. 5. Average accuracy according to spread’s values cluster techniques. It is able to reduce large data sets to a smaller number of codebook vectors (cluster For this value of spread we test our network using centres) suitable for data compressing. the four extracted time features, see table 6 for LVQ network used in this work has 4 neurons in the results. first competitive layer and one neuron for each class in the second linear layer. In comparison with Table 6: accuracy and number of correct instances for previous methods, LVQ needs only 30 to 40 epochs each class with extracted time features.
to converge, see table 8 and 9. Thumb Pointer Middle HC average
Acuracy 29.41 11.76 23.52 76.47 35.29
Table 8: Rate of classification (time extracted features) (Test) %
for each class and average value.
Classification accuracy (test data) % # of correct classified instances /17
competitive Thumb PointerMiddle HC average Thumb Pointer Middle HC Total
neurons Correct 5 2 4 13 24 / 68
6 76.47 0 5.88 41.17 30.88 (# / 17)
12 64.70 0 11.76 58.82 33.82
28 70.58 0 5.88 64.70 35.29 With the same method we test the classification
accuracy for the extracted time-frequency features with two values of spread: 0.7 and 1.5, see table 7. Table 9: Rate of classification (time-frequency extracted features) for each class and average value. Table 7: accuracy and number of correct instances Classification accuracy (test data) % (extracted features in time-frequency analysis domain) competitive ThumbPointer MiddleHC average for each class and average value with different values of neurons spread. 6 100 64.70 94.11 88.23 86.76 Classification accuracy (test data) % 12 100 64.70 94.11 88.23 86.76 Spread Thumb Pointer Middle HC average 28 100 64.70 100 88.23 88.23 value
0.7 88.23 88.23 76.47 82.35 83.82
1.5 76.47 70.58 64.70 76.47 72.05
6 Methods comparison
Spread # of correct classified instances /17 To resume the results obtained with the features value extracted in time-frequency analysis, we present for
Thumb Pointer Middle HC Total all methods their classification rate according to 0.7 15 15 13 14 57/68 some determined parameters. In fig. 6 four first 1.5 13 12 11 13 49/68 values of classification accuracy, with MLP method,
are presented according to number of neurons used
We get the best rate classification for Gaussian in hidden layer, which are: 5, 10, 20 and 50 neurons.
functions with spread value of 0.7 by extracted The second four values group given by
features in time-frequency analysis domain. unsupervised fuzzy subtractive clustering method corresponding to the different values of cluster
radius, 0.3, 0.4, 0.6 and 0.7. The two classification recognition as a control command of man–
accuracy values obtained with RBF method machine interface”. Med Eng Phys. 1996 corresponding to 0.7 and 1.5 spread parameter Oct;18(7):529-37. PMID: 8892237 [PubMed –
values. Finally LVQ method shows three rate of indexed for MEDLINE].
classification, corresponding to three values of  Hefftner G, Jaros G, “The electromyogram number of neurons (6, 12 and 28) in competitive (EMG) as a control signal for functional layer. neuromuscular, Stimulation, part I:
Autoregressive modeling as a mean of EMG
signature discrimination”. IEEE Trans Biomed
Eng 1988, 35(6), pp. 228–35.
 Zardoshti-Kermani M, Wheeler BC, Badie K,
Hashemi RM, “EMG feature evaluation for
movement, control of upper extremity
prostheses”. IEEE Trans Rehabil Eng
1995;Vol.3, No.4, pp. 324–333.
 C. Bonivento, A. Davalli, C. Fantuzzi, R.
Sacchetti, S. Terenzi, “Automatic tuning of
myoelectric prostheses” Journal of Rehabilitation Research and Development Vol. Fig. 6. Classification accuracy comparison between four 35 No. 3, July 1998, pp. 294-304. different methods using extracted features in time 5] K.Englehart, B. Hudgin, and P. A. Parker, “A [frequency domain analysis.
wevelet based continuous classification scheme for
multifunction myoelectric control”, IEEE
Transactions on Biomedical Engineering, vol. 7. Conclusion 48, no. 3, pp. 302-311, Mars 2001. As conclusion, first we can conclude that the  Z. K. Mahyar, W. C. Bruce, B. Kambiz, and M. identification methods cannot and don’t help to H. Reza, EMG feature evaluation for perform accuracy classification if the feature movement control of Upper extremity measures selected are not a relevant features. For prostheses”, IEEE Transactions on this reason the determination of a complete set of Rehabilitation Engineering, Dec 1995. discriminatory features is very important.  B. Hudgins, P. Parker and R. N. Scott, “New Both methods Neural Networks and Fuzzy Logic strategy for multi-function myoelectric are showed clearly that the selection of the feature control ”IEEE Transactions on Biomedical are of great importance to enhance the recognition Engineering, vol. 40, no. 1, pp. 82-94, Jan. rate of myoelectric patterns for the four movements, 1993. hand close and three finger (Thumb, pointer and  S. Haykin, Neural Networks: A Comprehensive middle). In this case the use of time-frequency Foundation, (2nd Edition 1998), ISBN: domain as features extraction domain is necessary 0132733501. Prentice Hall, 1998. to perform the identification.  T. Kohonen. Self organization and associate memory. Springer-Verlag, London, 3rd edition, 1989. Acknowledgement  T. Kohoenen. Improved versions of learning This work is supported by the DAAD PPP-program vector quantization. In International Joint 2005-2006 between Otto-von-Guericke-Universität conference on Neural Networks. Vol:1, p:545-Magdeburg, Germany, and University of Chemical 550, San Diego, 1990. Technology and Metallurgy, Sofia, Bulgaria.
 Chang G-C, Kang W-J, Luh J-J, Cheng C-K,
Lai J-S, Chen J-JJ, Kuo T-S, “Real-time
implementation of electromyogram pattern