Neural

Networks with AR Model Coefﬁcients Applied to

the EMG Signal Classiﬁcation

Marek Kurzynski and Andrzej Wolczowski

Wroclaw University of Technology, Dept. of Systems and Computer Networks

Wyb. Wyspianskiego 27, 50-370 Wroclaw, Poland

Abstract. The paper presents a concept of hand movements recognition on the

basis of EMG signal analysis. Signal features are represented by coefﬁcient of

autoregressive (AR) model, and as classiﬁer the MLP and Adaline networks

are applied. The performance of the proposed method was experimentally com-

pared against four different classiﬁers using real datasets. The systems developed

achieved the highest overall classiﬁcation accuracies demonstrating the potential

of neural network classiﬁers based on AR coefﬁcients for recognition of EMG

signals.

1 Introduction

The activity of human organism is reﬂected in characteristic biosignals, which can be

measured and next can be applied to the control of the work of technical devices. Elec-

trical potentials accompanying skeleton muscles (called EMG signals) are an example

of such biosignals. They can be detected and registered through the skin and used to the

control of bio-prosthesis.

Although in the last decade many attempts have been made to determine the hand

movements on the base of EMG signal analysis ([5, 10, 11, 13]), the reliable recognition

of kind of grasp is still a hard problem. The difﬁculty increases along with the prosthesis

dexterity (prosthesis movement repertoire), therefore it is still a need for research in

developing EMG signal recognition.

The paper presents a concept of recognition of hand movements (type of grasp) on

the base of EMG signal analysis. Signal features are represented by autoregressive (AR)

model coefﬁcients, and as classiﬁer the MLP and Adaline network are applied.

The performances of proposed classiﬁcation systems were compared against four

(statistical (Bayes, kernel), fuzzy and k-nearest neighbours) classiﬁers using real datasets.

For the purpose of experimental investigations a special measurement stand was elabo-

rated which allow us synchronous recording the image of the moving hand and multi-

channel registration of EMG signals.

The paper is divided into three sections and organized as follows. In section 2 we

provide an insight into the analysis of EMG signals which is the basis for the recognition

of grasps. In Section 3 computer experiments on real data are described and their results

are discussed.

Kurzynski M. and Wolczowski A. (2010).

Neural Networks with AR Model Coefﬁcients Applied to the EMG Signal Classiﬁcation.

In Proceedings of the 6th International Workshop on Artiﬁcial Neural Networks and Intelligent Information Processing, pages 81-86

 SciTePress

2 EMG Signal Analysis

The recognition of hand movement on the basis of the myopotentials comprises three

stages [13]: (1) the acquisition of the EMG signal; (2) extraction of the features differ-

entiating the movements; (3) classiﬁcation of the signal.

Each stage has an inﬂuence on the quality of the whole process, i.e. reliability of

the grasping movement recognition.

2.1 EMG Signal Acquisition – the Measurement Stand

The block diagram of the designed measurement stand for EMG signal acquisition

and identiﬁcation of the relation between the hand movement and simultaneously cre-

ated myopotentials, is presented in Fig. 1. The stand includes: (1) a video camera for

recording the image of the moving hand; (2) specially designed 8-channel EMG sig-

nals measuring circuit (Bagnoli Desktop EMG System, DelSys); (3) the PC computer

recording the results of the acquisition, equipped with high ﬁdelity measurement board,

containing 8 independent A/D converters (24 bits per channel) and USB port for USB

video camera; (4) an application for synchronous recording of the video and EMG data

streams and their analysis.

Fig.1. The measurement system for identifying the relation between the hand movement and

EMG signals.

2.2 Features Extraction

The extraction of features consists in determining such parameters that best differentiate

the received signals for the sake of movement recognition. The extraction of features

can be accomplished using various techniques including signal amplitude, EMG fre-

quency characteristic and power spectrum analyzed by fast Fourier transform (FFT)

method [6], the integral of the absolute value (IAV) and zero crossing signal [6, 7], time

and frequency histograms [11], among others. In this paper it is proposed an efﬁcient

method to determine the input features based on autoregressive (AR) model.

The AR model belong to a group of linear prediction methods that attempt to predict

an value y

of a time series of data {y

}based on the previous values (y

n−1

, y

n−2

, . . .).

Deriving the linear prediction model involves determining the coefﬁecients (a

, a

, . . . , a

)

in the equation:

ˆy

k=1

n−k

, (1)

where ˆy

is the estimated value of signal in a time n, a

are the AR coefﬁcients and p

is the order of AR model.

Several estimators of AR coefﬁcients are well known in the ﬁeld of signal process-

ing. We chose the Burg algorithm because of its many remarkable advantages (it does

not apply window data, minimizes forward and backward prediction errors, gives high

resolution for short data records, always produces a stable model) [9]. The Burg algo-

rithm estimates the AR coefﬁcients by ﬁtting an autoregressive linear prediction ﬁlter

model of a given order to the signal. Consequently, the Burg algorithm determines for

each channel the set of p AR coefﬁcients, which create the feature vector describing the

EMG signal (r is the number of channels):

x = [a

, a

, . . . , a

, a

, . . . , a

, a

, . . . , a

]. (2)

2.3 Classiﬁcation

Two types of artiﬁcial feedforward neural networks were used in this study for classiﬁ-

cation of EMG signal: multilayer perceptron (MLP) and Adaline network.

1. The MLP Classiﬁer (MLP). The network consists of the input, hidden, and output

neuron layers. The input layer plays the role of a data buffer so that the data are

normalized to belong to the [0, 1] range. There have been various numbers of input

and hidden layer neurons, depending on the actual quantities of data. The number

of output layer neurons is equal to the number of classes (types of grasps). The ﬁnal

classiﬁcation is made according to the maximum rule. Both the hidden and output

layer neurons have the sigmoid transition function. Neurons of the successive layers

are connected on the each-to-each basis. In the experiments, the corresponding lay-

ers were trained by means of the error back propagation method with momentum

term.

2. The Adaline Classiﬁer (ADA). The single layer neural network that contains neu-

rons with (positive) linear transfer functions. As previously, the number of neurons

is equal to the number of classes and the ﬁnal classiﬁcation is made according to

the maximum rule. In the experiments the Adaline network was trained by Widrow

and Hoff learning procedure, also known as the delta rule.

3 Experiments

The proposed methods of EMG signal classiﬁcation based on ANN techniques were

experimentally tested and their performances were compared against the four following

pattern recognition techniques: (1) Naive Bayes method (NB) [4]; (2) Parzen classiﬁer

with the Gaussian kernel and the optimal smoothing parameter (PAR) [4]; (3) 5-nearest

neighbours classiﬁer (5-NN) [4] and (4) classiﬁer based on fuzzy relations (FR) [13].

3.1 Experimental Setup

The experiments were carried out on healthy persons. The electrodes, connected to

the respective measuring channels, were put over the following forearm muscles: (1)

the extensor muscle of the ﬁngers, (2) the radial extensor of the wrist, short, (3) the

superﬁcial ﬂexor muscle of the ﬁngers, (4) the ulnar ﬂexor muscle of the wrist, (5) the

extensor muscle of the thumb, short, and (6) the ﬂexor muscle of the thumb, long (see

Fig. 2).

The experiments were conducted in MATLAB using PRTools and NN Toolbox.

Fig.2. The layout of the electrodes on the forearm.

In experiments ﬁve different types of grasps (classes) presented in Fig. 3 were cho-

sen for recognition from the set deﬁned by Schlesinger ([8]): 1) palmar, 2) tip, 3-4)

cylindrical and cylindrical tight, 5) spherical. Our choice is deliberate one and results

from the fact that the control functions of simple bioprosthesis are hand closing/opening

and wrist pronantion/supination, however for the dexterous hand these functions differ

depending on grasped object [2].

1 2 3-4 5

Fig.3. Types of grasps recognized in experiment.

Each measurement lasted 2.5 s and was preceded with a 10 s break. In that way for

the single grasp movements the discrete signals were obtained each of a size of 2500

samples (1 kHz sampling frequency) × 6 channels, together with the video sequences

related to them, that picture the movement types (classes). The 300 measurements (60

measurements for each grasp type (class)) were created, and next gathered EMG signals

were subjected to the feature extraction procedure for different orders of AR model p

equal to 2, 3, 5, 7 and 10. Consequently, we got 5 datasets, each containing 300 pat-

terns described by 12, 18, 30, 42 and 60 features, respectively. The training and testing

datasets were extracted from each dataset using two-fold cross-validation method.

The ADA classiﬁer comprised 5 neurons which inputs number was equal to the

number of features (different for each dataset). Similarly, the MLP classiﬁer comprises

5 neurons in the output layer and the number of input neurons (hidden neurons) was

equal to 12 (8), 18 (10), 30 (15), 42 (20) and 60 (30) for the successive datasets, respec-

tively. The number of epochs in the learning procedure for the both ANN classiﬁers was

equal to 200.

3.2 Results and Discussion

Classiﬁcation accuracies (i.e. the percentage of correctly classiﬁed objects) for meth-

ods tested are listed in Table 1. The accuracies are average values obtained over 10 runs

(5 replications of two-fold cross validation). Statistical differences between the perfor-

mances of the ADA, MLP classiﬁcation methods and the four classiﬁers were evaluated

using Dietterich’s 5x2cv test [3]. The level of p < 0.05 was considered statistically sig-

niﬁcant. In Table 1, statistically signiﬁcant differences are given under the classiﬁcation

accuracies as indices of the method evaluated, e.g. for the dataset with p = 5 the MLP

classiﬁer produced statistically different classiﬁcation accuracies from the NB, 5-NN

and FR methods. The row ”‘Mean”’ contains results averaged over all datasets.

Table 1. Classiﬁcation accuracies of classiﬁers compared in the experiment (description in the

text). The best score for each dataset is highlighted.

Classiﬁer / Mean (SD) accuracy [%]

AR order NB

PAR

5-NN

ADA MLP

p = 2 73.2(5.2) 82.8(2.8) 86.5(4.2) 72.4(6.3) 84.2(3.1) 87.7(2.2)

1,4 1,2,4

p = 3 79.3(4.6) 90.2(1.9) 94.2(1.6) 80.6(2.4) 91.0(1.3) 93.6(1.2)

1,4 1,2,4

p = 5 81.5(2.2) 97.6(0.3) 85.3(1.4) 83.5(3.6) 94.8(1.1) 97.4(0.6)

1,4 1,3,4

p = 7 80.4(2.5) 98.2(0.7) 95.8(0.9) 87.2(1.3) 98.3(0.2) 100 (0.0)

1,3,4 1,3,4

p = 10 82.7(2.3) 98.1(0.5) 96.9(0.4) 91.5(1.2) 100(0.0) 100(0.0)

1,2,3,4 1,2,3,4

Mean 79.4(3.4) 93.4(1.2) 93.8(1.7) 83.0(2.9) 93.7(1.1) 95.7(0.8)

1,4 1,2,3,4

The MLP classiﬁer achieved the highest overall classiﬁcation accuracy averaged

over all datasets – it outperformed the NB, PAR, 3-NN and FR classiﬁers by 16.3%,

2.3 %, 1.9% and 12.7% on average, respectively. The ADA neural network that was

the third best-scoring classiﬁer, outperformed the NB, PAR and FR systems by 14.3%,

0.3% and 10.7% on average, respectively. The both ANN-based classiﬁers produced

statistically signiﬁcant higher scores in 29 out of 40 cases (5 datasets × 4 classiﬁers

× 2 systems developed). The ADA and MLP classiﬁers also achieved the highest clas-

siﬁcation accuracy (i.e. 100%) when the datasets with 42 and 60 features were used.

Furthermore, they produced the best stability (the SD values of 1.1% and 0.8% aver-

aged over all datasets), followed by the PAR classiﬁer (1.2%). Results obtained indicate,

that proposed methods of grasping movement recognition based on the AR model as an

EMG signal feature extraction procedure, produced accurate and reliable decisions, es-

pecially in the cases with greater number of features.

References

1. Boostani R., Moradi M., Evaluation of the forearm EMG signal features for the control of a

prosthetic hand, Physiological Measurement 24 (2003), 309-319

2. De Luca C., J., Adam R., et al., Decomposition of surface EMG signals, Journal of Neu-

ropsychology 96, (2006) 1646-1657

3. Dietterich T.G., Approximate statistical tests for comparing supervised classiﬁcation learn-

ing algorithms, Neural Computing 10 (1998) 1895-1923

4. Duda R., Hart P., Stork D., Pattern Classiﬁcation, John Wiley (2000)

5. Ferguson S., Dunlop G., Grasp recognition from myoelectric signals, Proc. 2002 Aus-

tralasian Conference on Robotics and Automation (2002) 83-87

6. Kuribayachi K., Okimura K. and Taniquichi T., A discrimination system using neural net-

work for EMG-controlled prostheses, Proc. Int. Conf. IEEE Robot and Human Communica-

tion (1992) 63-68

7. Khoshaba T., Badie K., and Hashemi R., EMG pattern classiﬁcation based on back propaga-

tion neural network for prosthesis control, Proc. Ann. Int. Conf. IEEE Eng. Med. Biol. Soc.

(1990) 1474-1475

8. Reaz M., Hussain M., Techniques of EMG signal analysis: detection, processing, classiﬁca-

tion and applications, Biological Procedures Online 8 (2006) 11-35

9. Schloegl A., A comparison of multivariate autoregressive estimators, Signal Processing 9

(2006) 2426-2429

10. Wang, G., Zhiguo, Y., Xiao, H., et al., Classiﬁcation of surface EMG signals using harmonic

wavelet packet transform, Physiol. Measurement 27 (2006) 1255-1267

11. Wojtczak P., Amaral T., Dias O., Hand movement recognition based on biosignal analysis,

Engineering Applications of Artiﬁcial Intelligence 22 (2009) 608-615

12. Wolczowski A., Krysztoforski K., Control-measurement circuit of myoelectric prosthesis

hand, Acta Bioengineering and Biomechanics 4 (Suppl.) (2002) 576-578

13. Wolczowski A., Kurzynski M., Human-machine interface in bioprosthesis control using

EMG signal classiﬁcation, Expert Systems 27, (2010) 53-70