Biceps Activity EMG Pattern Recognition Using Neural Networks

Transcription

1 Biceps Activity EMG Pattern Recognition Using eural etworks K. Sundaraj University Malaysia Perlis (UniMAP) School of Mechatronic Engineering 0600 Jejawi - Perlis MALAYSIA kenneth@unimap.edu.my Abstract: - This paper presents a study of EMG signals pattern on an activated muscle. The main rationale is that the pattern of the EMG signal produced may differ depending on the activity of the muscle movement. Therefore, the purpose of this study was to demonstrate the effectiveness of the neural network on recognizing the pattern of certain activities evoked by muscle. Experiments were carried out on a selected muscle. Five subjects were asked to perform several series of voluntary movement with respect to the muscle concerned. From the EMG data obtained, four statistical features are computed and are applied to a feed-forward back-propagation neural network. Overall, the network can discriminate different EMG signal patterns with a successful rate of up to 88%. Key-Words: - EMG, eural etworks, Pattern Recognition Introduction Electromyography (EMG) is one of the most robust non-invasive physiological measures of muscle contraction. Basically, EMG is a complicated signal which is controlled by the nervous system and is dependent on the anatomical and physiological properties of muscles [1]. Figure 1 shows a neuronal action of a muscle innervated by the motor neuron originating from the spinal cord [4], [6]. Studies have shown that even paralyzed people can produce discernible EMG signal through self-effort []. In the most general form, EMG has been used to evaluate muscle activity for function, control and learning [3]. The analysis of EMG signal has utilized signal processing and artificial intelligent skills in order to develop signal classification. These systems will facilitate human-computer interaction and could be further integrated to various intelligent systems and rehabilitation machines such as intelligent prostheses and medical robots to help the disabled unfortunates to lead a way of life that will create opportunities to live with dignity, peace and longer life. In this paper, we present a study of a classification system using back-propagation neural network. This supervised learning paradigm is well applicable to pattern recognition (classification) and regression (function approximation). From the EMG pattern generated by a group of neurologically intact subjects, we aim to study the pattern of EMG responses elicited through their voluntary contraction. Besides this, we are interested to study the discrimination rate of the system to classify EMG signal pattern accordingly to the specific muscle activity elicited. Furthermore, we would like to study the effectiveness of using statistical features in classifying muscle activities. Fig. 1. Innervated muscle fiber by motor neuron. Pearson Education [7] ISS: ISB:

2 Data Acquisition Five subjects ( males and 3 females) aged between 4 to 8 years (mean weight: 59.3kg and mean height: 160.4cm) took part in the study. Subjects gave their informed consent after hearing a clear explanation of the objectives and procedures of the study. Firstly, subjects are asked to remove any watch and jewelry from the wrist. Then, subjects skins are cleaned to decrease the resistance of the outer layer of skin. A differential surface electrode is placed over the belly of the Biceps Brachii muscle, as shown in Figure. This muscle was preferred because it is easily recognized and accessible when placing the electrode while the reference electrode is placed at bony prominence around the wrist where the area is electrically neutral. During the experiment, subjects are requested to perform several muscle tasks against a relaxed background through voluntary contraction. Fig. 3. Hardware connection for the proposed system. Feature Extraction Important and significant features have been extracted from the EMG raw data for pattern classification purposes. In this study, statistical features are chosen for its uncomplicated computation and advantageous for real time applications. Let the signals recorded from the EMG be designated by and n represent the values of the n-th sample of the raw signals, where n = 1,, with = 1500 (1500 samples corresponds to 3 seconds of the EMG recording). The following four statistical features are computed from the raw EMG data: Means of raw signals Fig.. The experimental setup shows the connection of EMG sensors from the Biceps muscle. All data acquisition is done by using the I USB 651 and recorded with the MATLAB software for further processing. Data are sampled at a rate of 1 khz and a gain of 100. The experiment setup used is shown in Figure 3. In order to study the pattern of the EMG signal of forearm muscle, five series of activities were carried out; rest, slow weak contraction, slow strong contraction, fast weak contraction and fast strong contraction. Each respective series will give 10 recordings. A recording is acquired for 3 seconds time. A short break is given in between each recording while a longer break is provided after a complete particular series of acquisition. 1 (1) n n 1 Standard deviation of raw signals 1 n () n 1 Variance of raw signals 1 n (3) n 1 Maximum of raw signals The largest value from the raw signals, ISS: ISB:

3 max[ ] n (4) n The above mention features can represent the pattern of typical EMG signal for certain activity. These statistical features formed an input vector for the neural network. f ( ) (5) Classification A two layered feed-forward back-propagation neural network is used for classifying five types of Biceps Brachii muscle activities. It is composed of 4 neurons in the input layer, 50 neurons in the hidden layer and 4 neurons in the output layer as shown in Figure 4. Since there is no clear algorithm to define the number of neurons of the hidden layer, the number of neurons in the hidden layer has been defined after a set of trials with different number of neurons as shown in Table 1. The number of hidden neurons is determined based on the highest classification achieved by the neural network build. the training stage: (1) a total mean squared error of is achieved or () the training will stop after epochs. The vector (a, b, c, d) of the output layer defines five types of muscle movement as shown in Table. Table 1. Experiments carried out to study the number of neurons in the hidden layer and the classification rate achieved. umber of hidden neurons 10 Time (s) Regression (R) Rate of classification (%) Average Average Average Average Table. Desired network vector response with respect to class of movement. Fig. 4. The architecture of the proposed neural network comprised of Tansig and Logsig neurons. The EMG classification is divided into two stages: training and testing. The numbers of input learning data and testing data have been divided according to a ratio of 7:3, which are 175 and 75 respectively. Two alternative stopping criteria have been applied during Class of movement a b c d Rest Slow weak contraction Slow strong contraction Fast weak contraction Fast strong contraction Results & Discussion An overall of 88% of classification rate is obtained from testing data set by using the proposed neural network. The results for the classification of 5 types of muscle movements are shown in Table 3. The results show that the neural network is able to ISS: ISB:

4 discriminate all rest activity of all subjects, which achieved 100% of correct classification. Besides that, this classification system performed well to discriminate slow weak and fast weak contraction; 93.33% of correct classification. Table 3. Classification rate (%) for each monitored activity. Input\Target Rest Slow weak Slow strong Fast weak Fast strong Rest (b) Slow weak Slow strong Fast weak Fast strong For fast strong EMG signal discrimination, the system was capable to reach 86.67% recognition. However, this system could only discriminate 66.67% of slow strong muscle activity. Figure 5 shows the raw EMG signal acquired by the DAQ recorded via the electrode. After signal processing by MATLAB software, the signal obtained is as shown in Figure 6. From the graph of Voltage (V) vs. Samples, we can observe that there exists a special signal pattern when the biceps muscle contract. A peak is shown during each muscle contraction; the differences are solely the magnitude and the contraction cycle time. (c) (d) Fig. 5. Raw biphasic EMG data. (a) Slow weak (b) Slow strong (c) Fast weak and (d) Fast strong biceps contraction. On the other hand, weak and strong contractions play the role to determine the magnitude of the EMG signal; it is related to the number of fibers within the motor units that are elicited during the contraction. (a) From previous studies, the EMG signal pattern is governed by (1) the number of active motor units (motor unit recruitment) and () the rate at which the actual motor units generate action potentials (rate ISS: ISB:

5 coding) [1]. There are two main factors that influence the use of EMG: biological factors and technical factors. The biological factors consist of physiological variability, age, sex, skeletal morphology, psychological factors, skin thickness and weight. Technical factors are electrode placement, position and inter-electrode distance and statistical methodology [4], [5]. Owing to the above factors, the signal pattern obtained from each subject with respect to the respective muscle activities varies with each experiment. Fig. 6. Superimposed EMG patterns for different types of Biceps muscle movements. Conclusion From the study, we found out that there is dissimilarity in the EMG patterns elicited by different activities. These patterns can be further processed to produce a control command. Fives types of biceps muscle activities are successfully identified using the signal pattern generated from raw surface EMG data. The overall classification rate is about 88% while the correct classification rate for each respective activity achieves more than 66.67%. Thus, the experimental results also prove that suitable statistical features can be used in classifying muscle activities since they can be easily computed and advantageous in real time applications. Acknowledgements The authors would like to thank the Ministry of Higher Education (MoHE) for the financial support and the R&D department of University Malaysia Perlis for the administrative support for this project. References: [1] M. B. I. Reaz, M. S. Hussain, M. Yasin, Techniques of EMG signal analysis: detection, processing, classification and applications, [] J. B. Walker, B. Scroggins, M. Morse, Are paralyzed people really paralyzed? Probably not according to EMG analysis, IEEE International Conference on Engineering in Medicine and Biology Society, pp 1137, [3] L. S. Gary, M. K. Loretta, A guide for use and interpretation of kinesiology electromyography data, Physical Therapy, 80(5), pp , 000. [4] Y. L. Chong, K. Sundaraj & Z. Ibrahim, Development of a low cost DAQ system for biphasic EMG sensors, Malaysian Universities Conference on Engineering and Technology, 008. [5] D. K. Gary & P. O. Jeffrey, The clinical usefulness of surface electromyography in the diagnosis and treatment of temporomandibular disorders, Journal of American Dental Association, 137(6), pp , 006. [6] A. P. Ltd., Electromyograpghy (EMG), ADInstruments Reference Manual. [7] MusclePhysiology, ISS: ISB: