EMG signal classification using neural network with AR model coefficients J. Tomaszewski*, T. G. Amaral**, O. P. Dias***, A. Wo ł czowski*, M. Kurzy ń ski*,**** *Institute of Informatics, Automatics and Robotics, Faculty of Electronics, Technical University of Wroclaw, Wroclaw, Poland (email: jacek.tomaszewski@student.pwr.wroc.pl; andrzej.wolczowski@pwr.wroc.pl) **Escola Superior de Tecnologia de Setúbal/IPS, CESET, ISR – Pólo Coimbra, 2910-761 Setúbal, Portugal (Tel: +351-265-79000; e-mail: tamaral@est.ips.pt) ***Escola Superior de Tecnologia de Setúbal/IPS, CESET, INESC – Lisboa, 2910-761 Setúbal, Portugal (Tel: +351-265-79000; e-mail: pdias@est.ips.pt) ****Chair of Systems and Computer Networks (email: marek.kurzynski@pwr.wroc.pl) Abstract: In this work authors proposes two approaches for discriminating between five predefined grasps using the EMG signals. In the first approach, the signal energy and the number of zero-crossing are used as signal features. In the second, the signal is modeled using the AR model and its coefficients became the features. Feature vectors created from both approaches are processed by a neural network with a linear transfer function, which classifies them as one of above mentioned grasps. From the experimental results, pattern identification using the AR model obtained good efficiency for the studied grasps. 1. INTRODUCTION Electromyographic signals are biomedical signals that occur in the muscle tissue during contraction or at rest due to generation of electrical potentials by muscle cells. They are function of time and can be described by their amplitude, frequency and phase. During certain movements, the contraction and the distention of the muscles originate specific signal patterns. Thus EMG signals can be used to control active prostheses of human limbs (Wołczowski, 2001). EMG signals can be acquired directly from muscles by intramuscular electrodes or from the surface of the skin by surface electrodes (sEMG). In this approach latter method is used. It is much more convenient, but acquired signal is contaminated with more noise and interacts with signals from other motor units. Therefore reliable pattern classification algorithms must be used. In this paper author proposes two methods of pattern identification. Both use linear neural network for classification of two different feature vectors. First uses the signal energy and number of zero-crossing for characterizing the signal, while second use the coefficients computed in the autoregressive model. 2. ANALYSED GRASPS EMG signal of five different grasps were recorded for analysis as shown in figure 1. The following grasps were chosen from a set of movements defined by G. Schlesinger: Palmar, Tip, Cylindrical, Cylindrical tight, and Spherical (Schlesinger, G. 1919). Figure 1: Grasps recognized in experiment. 3. MEASUREMENT STAND Measurement stand used in experiment consisted of three parts: ● AD converter connected to PC, ● galvanic separator, ● measurement interface - EMG electrodes. The AD converter Ni 4477 card from National Instruments was used. Card contains 8 analog input channels with