Neuro-Fuzzy Surface EMG Pattern Recognition For Multifunctional Hand Prosthesis Control M.Khezri, M. Jahed Electrical engineering Department Sharif University of Technology Tehran, Iran Email: khezri@ee.sharif.edu , jahed@sharif.edu N. Sadati Electrical engineering Department Sharif University of Technology Tehran, Iran Email: sadati@sina.sharif.edu Abstract---- Electromyogram (EMG) signal is an electrical manifestation of muscle contractions. EMG signal collected from surface of the skin, a non-invasive bioelectric signal, can be used in different rehabilitation applications and artificial extremities control. This study has proposed to utilize the surface EMG (SEMG) signal to recognize patterns of hand prosthesis movements. It suggests using an adaptive neuro-fuzzy inference system (ANFIS) to identify motion commands for the control of a prosthetic hand. In this work a hybrid method for training fuzzy system, consisting of back-propagation (BP) and least mean square (LMS) is utilized. Also in order to optimize the number of fuzzy rules, a subtractive clustering algorithm has been developed. The myoelectric signals utilized to classify, were six hand movements. Features chosen for SEMG signal were time and time-frequency domain. Neuro-fuzzy systems designed and utilized in this study were tested independently and in a combined manner for both time and time-frequency features. The results showed that the combined feature implementation was the best in regard to identification of required movement tasks. The average accuracy of system for the combined approach was 96%. I. INTRODUCTION The myoelectric signal (MES), recorded at the skin surface, has become an important tool in rehabilitation for amputees. The MES gives us information about the neuromuscular activity from which it originates, and this has been fundamental to its use in clinical diagnosis, and as a source of control for assistive devices and schemes of functional electrical stimulation. It has been proposed that the electromyographic signals from upper limb musculature can be used to identify motion commands for the control of an externally powered prosthesis hand [1]. Extracted information from EMG signals which are represented in a feature vector is chosen to minimize the control error [2],[3]. In order to achieve this goal, a feature set must be chosen which maximally separates the desired output classes. Extraction of accurate features from EMG signals is the main kernel of classification systems and is essential to the motion command identification. SEMG signal for prosthesis application is generally acquired by placing one or more differential electrode on the skin of the user. The non-stationary nature of SEMG signal makes it difficult to extract feature parameters precisely with the block processing stationary model such as autoregressive (AR) model [4]. Also it is very difficult for one feature parameter to reflect the unique feature of the measured SEMG signals to a motion command perfectly. Therefore in order to increase recognition rate of this system, two types of features which are time domain and time-frequency representations have been used. In time domain we use three major feature of SEMG signal such as mean absolute value (MAV), slope sign changes (SSC) and autoregressive (AR) model coefficients of signal and in time-frequency domain, zero crossing (ZC) of wavelet transform have been used [3]. Therefore our feature set composed of four features. Once a feature set has been constructed, they are fed to a classifier for discriminating among six motion commands of human hand. The movements that we apply are Hand opening and closing, wrist radial flexion and extension, pinch, and thumb flexion. The system presented in this work is based on a new approach, namely a neuro-fuzzy classifier. For training this system, a hybrid method including backpropagation and least mean square will be utilized. Also we use subtractive clustering method to specify fuzzy system rules. After implementing SEMG pattern recognition system, we acquire classification rate of this system which varies from 88% to100%. In the first case only time domain features and in another case compound features have been used. This fuzzy algorithm demonstrated success in pattern recognition of SEMG signal and allows for suitable control in multifunctional prosthesis hands. II. SEMG ACQUISITION AND PREPROCESSING The EMG signal is the electrical manifestation of the neuromuscular activation associated with a contracting muscle. It is a complicated signal influenced by various factors such as physiological and anatomical properties and characteristics of instrumentation. This signal differs from one person to another. A good acquisition of the SEMG signal is a prerequisite for good signal processing. In this work we use two channels of differential surface electrodes for collecting SEMG signal [1]. This Signal is easily affected by undesired signal that come from different sources such as 50/60 Hz electromagnetic interference from power lines. In addition, for surface electrode instrumentation, complicating issues may arise due to its coupling with skin (e.g. impedance of the skin may vary as function of the moisture of the skin, the superficial skin oil content and the density of dead cell layer).We place differential electrodes on the forearm under the elbow and place reference electrode on the wrist. After the acquisition, SEMG signal is filtered generally using a band-pass filter and amplified by 269 1-4244-0755-9/07/$20.00 '2007 IEEE