Classification of Forearm Multichannel Electromyogram Signals by a Self-Organized Neuro-Fuzzy Structure

Hish rate classification of Electromyogram (EMG) signals for controlling of prosthetic hands is still a hot topic among the rehabilitation research titles. Specially, when the degree of freedom in artificial hands increases, the classification rate decreases dramatically. In this paper, a new five layer classifier based on Neuro-Fuzzy-Genetic structure was introduced to increase the classification accuracy of EMG signals. The proposed classifier has a self- organized structure, which adaptively creates new rules according to the input features and trains the fuzzy rule weights based on the back propagation method. Finally, the genetic algorithm (GA) was employed for the final tuning stage. In this study, six subjects were asked to perform 9 different movements and their EMG signals were caught during the tasks from the six different forearm muscles. In order to remove the noises, the signals were filtered. Then the integral absolute average (IAV), Cepstrum coefficients and Wavelet Packet Coefficients with entropy pruning were extracted from the filtered signals as features. We used principal components analysis (PCA) for dimensionality reduction (234 to 10). The dimensionality reduction by PCA simplifies the structure of the classifier and reduces the processing time for the pattern recognition. The proposed classifier was applied on the features and the results were led to higher than 96.7% classification rate for the 9 classes of movement. To make a comparison, support vector machine (SVM) was employed (76% classification rate for 9 classes) and the results showed a drastic supremacy of the proposed method.