NEURAL NETWORK METHODS IN APPLICATION FOR MYOELECTRICAL SIGNALS CLASSIFICATION

摘要

This research investigates the problem of the movement classification by surface myoelectrical signals (MES), used for electromyographical (EMG) control of powered upper limbs, and also for biometric identification of the person. On of the solutions in this task is using pattern recognition approach. In this case the success of the myoelectric control scheme depends largely on the classification accuracy. The main target of the research was comparison of various neural network classifiers, such as multi layer perceptron with back-propagation learning algorithm (BPG), neural networks with radial-basis functions (RBF), probabilistic neural networks (PNN) and Kohonen's self-organised maps (SOM). Fundamental to the success of chosen method was the sheme, which involves a wavelet based feature set, dimensionally reduced by principal components analysis (PCA), and classified by SOM classifier. It was also detected that the best accurate performance is possible when using 30 components as input vector for classifier, and four channels of myoelectric data greatly improve the classification accuracy, as compared to one channel.