EMG-based Finger Movement Classification for Prosthetic Hand Control

Abstract

Millions of people around the world are suffering from the consequences of amputation. In particular, advanced prosthetic hands could retrieve some functionality and improve quality of life of those who lost it. The most challenging part in the prosthetic hands is their finger control. Ideally, when the brain sends the signal for moving a finger, the corresponding finger is detected in a very short time. Therefore, not only the accuracy but time to reach that accuracy should be considered. EMG signal is found as a promising mechanism to control prosthetics. Unfortunately, EMG signal is easily contaminated with noise and various artifacts factors that deteriorates the accuracy of classification. In this research, we addressed key challenges in EMG classification. First, we proposed 5 new features to enhance the accuracy. Second, we advocated EVM as a powerful classifier and compared its performance with other EMG classifiers in the literature. Third, for online classification, we employed various recurrent neural network (RNN) structures, and showed that bidirectional-RNNs with sequential inputs could achieve higher accuracies in a shorter time. Fourth, to make the system robust against noise, we proposed a convolutional neural network feature learning (CNNFL) structure combined with EVM. Fifth, to address the drawback of the Bayesian fusion method, we proposed a novel postprocessing technique. Overall, in this research, we have customized machine learning techniques for EMG analysis that can produce accurate classification in realtime and real world noisy condition.