Classification of sEMG Signal-Based Arm Action Using Convolutional Neural Network

Abstract

Prosthetic arms are rapidly gaining pace because of their use in the development of robotic prosthesis. The surface electromyography (sEMG) signal acquired from the residual limb of amputee helps to control the movement of prosthesis. In this work, the sEMG signal is acquired using a dual-channel amplifier (Olimex EMG shield) from the below-elbow muscles of one amputee for six actions. The acquired signal is pre-processed using band-pass and band-stop filter to eliminate the noise in the signal. The classification is accomplished using machine learning and deep learning approach. The testing of machine learning algorithm and deep learning are implemented in Raspberry Pi 3 embedded in a Python script. In the machine learning approach, 11 relevant time domain features are extracted from pre-processed signal that are fed as input to linear support vector machines for classification. In the second approach, the signals are converted into images for deep learning analysis. Convolutional neural network (CNN) is used for classification of six hand actions. The model is trained and tested by varying a number of steps per epoch and number of epochs, and accuracy is compared with linear support vector machine (SVM). Results demonstrate that the mean accuracy of linear support vector machine is observed to be 76.66%, whereas for CNN model with 1000 steps per epoch with 10 epochs, it is found to be 91.66%. The classification accuracy of CNN is higher than machine learning model. Thus the usage of proposed methodology with CNN model could act as an indigenous system for sEMG signal acquisition and to activate the signal-controlled prosthetic arm which aids in rehabilitation of the amputees.