sEMG-based recognition of composite motion with convolutional neural network

Surface electromyography (sEMG) signals are expected to help recognize motions precisely and timely for its generation origins from muscular contractions. In most cases, existing researches about sEMG-based motion recognition cannot guarantee comprehensively excellent performance and apply flexibly to different types of motions. This paper proposes a new initiative via deep learning to recognize general composite motions, which processes sEMG signals as images. Inspired by several definitions of “sEMG Image” for static gestures, we define a novel “sEMG Image” to represent composite motions, which can make different cooperation of muscles reflected on image textures. With a well-designed convolutional neural network (CNN), this method can obtain effective filters automatically to extract features of texture by the training of considerable data. The results from two experiments of different composite motion recognition (including gentle writing motions and drastic sign language motions) indicate that this method embraces high accuracies and strong generalization ability for several influence factors. In addition, two techniques are proposed to further optimize this method: pre-train the network with an irrelevant dataset to reduce the demand for data and speed up the convergence; fuse multiple sensors with simple modifications for CNN to improve performance greatly.