Classification of multichannel surface-electromyography signals based on convolutional neural networks

Abstract Electromyography is a science that studies or detects bioelectrical activity of muscles to analyze skills and morphological changes of the neuromuscular system and contributes to studies on the neuromuscular system. Surface electromyography (SEMG) signal is a bioelectrical signal emitted when nervous and muscular activities are recorded from the surface of human skeletons by means of poles, which can reflect the functional state of nerves and muscles under non-invasive conditions on a real-time basis. SEMG signals found a wide application in different fields including prosthesis control, sports medicine, rehabilitation medicine, and clinical diagnosis. However, how to efficiently exact features from SEMG signals to realize accurate recognition of action modes is a key issue for the practice of electromyography-controlled prostheses and to achieve precision of rehabilitation treatment. Deep learning reveals drastic changes in many fields of machine learning, including machine vision and voice recognition, over the past few years. We use convolutional neural networks (CNNs) to extract deep features from SEMG signals and classify actions. CNNs exhibit good translation invariance due to its characteristics of local connection and weight sharing. If SEMG signals were applied in the modeling of electromyography signal recognition, then the diversity of electromyography signal itself can be overcome using invariance in convolutions. Therefore, in this study, the spectrogram obtained by analyzing electromyography signals is proposed to be used as an image. Intensively used deep convolutional networks in the image were also adopted to conduct the gesture motion recognition of SEMG signals.