Research On Gesture Recognition Method Based On Deep Learning And Surface EMG Signal

With human-computer interaction technology widely used in sign language recognition,intelligent prosthetic limb control,rehabilitation of bones,EMG signals have become a research hotspot.In order to achieve a more accurate recognition of human movement intentions,to prejudge human gestures and therefore to react in advance,EMG signal and deep learning network for recognition and classification has been employed as the main method in this research.Natural control methods based on surface electromyography(s EMG)and pattern recognition are promising for hand prostheses.However,the robustness of the control provided by the Scientific Research Institute is still not sufficient for many real-life applications,and commercial prostheses can only provide natural control of a few movements.In recent years,with the wide application of deep learning technology,multiple machine learning fields including computer vision and speech recognition have been completely changed.The paper is based on the exploration of the application background of a large number of complete subjects and amputee subjects to control the robot hand through EMG signals,to investigate the surface EMG signal gesture recognition methods through deep learning technology.The purpose of the paper is to evaluate the performance of the deep learning algorithm that uses RMS rectified surface EMG signals for EMG control.Three different network architectures were tested: convolutional neural network(CNN),long-short-term memory network LSTM in a recurrent neural network(RNN)model,and the combination of the two labeled as CNN-LSTM in order.17 hand movements of 40 complete subjects from dataset 2 of the Swiss public dataset Ninapro and 11 amputees from dataset 3 are classified.Experiments show that the CNN network model provides the best recognition results with the average corresponding recognition accuracy of complete subjects(data set 2)reaching 82.60%,and the amputated subjects(data set 3)reaching 71.58%.For LSTM network,the model recognizing rate is relatively decreased.The average recognition accuracy rate of complete subjects(data set 2)is 64.11%,and the average recognition accuracy rate of amputated subjects(data set 3)is 53.37%.For the CNN-LSTM network model,the average recognition accuracy rate of complete subjects(data set 2)is 62.74%,and the average recognition accuracy of the amputated subject(data set 3)is 52.29%.Since most of the end users of the EMG control system are amputees,the amputee subject data with the highest recognition accuracy in the data set is selected to optimize the initial parameter training batch size of the CNN network structure,with the cross-entropy loss function as the optimization goal Finally,all subjects in the amputee data set(data set 3)are tested,and the average recognition accuracy rate is 71.98%.