A Convolutional Neural Network Classification Method Based on Non-Ideal sEMG Signals for Human-Robot Interaction System

摘要

Surface Electromyography (sEMG) is a widely utilized biological signal in the field of gesture recognition. The recognition performance of sEMG decreases due to muscle fatigue during prolonged exercise. To address this issue and enhance recognition accuracy, we propose an optimized Convolutional Neural Network classification algorithm (CBLSA) in this paper. The algorithm transforms the EMG data signal into a two-dimensional grayscale image. We introduce bidirectional Long Short-Term Memory network (BiLSTM) and Self-Attention mechanism to reinforce the activation of EMG signals across different channels, thereby minimizing the impact of signal weakness caused by muscle fatigue. The proposed algorithm is evaluated on both ideal conditions and muscle fatigue conditions using the public dataset Senic. Experimental results demonstrate that our algorithm achieves an accuracy rate of 98.61 % and 95.64 % respectively for these two scenarios, representing significant improvements compared to traditional machine learning methods and unoptimized convolutional neural networks.