Gait Phase Recognition in Multi-Task Scenarios Based on sEMG Signals

In the human–exoskeleton interaction process, accurately recognizing gait phases is crucial for effectively assessing the assistance provided by the exoskeleton. However, due to the similarity in muscle activation patterns between adjacent gait phases, the recognition accuracy is often low, which can easily lead to confusion in surface electromyography (sEMG) feature extraction. This paper proposes a real-time recognition method based on multi-scale fuzzy approximate root mean entropy (MFAREn) and an Efficient Multi-Scale Attention Convolutional Neural Network (EMACNN), building upon the concept of fuzzy approximate entropy. MFAREn is used to extract the dynamic complexity and energy intensity features of sEMG signals, serving as the input matrix for EMACNN to achieve fast and accurate gait phase recognition. This study collected sEMG signals from 10 subjects performing continuous lower limb gait movements in five common motion scenarios for experimental validation. The results show that the proposed method achieves an average recognition accuracy of 95.72%, outperforming the other comparison methods. The method proposed in this paper is significantly different compared to other methods (p < 0.001). Notably, the recognition accuracy for walking in level walking, stairs ascending, and ramp ascending exceeds 95.5%. This method demonstrates a high recognition accuracy, enabling sEMG-based gait phase recognition and meeting the requirements for effective human–exoskeleton interaction.