Dynamic Muscle Fatigue State Recognition Based on Deep Learning Fusion Model

The timely and accurate recognition of dynamic muscle fatigued states in rehabilitation training helped in tailoring more scientifically effective training plans for patients. In this study, a CNN-LSTM-Transformer (CLT) model based on surface electromyographic signals (sEMG) was proposed to address the limited classification and low identification accuracy observed in existing muscle fatigued state recognition methods. This model combined the traditional convolutional neural network (CNN), long short-term memory network (LSTM), and Transformer encoder to achieve accurate classification of dynamic muscle fatigued. First, elbow flexion-extension fatigued experiments were conducted on 20 healthy participants, who were divided into four fatigued states based on the degree of fatigued. Subsequently, the acquired sEMG signal data was preprocessed and two nonlinear features, approximate entropy (ApEn) and permutation entropy (PE), were extracted as input features for machine learning. Finally, a CLT fatigued recognition model was constructed using the preprocessed sEMG signal data and compared with CNN, LSTM, and random forest (RF) models. The results showed that the CLT model had higher accuracy in recognizing muscle fatigued states compared to CNN, LSTM, and RF models, with a respective increase of 4.1%, 5.06%, and 8.39%. This model had good classification performance and could be used to monitor muscle fatigued during upper limb rehabilitation processes.