Prediction Method of Lower Limb Muscle Fatigue Based on Combining Random Forest and Gated Recurrent Unit Neural Network

In this paper, the traditional fatigue state classification method is abandoned, and a neural network model is established to predict the variation of muscle fatigue by using the extracted muscle fatigue characteristics. This is of great significance for the subsequent use of muscle fatigue characteristics to compensate for the changes of sEMG signals caused by muscle fatigue in continuous motion, so as to achieve the compliance control of the exoskeleton. In the muscle fatigue experiment, we selected 7 representative subjects and collected the data of each subject from non-fatigue state to fatigue state during the dynamic contraction of the lower limb, fifteen sets of data were collected for each subject. In this paper, a muscle fatigue prediction method combining random forest (RF) and gated recursive unit (GRU) neural network is proposed, in the experiment, 75 sets of data from the first 5 subjects were used for model training, and 30 sets of data from the last 2 subjects were used for model test, and each set of data was predicted separately. In order to verify the generalization of the proposed model, 20 experiments are carried out. The experimental results show that compared with the traditional recursive neural network (RNN), long and short term memory (LSTM), GRU and multi-layer feedforward and back propagation neural network (BPNN), the proposed model has the advantages of higher prediction accuracy and better generalization.