Real-time Sports Fatigue Monitoring Using LSTM-based Neural Networks and Wearable Technology

With the wide application of surface electromyography (sEMG) in the field of exercise fatigue monitoring, the existing studies are facing the dual challenges of data quality and classification accuracy. The original electromyography (EMG) signal is susceptible to noise, data imbalance, and individual differences, which makes it difficult for traditional classification algorithms to accurately capture the characteristics of muscle activity, while deep learning methods have the ability of sequence modeling, but their generalization performance in complex physiological signal scenarios still needs to be improved. Aiming at real-time state monitoring of athletes in the process of sports to prevent excessive fatigue and sports injuries, protect health, and prevent accidents, this paper focuses on wearable sports fatigue monitoring technology, hoping to achieve real-time sports state monitoring through noninvasive detection technology and identification of sports fatigue. Through real-time monitoring of electrocardiograph (ECG) signals and EMG signals on related muscles during human exercise, some necessary filtering and preprocessing of the collected body surface bioelectric signals are carried out to extract strong fatigue-related feature indicators. Then, recurrent neural network (RNN) based on logarithmic short-term memory (LSTM) is applied to the classification and recognition of fatigue degree and state during exercise. The experimental results show that during exercise, the fatigue degree of exercisers is determined by exercise intensity and time, and the subjective fatigue degree also increases. At the same time, it affects the values of heart rate, high- and low-frequency power (HF), root mean square (RMS), and mean power frequency (MPF), which can be used as indicators of exercise fatigue state recognition. For the experiment of fatigue recognition during sports state, the experimental results show that LSTM neural network can use historical information to mine the effective information of physiological signal features, which can be used to improve the accuracy of the sports fatigue state classification. Compared with the traditional support vector machines (SVM) algorithm, the time domain of LSTM model is 82.67%, the frequency domain is 76.58%, and the time–frequency domain is 90.63%. It effectively overcomes the interference caused by data noise and feature complexity and provides a reliable scheme for high-precision motion fatigue monitoring.