Abstract:

In the process of exoskeleton design， the evaluation of assistance performance directly impacts the overall structural safety and efficiency. Addressing the current issue of predominantly utilizing single metrics for performance evaluation， a method based on multi?source physiological signals （surface electromyography， photopretismography， and respiration） for LSTM prediction of motion energy consumption was proposed. This method involves preprocessing and feature extraction of physiological signals， followed by prediction using a 6?layer LSTM model and validation through K-fold cross?validation. Comparative experiments with DT and SVM were conducted. Finally， an online energy consumption monitoring system was established， serving as a basis for evaluating exoskeleton assistance performance. Results indicate the feasibility of utilizing multi?source physiological signals for fusion prediction， with an RMSE of 0.073 kJ for the three?source signal. The LSTM model achieves a 39.53% and 15.68% reduction in RMSE compared to DT and SVM， respectively. The total energy consumption error on the test set is 23.98 kJ， demonstrating the superior performance of the LSTM model for total energy consumption prediction and its suitability for exoskeleton assistance performance evaluation.

Key words: energy consumption prediction, LSTM, exoskeleton performance evaluation, signal fusion, online monitoring