A Multi-Branch Adaptive Model with Hybrid Time—Frequency Loss to the Enhanced Joint Moment Prediction of Prosthetic Control and Human Motion Applications

Abstract

In the fields of prosthetic control and rapid response prediction for human motion, the accurate prediction of joint moments is crucial for understanding and simulating human behavior. However, traditional time series models, especially when trained using small batches and limited data for single-time step predictions, frequently encounter difficulties in managing long-term dependencies. This deficiency significantly compromises their ability to generalize and maintain predictive accuracy over extended periods. To address these challenges, an innovative model called Multi-Branch Adaptive Encoding (MAE) has been introduced. This model features an adaptive weight module that employs a multi-branch input strategy to dynamically allocate weights to different surface electromyography (sEMG) signals and joint angles, thereby optimizing the processing of small sample data. Additionally, a feature extraction encoder, named Simplified Feature Transformer (SFT) has been designed. This encoder substitutes traditional attention mechanisms with a Multilayer Perceptron (MLP) and omits the decoder component to enhance the model’s efficiency and offer significant advantages in small-batch training and long-term prediction capabilities. A Hybrid Time–Frequency Loss (HTFLoss) has also been introduced to complement the MAE model. This approach significantly enhances the model’s ability to handle long-term dependencies. The MAE model and HTFLoss demonstrate an increase in Variance Accounted For (VAF) of 0.08 ± 0.03, a reduction in Root Mean Square Error (RMSE) of 1.77 ± 0.735, and an improvement in the coefficient of determination (R²) of 0.09 ± 0.05, indicating substantial superiority. These enhancements highlight the extensive potential applications of the model in the fields of rehabilitation medicine, and human-machine interaction. The improved predictive accuracy and the ability to manage long-term dependencies make this model particularly valuable in designing advanced prosthetic devices that can better mimic natural limb movements, thereby improving the quality of life for amputees.

Keywords

• prediction joint moment
• long-term dependencies
• time series models