Multisource Information Fusion for Continuous Prediction of Joint Angles Using TCN Combined With Temporal Pattern Attention Mechanism

Continuous motion intention prediction is valuable for human-machine interaction (HMI), however, the accuracy and the efficiency of the existing related results are far from satisfactory. This study proposes a novel continuous motion-intention prediction method that combines multi-source information fusion with an improved temporal convolutional neural network (TCN), enhanced by the introduction of the temporal pattern attention (TPA) mechanism. By integrating the temporal features of surface electromyography (sEMG) and mechanomyography (MMG) signals, we fully exploit their synergistic effect in movement intention prediction. Using the TCN network as the continuous motion prediction model improves the training efficiency through parallel computation and a simple network structure. TCN avoids the possible gradient vanishing problem, and allows for parameter tuning according to different tasks. By integrating the TPA mechanism with the TCN, the model's ability to recognize human motion intention in sequential data is significantly improved by focusing on key time steps. This enhancement increases prediction accuracy and strengthens the model's ability to capture long-term dependencies in time series data. Experiments are conducted to show the effectiveness and the advantages of the proposed TPA-TCN based motion prediction in comparison with the related results.