sEMG-Based Lower Limb Joint Angle Prediction With Muscle Synergy Spatio-Temporal Fusion and DWCIT Model

摘要

In lower limb rehabilitation training, using surface electromyography (sEMG) signals for continuous joint angle prediction is crucial for controlling lower limb exoskeleton robots. This paper proposes a method for continuously estimating lower limb joint angles during walking based on sEMG. The method integrates spatio-temporal information of muscle synergies extracted from the sEMG signals and employs the Depthwise Convolutional iTransformer (DWCIT) model for joint angle prediction. First, the sEMG data are decomposed into a muscle synergy matrix and an activation coefficient matrix using non-negative matrix factorization (NMF). Then, these matrices are subsequently fed into the DWCNN model to extract spatial and temporal features, respectively. Finally, the iTransformer model is utilized to capture the inter-channel correlations among multiple feature channels and perform information fusion, thereby facilitating the accurate continuous prediction of hip and knee joint angles. sEMG and joint angle data were collected during the experiment from eight participants walking at different speeds. The results indicate that the W-HST features with the DWCIT model achieved an R² of 0.956 and an RMSE of 3.26°. Compared with using muscle synergy features (W, H) and time-domain features (RMS, MAV, IEMG) respectively, this represents an improvement in R² of at least 3.24% and 4.48%, and a reduction in RMSE of 0.97° and 1.31°. On average, our algorithm DWCIT outperformed other algorithms by improving R² by 3.68% and reducing RMSE by 1.00°. These findings demonstrate that the deep learning model based on muscle synergy effectively captures the complex nonlinear relationship between muscle activity and hip-knee joint motion.