Fast prediction of lower-limb joint angles from muscle synergies with DLSSANet

Abstract Fast and accurate prediction of lower limb joint angles is essential for exoskeleton control, enabling more effective assistance for patients during walking. Compared with consequential signals such as inertial measurement units (IMUs) and torque, pre-movement surface electromyography (sEMG) signals offer superior predictive performance. However, existing sEMG-based prediction methods rely on manually extracted feature inputs, which result in long prediction times, low accuracy, and a lack of physiological interpretability. To address these limitations, this paper proposes a method that combines muscle synergy and deep spatial-temporal learning for predicting lower limb joint angles. Non-negative matrix factorization is applied to the sEMG data to extract the spatial factor matrix W , which is used as a compact synergy-inspired low-rank feature and fed into a deep neural network for regression. A DLSSANet model, consisting of depthwise separable convolution, an LSTM, and separable self-attention, is designed to extract both spatiotemporal and global features for predicting lower limb joint angles. sEMG and IMU data collected from eight participants walking at various speeds were used for model evaluation. The results demonstrate that the muscle synergy and activation coefficient matrices extracted from sEMG signals exhibit significant correlations with joint angles. Notably, the muscle synergy matrix with a smaller time window achieves the highest prediction accuracy, with an R 2 of 0.937 and an root mean squared error of 2.93 ∘ , outperforming traditional temporal features and the activation coefficient matrix. The total prediction time is 5.30 ms on the CPU for exoskeleton control and 6.25 ms on the NPU for edge computing.