Muscle Synergy-driven TimesNet Method for Continuous Recognition of Lower Limb Motion Patterns

Abstract Accurate recognition of human lower limb motion patterns is essential for the effective control of intelligent assistive robots. However, most existing methods in recent studies relied on the complete gait cycle as input, limiting their ability to perform continuous recognition throughout the entire motion process. To address this limitation, this study proposed a muscle synergy-driven TimesNet model with low computational complexity to enable efficient and accurate continuous recognition of lower limb motion patterns. Surface electromyography (sEMG) signals were collected from 11 muscles of the dominant legs of 15 subjects during five types of lower limb movements. The 11-channel preprocessed sEMG signals were decomposed into 8 muscle synergies using principal component analysis (PCA). To comprehensively extract features from the one-dimensional time series data, a deep learning model known as TimesNet was employed to capture both intra-period and inter-period information from the synergy data. The proposed method achieved an average recognition accuracy of 0.91, with computational cost measured at 0.98 billion FLOPs, demonstrating state-of-the-art performance with low computational complexity. These results indicated that the proposed method enables efficient and accurate continuous recognition of lower limb motion patterns. This study provided a more practical and real-time applicable solution for motion pattern monitoring in the control of assistive robotic devices, such as exoskeletons and intelligent prostheses.