Locomotion Modes Recognition Driven by Time-Variant Muscle Synergy Feature

Muscle synergy has been proven to be simple and intuitive EMG features for myoelectric interface to simplify human machine interaction in upper limb. However, suffered from poor discriminability result from similar feature between locomotion modes, spatial muscle synergy hasn’t been widely applied in myoelectric interface for lower limb assistance. Recent studies suggested to provide synergy-based neural controller with extra time-variant information to improve their dynamic performance, but it remains to be clarified that synergetic pattern in either invariant spatial or variant time domain, is superior for locomotion mode recognition. Instead of traditional locomotion mode recognition with spatial muscle synergy, we extracted the synergetic pattern in both spatial and temporal domains via non-negative matrix factorization (NMF) algorithm, and their performance together with similarities between locomotion modes, that is evaluated by recognizing five locomotion modes using four common machine learning algorithms. The results showed that the performance of time-variant muscle synergy pattern on machine learning classifiers was significantly better than that of spatial time-invariant features. All the used classification models based on time-variant features achieved the better results, with a classification accuracy of 93% for machine learning method and 97% for deep learning method, which supported the view of taking time-variant muscle synergy pattern into consideration for lower limb myoelectric interface. The proposed method improves the classifier performance of commonly used synergy-based locomotion modes recognition, and provides a new research idea that encourages future HMI systems or instruments to develop more dynamic-sensitive measurement setup for synergy-based myoelectric interface.