sEMG-Based Multi-view Feature-Constrained Representation Learning

The information contained in sparse multi-channel sEMG signals is limited, which requires a high standard for feature representation learning. Based on this, some solutions involve using multiple features to construct multiple views in order to extract more information, but previous methods mostly focus on divergent learning and merging shared information, without considering the inherent characteristics of multi-view data, leading to often existing distribution gaps and information redundancy in feature representation. In view of this problem, we propose a multi-view feature constraint representation learning method based on sEMG, which is divided into three parts. The first part constructs a dedicated encoder for each view, the second part focuses on extracting the shared and specific representations of different views, and uses multiple constraints to balance the sharing, complementation, and redundancy of information. The third part constructs a multi-feature low-rank fusion module to get the final representation for different downstream tasks. We take patient classification as the downstream task. Experiments conducted on DB1, DB5, Myo three datasets show that our method has significant advantages.