Analysis of Typical Features Using an Angle-to-sEMG Decoder Model

Current prosthetic control methods mainly rely on extensive data collection and the training of classification models. Nevertheless, achieving a balance between learning efficiency and classification accuracy remains a persistent challenge. In this paper, we introduce a novel prosthetic control approach that divides the processing of surface electromyography (sEMG) signals into two stages to predict corresponding joint angles. Initially, a pretrained model is utilized to infer the associated joint angles based on typical features. This methodology necessitates the mapping of individual sEMG signals to these typical features through a compact mapping network during application. we think that it will enhance hand motion recognition accuracy while minimizing learning duration though a pre-training approach. This study represents the initial phase, which involves the acquisition and evaluation of the similarity among typical features linked to joint angles. To this end, we constructed angle-to-sEMG models (decoders) and input the same unprocessed continuous joint angle data into the decoders of five participants. Subsequently, we utilized the resultant typical features as inputs for the encoders of different individuals. The findings indicate that, notwithstanding interindividual variability, there exists a degree of resemblance in the typical features associated with joint angles, thus enabling the generation of angle data resembling the true angles even across diverse subjects.