Stroke patient cross-subject gesture recognition method based on prototype network and high-density surface electromyography

Abstract

This invention belongs to the field of stroke rehabilitation engineering technology, specifically relating to a cross-subject gesture recognition method for stroke patients based on a prototype network and high-density surface electromyography (HD-sEMG). The method includes subject screening and preparation, HD-sEMG signal acquisition, signal preprocessing, macro- and micro-feature extraction, feature dimensionality reduction and standardization, prototype network model construction, data augmentation and signal optimization, and model training and validation. This invention improves feature discriminativeness through macro- and micro-feature fusion; simplifies the Transformer encoder to achieve high-quality feature embedding; utilizes a multi-loss collaborative framework to achieve a dynamic balance between cross-domain generalization and classification discriminativeness; enhances the model's anti-interference ability under small sample conditions through physiological data augmentation and signal optimization; and requires only a very small number of patient calibration samples to achieve accurate gesture recognition, significantly reducing calibration costs in clinical applications. This provides a feasible solution for the clinical translation of intelligent rehabilitation robots and closed-loop rehabilitation systems, promoting the intelligent development of stroke rehabilitation.

Images