A PSR-Enhanced MSCNN-ViT Framework for Multi-Channel Surface Electromyography-Based Hand Gesture Recognition

Surface electromyography (sEMG) signals have shown great potential for decoding human motion intent through deep neural network training, which is critical for applications in prosthetic control and rehabilitation training. Therefore, research on sEMG-based gesture recognition carries substantial academic significance and societal value. However, existing methods often suffer from low feature learning efficiency and limited recognition performance due to the inherently non-stationary nature and small-sample characteristics of sEMG data. Although various feature selection and model design strategies have been explored, the process often involves complex architecture configurations and numerous combinations of features, leading to increased computational workload and suboptimal recognition results. To address this issue, we propose a novel hybrid architecture, termed PSR-MSCNN-ViT, which integrates phase space reconstruction (PSR)-based data augmentation with a multi-scale convolutional neural network (MSCNN) and a Vision Transformer (ViT). The PSR technique expands the non-stationary sEMG signals into a higher-dimensional state space, enhancing their nonlinear feature representation. To effectively model these high-dimensional dynamics, the MSCNN captures both phase trajectory and temporal dependency features, while the ViT module is employed to refine long-range attention allocation. This design not only improves parameter efficiency but also significantly enhances gesture recognition accuracy. Experimental results on the Ninapro DB5 dataset demonstrate that our proposed method achieves a classification accuracy of 76.02%, outperforming existing state-of-the-art approaches.