Wavelet-Based Spatio-Temporal Fusion Network for Electromyographic Decoding in Bionic Hand Control Systems

Surface electromyography (EMG) signals are essential for facilitating intuitive interactions between humans and bionic hands. However, their inherent non-stationarity, low signal-to-noise ratio, and significant inter-individual variability present considerable obstacles to precise decoding. To overcome these challenges, this study proposes a novel recognition framework combining wavelet packet decomposition and a dynamic graph convolutional-Transformer model. The process starts with multi-layer wavelet packet decomposition and adaptive threshold denoising, effectively removing noise while retaining critical signal features. Subsequently, a dynamic graph convolutional network is employed to capture spatial interactions among multi-channel electrodes, and a Transformer encoder models long-term temporal dependencies within the signals. By integrating these methods, the model generates a fused feature representation that incorporates both spatial and temporal correlations. Experimental results demonstrate that the proposed model provides enhanced robustness to noise and achieves greater classification accuracy compared to conventional methods.