Multisource Adversarial Feature Disentanglement Method for Cross-Subject Gesture Recognition Using sEMG Signals

Cross-subject gesture recognition remains a challenge in domain adaptation (DA), as most existing methods focus on extracting domain-invariant (DI) class-relevant features, treating all source domains as a single entity. These methods often neglect the importance of domain-specific (DS) class-relevant features, which are crucial for capturing individual subject variations. To address this, we propose a novel multisource adversarial feature disentanglement (MSAFD) method that separates DI and DS features, improving gesture recognition accuracy and robustness. The MSAFD method introduces a two-stage transfer learning (TL) strategy. In the pretraining stage, a multisource adversarial network extracts generalized DI features by minimizing domain discrepancies across multiple source domains. In the calibration stage, a pretrained DI feature extractor captures shared features between source and target domains. Simultaneously, a DS feature extractor, enhanced with an orthogonal projection layer, extracts target DS features by projecting target domain features orthogonally to the DI features. This ensures that DS features capture only the unique variations of the target domain. The disentangled DI and DS features are then processed through an attention module to improve gesture classification. This approach enables precise feature separation, preserves critical DS information, and enhances generalization across subjects. Experimental results on the self-collected XDMyo and public MYO datasets show that MSAFD outperforms benchmark methods, achieving the highest classification accuracy and lowest standard deviation. Real-time experiments demonstrate that the method achieves an average gesture recognition latency of $111.86~\pm ~0.73$ ms with a recognition accuracy of 98.33% $\pm ~1.03$ %, highlighting its suitability for real-time gesture recognition systems.