Multisource Adversarial Feature Disentanglement Method for Cross-Subject Gesture Recognition Using sEMG Signals
Kejia Su, Kai Liu, Bo Wan, Hanbing Qiao, Jiayang Huang, Min Feng, Jinhui Liu · IEEE Transactions on Instrumentation and Measurement · 2025
Cross-subject gesture recognition remains a challenge in domain adaptation (DA), as most existing methods focus on extracting domain-invariant class-relevant features, treating all source domains as a single entity. These methods often neglect the importance of domain-specific class-relevant features, which are crucial for capturing individual subject variations. To address this, we propose a novel Multi-Source Adversarial Feature Disentanglement (MSAFD) method that separates domain-invariant and domain-specific features, improving gesture recognition accuracy and robustness. The MSAFD method introduces a two-stage transfer learning strategy. In the pre-training stage, a multi-source adversarial network extracts generalized domain-invariant features by minimizing domain discrepancies across multiple source domains. In the calibration stage, a pre-trained domain-invariant feature extractor captures shared features between source and target domains. Simultaneously, a domain-specific feature extractor, enhanced with an orthogonal projection layer, extracts target domain-specific features by projecting target domain features orthogonally to the domain-invariant features. This ensures that domain-specific features capture only the unique variations of the target domain. The disentangled domain-invariant and domain-specific features are then processed through an attention module to improve gesture classification. This approach enables precise feature separation, preserves critical domain-specific 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 ms 1 0.73 ms with a recognition accuracy of 98.33% ± 1.03%, highlighting its suitability for real-time gesture recognition systems.