A Data Augmentation Method for sEMG Signals Based on CWGAN-GP Utilizing Multi-scale Feature Fusion Attention Mechanism

This paper introduces a novel GAN-based data augmentation method for surface electromyography (sEMG) signals. By integrating channel attention mechanism and multi-scale feature fusion as core units for feature extraction, this method is implemented within the CWGAN-GP framework. The model’s generative performance is evaluated using both similarity metrics and classifier accuracy. Experimental results show that the data generated by this method outperforms other approaches, achieving improvements of over 2.1% in CC and over 1.1% in RMSE. Additionally, when evaluated with a practical classifier, the dataset augmented by this method achieved an average accuracy of 0.825. The best augmentation effect was obtained when approximately 37.5% of the generated data was incorporated. In summary, this proposed sEMG signal augmentation method provides a promising solution to improve gesture recognition classifier performance and address the scarcity of sEMG data, thus offering substantial support for the widespread adoption of human-machine interaction devices, such as myoelectric prostheses, in real-world applications.