MVMD-TCCA: A method for gesture classification based on surface electromyographic signals

Abstract

Gesture recognition plays a fundamental role in enabling nonverbal communication and interaction, as well as assisting individuals with motor impairments in performing daily tasks. Surface electromyographic (sEMG) signals, which can effectively detect and predict motor intentions, are integral to achieving accurate gesture classification. This paper proposes a method, the multivariate variational mode decomposition and the two-channel convolutional neural network with added attention mechanism (MVMD-TCCA), to enhance the accuracy of gesture classification for motor intention recognition. The MVMD technique is utilized to decompose and fuse sEMG signals, enriching signal content and improving feature representation. To further optimize gesture classification performance, the convolutional block attention module (CBAM) and CrissCross attention mechanism are integrated into the neural network, enabling superior learning of local and spatial features. The experimental results show that the MVMD-TCCA method achieves an average classification accuracy of 85.09 % on the NinaPro DB2 dataset, representing a 13.46 % improvement compared to the use of the original signal, and an average classification accuracy of 97.90 % on the dataset collected from 15 subjects, reflecting a 1.70 % improvement over the original signal. These findings underscore the critical role of accurate gesture classification in facilitating daily task assistance for cerebral infarction patients, demonstrating the potential of the proposed approach.