NewRep: a classification model for online pattern recognition based on HD-EMG signals

High density (HD) electromyography (EMG) signals have been widely used for pattern recognition (PR) in myoelectric control, particularly for gesture recognition. Although neural networks generally outperform traditional machine learning models in EMG PR, the limited accuracy in recognition remains a critical challenge. To further improve the online classification of neural networks in EMG PR, this study establishes a new neural network named NewRep by introducing the self-attention mechanism module into the existing RepViT, a lightweight convolutional neural network inspired by vision transformer designs. Online gesture classification was performed with NewRep, RepViT, and linear discriminant analysis. Eighteen non-disabled subjects were recruited and tested in this study. For each subject, 256 channels of EMG signals were acquired from four forearm muscles: flexor digitorum superfcialis, flexor carpi radialis, extensor digitorum, and extensor carpi radialis brevis. Twenty hand and wrist motions were performed during offline and online experiments. Five metrics were employed to evaluate performances of the three models, including offline classification accuracy, motion selection time, motion completion time, motion completion rate, and online classification accuracy. Compared with RepViT and linear discriminant analysis, the proposed NewRep showed a significant improvement in motion completion time (decreased by 0.053 s and 0.117 s, respectively) and online classification accuracy (increased by 2.91% and 6.24%, respectively). The results demonstrated that NewRep effectively enhanced the real-time decoding capability of myoelectric control, offering a promising approach for practical applications.