Movements Classification Through sEMG With Convolutional Vision Transformer and Stacking Ensemble Learning

Thanks to the powerful capability of the feature extraction, deep learning has become a promising technology for an increasing number of researchers to decode movements from surface Electromyography (sEMG) signals. The mainstream methods of deep learning are based on Convolutional Neural Networks (CNN) or Recurrent Neural Networks (RNN). However, the sequential features of sEMG signals will be ignored in CNN-based approaches while the training of the neural networks is much time-consuming in RNN-based approaches. To solve these problems, a novel convolutional vision transformer (CviT) with stacking ensemble learning is proposed in this paper, which has great potential in the fusion of sequential and spatial features of sEMG signals with the parallel training. In NinaPro DB2, the proposed method achieves 80.02% with the window length of 200ms. In the subset of NinaPro DB2 (Exercise E1), the proposed method achieves 83.47% and 84.09% with the window length of 200ms and 300ms respectively. In the subsets of NinaPro DB5 (Exercise A, Exercise B), the proposed method achieves 76.83% and 73.23% respectively. The experimental results demonstrate that the proposed CviT has better performance than most current approaches. In addition, the successful application of Transformer in sEMG-based movements classification provides a significant reference for the application of Transformer in other biological signals.