Deep Learning Enhances the Robustness of Online HD-sEMG Decomposition Against Electrode Detachment

Electrode detachment introduces noise into acquired surface electromyography (sEMG) signals, thereby compromising sEMG decomposition accuracy. In this study, we propose for the first time the deep learning-based technique for mitigating the effect of electrode detachment on sEMG decomposition. We constructed a bidirectional gated recurrent unit (Bi-GRU) network with data augmentation to enhance the robustness of high-density sEMG (HD-sEMG) decomposition. Data augmentation was achieved by simulating HD-sEMG data with electrode detachment, thereby enhancing the diversity of the training dataset. Randomly selected channels of sEMG were set to zero to simulate the electrode detachment. Two levels of electrode detachment interference were simulated (0-4 and 0-8 random-selected detached channels). Both simulated and experimental signals were utilized to evaluate model performance. The progressive FastICA peel-off (PFP) algorithm was employed to decompose sEMG signals. Both raw data and the derived motor unit (MU) firing sequences were segmented into training samples. The online PFP method using separation vectors served as the comparative method. Experimental results demonstrate that our proposed Bi-GRU network with data augmentation effectively enhances the robustness of HD-sEMG decomposition, achieving the highest matching rate among all methods under electrode detachment interference. This study further confirms the great potential of deep learning techniques to achieve robust sEMG decomposition, with potential applications in prosthetic control and rehabilitation medicine.