Optimizing sEMG-Based Gesture Recognition Under Nonideal Conditions Through a Robust Deep Learning Approach

Gesture recognition schemes based on surface electromyography (sEMG) signals and used for myoelectric control have been widely acknowledged and investigated. However, non-ideal conditions affect the practical applicability of sEMG-based human-machine interfaces (HMIs). The impacts of these disturbances on deep learning methods have not yet been thoroughly explored, and the adaptability and robustness of these approaches still need to be evaluated and quantified. In this paper, a multibranch enhanced spatial-temporal feature-based hybrid neural network (MESTNet) structure is proposed to extract intention information from sEMG maps. Convolution layers with squeeze-and-excitation blocks are used to enhance the critical features and reduce redundant features. Then, the interdependency of the sEMG sequence is learned through a bidirectional gated recurrent unit (BiGRU). A domain adaptation approach is also introduced for addressing unknown non-ideal factors to explore the generalizability. Two different modeling methods were employed to compare the effects of five non-ideal conditions on MESTNet and several comparison methods. The performance of the models was evaluated on the public SeNic dataset, where MESTNet outperformed the other classifiers, with average classification accuracies of 91.89% and 73.43% produced for different modeling methods under all non-ideal conditions. The results indicate that the proposed MESTNet outperformed the competing approaches on this dataset. Furthermore, analyzing the accuracy variations exhibited under different conditions provides valuable recommendations for designing interfaces for practical and clinical myoelectric control applications, which can inspire future sEMG-based HMI designs.