A novel SE-ResNet architecture for continuous estimation of wrist and hand movements from HD-sEMG

With the rapid technological advancements in flexible manipulators, it has emerged as a research hot spot to realize a more flexible control strategy driven by human movements. This study investigates the feasibility of utilizing deep neural networks and a squeeze-and-excitation (SE) module based on high-density surface electromyography (HD-sEMG) signals, so as to enhance the prediction accuracy for the two-degrees-of-freedom wrist and metacarpophalangeal (MCP) joint angles. Two residual networks, namely, a 32-layer ResNet (ResNet-32) and a 56-layer ResNet (ResNet-56), were proposed, and their performance was compared with that of an 8-layer convolutional neural network (CNN-8). A dataset was collected from 7 subjects for the evaluation of prediction performance, which contains HD-sEMG and kinematic signals corresponding to wrist and hand movements. The root mean square error (RMSE) was taken as the evaluation metric. The RMSE of ResNet-32 and ResNet-56 for simultaneous movements can reach 17.80 and 17.99 degrees, respectively, which is significantly better than that of CNN-8. The findings revealed that an appropriate increase in the number of network layers is sufficient to improve prediction performance, because an excessive layer accumulation may incur an undesirable computational overhead and few improvements. Further, the ResNet-32 model was augmented with SE module, named as SE-ResNet model. A support vector machine (SVM) was introduced for comparative analysis. The RMSE of the SE-ResNet model for simultaneous movements can reach 16.48 degrees, which is better than that of SVM and ResNet-32. In brief, our SE-ResNet model exhibits better prediction performance and robustness.