sEMG-based multi-joints motion estimation of lower limb utilizing deep convolutional neural network

The accuracy of the motion intention recognition is the security guarantee of human-machine interaction (HMI) control for lower limb rehabilitation exoskeleton (LLRE). Therefore, to advance the precision of the multi-joint motion intention recognition, the multi-channel surface electromyography (sEMG) signals of the subject with cycling and walking are collected, and the signals are processed with reasonable processing methods in this paper. Then, the deep convolutional neural network (CNN) model is constructed based on the processed sEMG signals to estimate the multi-joint angle of the lower limb. The feasibility and efficiency of the developed CNN model in the field of intention recognition of the lower limb multi-joint motion are verified by experimental simulation. Furthermore, compared with CNN model, the conventional back-propagation neural network (BPNN) model and radial basis function neural network (RBFNN) model, which demonstrates that the estimation accuracy of the developed CNN model is better than that of classical BPNN and RBFNN, and the root mean square errors (RMSE) of hip, knee and ankle joints estimated by utilizing CNN model are 3.8886°, 2.8199° and 3.1148°, respectively. It proves that the proposed CNN model can effectively recognize the motion intention of the lower limb multi-joint, which provides a theoretical basis for the research on HMI control of the LLRE.