sEMG-based lower-limbs motion recognition using deep learning with one dimensional convolutional neural network

Surface electromyographic (sEMG)-based motion recognition has been successfully applied in the exoskeleton, human–machine interaction, and rehabilitation engineering. To improve the accuracy of motion recognition and then achieve accurate control of the exoskeleton, this paper is devoted to proposing a one-dimensional (1D) convolutional neural network (1D-CNN) to classify seven daily lower-limb motions from the eight main muscles of human lower limbs. The proposed model architecture mainly consisted of three 1D convolutional layers, three max-pooling layers, and one global average pooling layer. The Teager–Kaiser energy operator was adopted for the starting and ending points detection of sEMG signals. To avoid overfitting, only two features (i.e., mean absolute value (MAV) and root mean square (RMS)) were extracted in this study and used as the input of the proposed model. The results show that the accuracy of motion recognition based on our proposed model had been improved to more than 97.0 ± 0.8% and was higher than that based on the deep neural network (DNN) and support vector machine (SVM) models (95.0 ± 0.6% and 92.8 ± 0.7%). The research results and proposed model of this study are significant for research into exoskeleton control based on sEMG signals.