Two-dimensional identification of lower limb gait features based on the variational modal decomposition of sEMG signal and convolutional neural network

Background: Gait feature recognition is crucial to improve the efficiency and coordination of exoskeleton assistance. The recognition methods based on surface electromyographic (sEMG) signals are popular. However, the recognition accuracy of these methods is poor due to ignoring the correlation of the time series of sEMG signals. Therefore, this paper proposes a two-dimensional recognition method of lower limb gait features based on sEMG signal decomposition under multiple motion modes to improve the accuracy and robustness of gait recognition. Methods: First, in order to obtain gait information of human lower limbs, gait experiments in different motion modes are carried out using the sEMG acquisition system with 7 channels. Then, the gait dataset of human lower limbs is expanded and transformed using the variational modal decomposition (VMD) algorithm and Gramian Angular Field (GAF). The processing not only enhances the data, improves the learning ability of classifiers and avoid the overfitting during the training of the convolutional neural network (CNN), but also effectively utilizes the feature extraction capability of the CNN and preserves the temporal correlation of the EMG. Finally, the gait features in four motion modes are recognized using the processed sEMG data and trained ResNet network. Results: The recognition results show that the proposed method in this paper has the highest recognition rate under four motion modes compared to BP neural network and CNN network based on original sEMG signal. This research is helpful for the effective implementation of intelligent control strategies and the coordination of human-exoskeleton system.