Surface electromyography decoding for continuous movement of human lower limb during walking

Abstract

To estimate the continuous movement of human lower limb during walking, a regression model which relates the surface electromyography (EMG) and the movement variables of the lower limb joints is constructed. The joint movement angles of lower limb are calculated accurately based on optical motion capture system, then the surface EMG signals are sampled from the main muscles directly concerned with the lower limb motion; the muscle activities are extracted, and a deep auto-encoder (DAE) network with restricted Boltzmann machines (RBM) is realized, by which the multi-channel processed surface EMG signals are encoded in low dimensional space and the optimal features are extracted. The nonlinear model mapping the EMG features to sagittal surface movement angles is established with back propagation (BP) neural network. Extensive experiments indicate that the features extracted with the deep auto-encoder (DAE) network are outperformed principal components analysis (PCA); the movement angles of lower limb joints can be estimated continuously and precisely with the regression models and the mean square error (MSE) between the estimated values and real values is reduced by 25%-35% compared with the traditional method. The proposed strategy is expected to develop human-machine interaction interface technology for the achievement of continuous bioelectric control and the improvement of motion stability between human and machine, especially for lower limb wearable intelligent equipment.