Decoding of lower limb continuous movement intention from multi-channel sEMG and design of adaptive exoskeleton controller

The utilization of robot-assisted rehabilitation training has shown promising results in promoting motor recovery in neurologically impaired patients. However, current methods are limited to predefined desired trajectories, disregarding individual variations. Therefore, this article introduces a subject-based active rehabilitation training framework for lower limb daily activities, focusing on integrating intention perception and compliance control. To accurately interpret human intention, this study proposes a divided Spatial-temporal Attention EMG Network (dSTA-EMGNet) model for the time-advancing prediction of the trajectory of the knee joint with multi-channel surface electromyographic signals. Subsequently, an admittance adaptive control scheme is formulated based on the Nonlinear Disturbance Observer (NDO) technique. Initially, an admittance model is utilized to ensure compliant behavior of the exoskeleton, with the NDO estimating the real-time torque resulting from human-exoskeleton interaction. Furthermore, a novel adaptive controller employing a radial basis function neural network is devised to address the feedforward compensation of dynamic uncertainties. Experimental findings indicate that the proposed dSTA-EMGNet exhibits superior predictive capabilities, as evidenced by a mean value of the coefficient of determination exceeding 0.982 ± 0.007 and an average absolute error lower than 2.597°±0.742°. Furthermore, the implemented control scheme shows commendable motion-tracking proficiency and exceptional compliance, affirming the efficacy of the proposed framework.