In order to judge the motor intention of patients with lower limb disorders, the relationship between surface electromyography (SEMG) signal and lower limb joint movement was analyzed through rehabilitation training of exoskeleton. The root mean square (RMS), mean absolute value (MAV), waveform length (WL) and variance (VAR) of SEMG signal are extracted as feature input signals, and the mapping relationship between SEMG signal and lower limb joint angle is established by using extreme learning machine (ELM); the output results are optimized and filtered to reduce the error of the model, and the continuous change of lower limb knee angle is predicted. Compared with the traditional back-propagation neural network and radial basis function neural network, the results show that the extreme learning machine has higher accuracy in predicting the change of lower limb joint angle through SEMG signal.