Aiming at the needs of bionic prosthetic motion recognition, a lower limb motion recognition method based on time-frequency generalized S transform and VL-MOBP neural network was proposed. First, time-frequency generalized S-transform was used to measure 4 kinds of surface electromyographic signals and knee flexion of the lower extremities of 22 male subjects aged between 20 and 40 years old, between 170 cm and 185 cm tall and weight between 50 kg and 75 kg. Using multi-resolution analysis of the frequency signal to obtain the time-frequency cumulative characteristic curve of the signal when the time and frequency resolution were good, then extracting the mean and standard deviation of the amplitude of the time-frequency cumulative characteristic curve as the feature vector, and using the VL-MOBP neural network to recognize the three movements of human lower limbs: Walking, standing, and sitting. The experimental results showed that the proposed lower limb movement recognition method can achieve good recognition results, with an average recognition accuracy of 96. 67%, which is about 56% higher than the wavelet transform and about 36% higher than the short-time Fourier transform. Effectiveness in motion recognition has been verified.