In order to achieve autonomous lower limb rehabilitation training for stroke patients, exoskeletons are gradually combined with BCI(brain computer interface). Still, multi-class MI(motor imagery) EEG(electroencephalogram) has always had characteristics of difficulty in extraction and low recognition accuracy. A multi-classification optimized support vector machine algorithm for EEG signals based on WICA(wavelet independent component analysis) and CSP(common spatial patterns) was proposed. PSO(particle swarm optimization) was used to train a support vector machine for classification. It proves that the algorithm can effectively extract EEG features and has a better recognition effect of motor imagery EEG signals. The feasibility of online and real-time EEG control was verified by combining motor imagery and exoskeleton devices.