In order to improve the gait recognition accuracy of prosthetic wearers， a new method of prosthetic gait recognition based on surface electromyogram signal is proposed. In this method， surface electromyogram signal（sEMG） is selected to solve the multi-classification problem in gait recognition of prosthetic wearers. sEMG is a higher order zero crossing analysis （HOC） square representing the characteristics of time series. The method is applied to feature extraction of EMG signals under different gaits， and multiple classification steps are established by combining with relevance vector machine （RVM）. Bat algorithm （BA） is used to optimize kernel function parameters of RVM classifier. Experimental results show that the proposed method is better than particle swarm optimization-relevance vector machines （PSO-RVM） and RVM for walking on the ground， upstairs， downstairs， uphill and downhill of five kinds of gaits.The recognition accuracy is higher than PSO-RVM and RVM and other methods.