Feature extraction is the key to better application of muscle-machine interface. The selection rationality of a single index is poor, so it is usually necessary to adopt the method of multi-index integration. However, this method will not consider the physical significance or the weight imbalance when selecting indicators, resulting in inappropriate selection, which directly affects the signal processing performance. Based on the idea of multiple eigenvalues obtained by different simultaneous time and frequency feature extraction methods, a feature subset algorithm based on variation coefficient is proposed, and the fitting index system is established according to different time-frequency domain indexes. In the experiment of recognition of 6 kinds of gestures on 10 subjects, the algorithm can effectively reflect the difference between gestures, and the average success rate of recognition of gestures can reach 98.5%.