In order to extract the circuit breaker fault characteristics more accurately and get more reliable fault diagnosis results, this paper proposed a method based on empirical wavelet transform and relevance vector machine applied to the vibration signal. Different fault vibration signals should be extracted firstly. Setting thresholds to initialize the segments of the signal in the frequency domain, and multiple modes with limited bandwidth can be obtained by EWT decomposition. Then calculate the sample entropies of the modes which regard as the import vector of RVM. Finally, build various models of different fault to diagnose the test sample. Compared to other methods, this method has a higher fault diagnosis rate and faster recognition speed.