The mechanical fault diagnosis results of the high voltage circuit breakers (HVCBs) are mainly determined by the feature vector and classifier used. In order to obtain more remarkable characteristics of signals and a robust classifier which is suitable for small sample classification, in this paper, a new mechanical fault diagnosis method is proposed. Firstly, the vibration signals of HVCBs are collected by a designed acquisition system, and the noise of signals is eliminated by a soft threshold de-noising method. Secondly, the empirical wavelet transform (EWT) is adopted to decompose the signals into a series of physically meaningful modes, and then, the improved time-frequency entropy (ITFE) method is used to extract the characteristics of the vibration signals. Finally, a generalized regression neural network (GRNN) is used to identify four types of vibration signals of HVCBs, while the smoothing parameter δ of GRNN is optimized by a loop traversal method. The experimental results show that by using this optimal classifier for fault diagnosis, the proposed fault diagnosis method has the better generalization performance and the recognition rate of unknown samples is over 95%, and the signal features obtained by the ITFE method are more significant than those of the traditional TFE method.