Abstract: Transformer voice print signals, which can be divided into transient voice print and steady voice print under different operating conditions, contain rich information. This paper analyzed the vibration mechanism and excavated the voice print characteristics to realize the fault identification based on the shedding fault of the end pad in a transformer. Firstly, a mass-spring-damping model was established to analyze the vibration mode and development law of the fault. Secondly, short time Fourier transform（STFT） was applied to the voice print signal, and Mel filter bank was used to realize the compressed sensing of time spectrum. Thirdly, the concept of signal stability was introduced based on time series spectrum entropy feature extraction algorithm. Finally, we calculated the stability of the voice prints signal data set collected from 162 transformers and the core loosening fault signal data set collected from the test to obtain the stability distribution. With the difference between the transient voice print distribution in end pad shedding and the steady state voice print distribution in normal condition, the end pad shedding fault can be identified by stability algorithm.