Abstract: In order to quickly and accurately classify gear failure, combined with pattern classification capabilities of Hidden Markov model (HMM), a novel fault diagnosis approach based on the Zoom Spectrum and Hidden Morkov Model is proposed. First, time-domain synchronous average signal of an interested gear is extracted from original signal and zoom spectrum is analyzed using the TSA signal. Then, the side-frequency bands of fundamental frequency and its harmonious amplitude are processed as a feature vector, which was proved sensitive in fault diagnosis in previous research. The HMMs are trained by maximizing the probability of given feature vectors and the gear failure types are identified by comparing the logarithmic likelihood probability value. Finally, the performance of the fault diagnosis scheme is validated using experimental data collected from gearbox.