ABSTRACT: In order to avoid the practical problems that traditional wavelet energy spectrum as the characteristics is insufficient for the diagnosis infer in bearings fault detection, a novel fault detection method based on the wavelet histogram signatures which capture all the first order statistics using a model based on hidden Markov model (HMM) is presented. In this approach, the statistics features of the multi-scale wavelet coefficients generated by the wavelet decomposition of the signals are analyzed, the detail wavelet histogram of the bearing vibration signals can be modeled by the hidden Markov model. The parameters of this model as diagnosis features are introduced to completely describe the wavelet coefficients’ first-order statistics. The scale and shape parameters of the model are estimated by the maximum likelihood method. Comparison of the performance of detection of the proposed approach with the method based on the wavelet energy spectrum and wavelet energy spectrum entropy is experimented. The results show the relative effectiveness of the introduced feature sets in the detection of the bearing conditions with some concluding remarks. The effects of the wavelet base selection, window width and classifier on the proposed method are conducted in the experiments, which evaluate the stability and robustness of the proposed method.