This paper proposes an intelligent diagnosis method for gearbox using local mean decomposition and discrete hidden Markov models, including local mean decomposition, the energy difference spectrum of singular value, multiscale sample entropy, and the discrete hidden Markov model. How to extract feature information effectively and identify the fault type is key to making a diagnosis in the presence of strong noise. Combined with the Kurtosis criterion and correlation coefficient, the product function that contains the main characteristic frequency is filtered out by local mean decomposition. Next, the filtered local mean decompositions are used to construct the Hankel matrix and complete singular value decomposition. The denoised and reconstructed signals are achieved by an energy difference spectrum of singular value. Furthermore, the feature information after denoising is extracted by multiscale sample entropy. After combining the discrete hidden Markov models, the mechanical condition is identified. Practical examples of diagnoses for four gear types used in the gearbox can accurately identify the gear types, and the recognition rates of the various types are above 92%. The experiments shown here verify the effectiveness of the method proposed in this paper.