A new approach for identifying the wear degree of gear based on Mixture of Gaussians Bayesian Belief Network (MoG-BBN) was proposed. The inference algorithm was established through combining the variable elimination algorithm with expectation maximization algorithm. Then, one can recognize the gearbox wear states through identifying the hidden state of MoG-BBN which best fits the observations. Aiming at the local convergence problem of expectation maximization, a modified algorithm was proposed. According to the non-linear dependencies between features, the curvilinear distance analysis was used for dimension reduction. Finally, the data of gear’s wear experiment were used to demonstrate the proposed methods. The results showed the classification accuracy was 99%.