A training algorithm, state-weighted synthesis Continuous Gaussian mixture Hidden Markov Model(CGHMM) synthesized by several CGHMMs, is presented according to the states distribution of observations. It can solve the problem that it was difficult to provide enough training data due to too many parameters in HMM model. The proposed method is applied in bearing fault diagnosis, and the average training time of 12.86 s, diagnosis time of 0.189 s, and diagnosis rate of 90% are obtained. The method based on state-weighted synthesis CGHMM is effective and feasible in bearing fault diagnosis and has a great prospect for application.