Rolling element bearing is a crucial component in rotating machinery, so the diagnosis of bearing fault has attracted a lot of attention in both scientific and application areas. The objective of fault diagnosis can be viewed as a separation of informative events concerning different types of fault in noisy measurements. Therefore, it is appealing to model the raw signal as a linear combination of few components with the prior knowledge and assumption. Hidden Markov model (HMM) is a probabilistic model of joint probability of a collection of random variables which represent the hidden states as state variables given the observation sequence. In this paper, an exploring study in rolling element bearing diagnostics based on HMM is investigated and a new fault separation scheme is proposed. We analyse the performance of the proposed scheme through numerical experiments and demonstrate its potentiality in industrial applications.