A technique of feature extraction based on vibration phase diagram for roller of rolling mill was discussed. Basic theory of roller fault diagnosis based on vibration phase diagram was presented, causes for phase diagram under different conditions was analyzed and a method of computing the eigenvalue for distribution density matrix was provided. To same one roller vibration under identical environment, the cooperation of phase point probability distribution is only relelation to the density of probability distribution, and unconcerned with noise. It is also different to phase point distribution density of different states. A experiment was conducted on ψ500×2 shrewd roller of rolling mill, the feature of roller signal was extracted based on vibration phase diagram and chose Hidden Markov Model (HMM) to build and identify mould. The experiment shows that this system has better identification rate, and proves that this technique is effective.