For the problems that there were a large number of redundant connections between the variables caused by data-based methods and the information loss between the variables caused by knowledge-based methods in the field of the propagation path identification on faults, a new method for the propagation path identification on faults that combined Hidden Markov Model (HMM) and Bayesian Network (BN) was proposed. First, the knowledge of the hot strip rolling process was constructed as a qualitative BN structure, and the dimensionality of the data in the hot strip rolling process was reduced by the principal component analysis method to obtain the observation sequence required for training model. Then, based on the normal historical data after dimensionality reduction and its log-likelihood value, the conditional probability table was established for BN to identify the propagation path. Finally, the fault data and their log-likelihood values were used as the likelihood evidence for BN to identify the fault propagation path. The experimental results show that this method can accurately locate the six variables where the fault occurs, without any misdiagnosis or missed detection, and can accurately identify the propagation path of faults.