Proper initialization of Hidden Markov models for industrial applications

Hidden Markov models (HMMs) are widely employed in the field of industrial applications such as machine maintenance. However, how to improve the effectiveness and efficiency of HMM-based approach is still an open question. The traditional HMMs learning method (e.g. the Baum-Welch algorithm) starts from an initial model with pre-defined topology and randomly-chosen parameters, and iteratively updates the model parameters until convergence. Thus, there is the risk of falling into local optima and low convergence speed because of wrongly defined number of hidden states and randomness of initial parameters. In this paper, we proposed a Segmentation and Clustering (SnC) based initialization method for the Baum-Welch algorithm to approximately estimate the number of hidden states and the model parameters for HMMs. The SnC approach was validated on both synthetic and real industrial data.