Analysis of the Health Status of Railway Vehicle Bearings Based on Improved HDP-HMM

Abstract

In order to solve the problem that state number of the hidden Markov model (HMM) must be specified in advance and the convergence result of HDP-HMM is sensitive to hyperparameters, a bearing health status analysis model based on improved HDP-HMM is proposed in this paper. Based on the Hierarchical Dirichlet Process (HDP) and Hidden Markov Models, the model uses the nonparametric properties of the hierarchical Dirichlet process to infer the number of hidden states, compensates for the defects of HMM, and utilizes Bayesian Optimization and Mann-Kendall criteria optimize its hyperparameters. At the same time, considering the ergodic topology of the traditional HDP-HMM is not suitable for the timing monitoring data of the bearings, we convert the HDPHMM topology into a left-to-right mode, which is more suitable for the needs of health status analysis. In addition, taking the nonlinear characteristics of the performance degradation process of railway vehicle bearings into consideration, we use GKPCA (greedy kernel principal components analysis) to extract features of bearing degradation. Finally, the model is verified by using the monitoring data collected during the train running such as the bearing temperatures. The experimental results show that the proposed HDP-HMM can effectively identify multiple health status of the railway vehicle bearings and has reliable performance. It provides an important basis for the state repair of the railway vehicles bearings under actual conditions.