A Multi-level Bayesian Network Based on Causality Analysis for Fault Diagnosis of Nonstationary Processes

Abstract

Finding the fault propagation pathways for nonstationary processes is a hard mission resulting from the propagation pathways determined by causality analysis can be twisted by nonstationary characteristics. Therefore, exploring an accurate causal relationship has drawn a special attention when an industrial process has nonstationary characteristics. However, the causal relationships were described inaccurately in previous work due to the assumption where all the processes are stationary. In the present work, a two-step causality construction method is developed to find the correct causal relationship for industrial processes with nonstationary characteristics. First, the cointegration relationship is explored for nonstationary variables and the local causality construction method is proposed for nonstationary variables to extract their strong causal relationship. Next, a global causality construction method is designed by stationary features, which are regarded as a bridge connecting stationary variables and nonstationary variables. When the global causality structure is done, a multi-level Bayesian Network (BN) model based on this structure is presented to locate the root fault cause and to depict fault propagation pathways for nonstationary processes. The performance of the presented model is verified by the Tennessee Eastman process.