Monitoring and Diagnosis for Multi-Mode Processes with Varying Operating Parameters: A Covariate-Adjusted Mixture Bayesian Network Approach

Process monitoring and diagnosis are crucial in many industrial applications for the timely detection of anomalies and the identification of responsible process variables. Over the past decade, Bayesian network-based methods have attracted increasing interest due to their diagnostic capabilities and compatibility with multivariate control charts. However, existing methods fail to account for process heterogeneity caused by both unobserved operational modes and observed operating parameters. To address this issue, we propose the Covariate-adjusted Mixture Bayesian Network (CaMBN) model. It uses a nonparametric mixture model to capture heterogeneous correlation structures among process variables across different operational modes, while incorporating covariates to adjust in-control parameters. A Generalized Expectation-Maximization (GEM) algorithm is further proposed for model estimation. Based on this model, a likelihood ratio-based chart and a Bayesian inference-based diagnosis procedure are proposed. The effectiveness of the proposed method is demonstrated through comprehensive numerical studies and real-world data applications to harbor tug engines. Supplementary materials for this article are available online.