Bayesian network based covariance monitoring

Covariance monitoring is an important problem in multivariate process control. Quite a few methods have been developed in the literature, among which a popular one called the TB method is based on regressions. The TB method is computationally intensive and has a limited capability for diagnosing the root causes of changes in the covariance matrix. This paper proposes a new method that integrates a Bayesian network (BN) for process causal relationship identification with covariance monitoring. Extensive simulation studies are performed, demonstrating that the proposed method outperforms the TB method in covariance monitoring and root cause diagnosis. Moreover, theoretical analysis is performed on the risks of the proposed method posed by the uncertainty in BN structure identification from observational data.