Fault Detection And Root Cause Diagnosis Using Dynamic Bayesian Network

This thesis presents two real time process fault detection and diagnosis (FDD) techniques incorporating process data and prior knowledge. Unlike supervised monitoring techniques, both these methods can perform without having any prior information of a fault. In the first part of this research, a hybrid methodology is developed combining principal component analysis (PCA), Bayesian network (BN) and multiple uncertain (likelihood) evidence to improve the diagnostic capacity of PCA and existing PCA-BN schemes with hard evidence based updating. A dynamic BN (DBN) based FDD methodology is proposed in the later part of this work which provides detection and accurate diagnosis by a single tool. Furthermore, fault propagation pathway is analyzed using the predictive feature of a BN and cause-effect relationships among the process variables. Proposed frameworks are successfully validated by applying to several process models.