Applying dynamic Bayesian Networks to process monitoring

[摘要] ENGLISH ABSTRACT: In efforts to reduce the impact of human error on the operation of chemical andmineral processing plants, reliable process monitoring solutions attempt to assistplant operators and engineers to detect and diagnose process faults before significantloss is incurred. An existing solution, the traditional multivariate statistical processmonitoring (MSPM) approach, is able to reliably detect abnormal process behaviourbut struggles to unambiguously identify the root cause of the abnormal behaviour.It was identified that this is caused by a lack of incorporation of existing processknowledge into the framework of the MSPM approach.It was proposed to investigate a different fault diagnosis approach which directlyincorporates process knowledge into its framework. Lerner et al. (2000) and Lerner(2002) present such an approach, using probabilistic methods to infer processbehaviour given a particular process model. This model is in the form of a dynamicBayesian network (DBN), and would contain various models which each describeparticular process behaviour given information about the operational status of variousprocess components. In particular, these DBN models were able to describe normalprocess behaviour in addition to highly specific abnormal process behaviour causedby, for instance, a sensor fault or a blocked pipe. Using optimised methods, theauthors could then use a DBN model to make predictions about process behaviour andinfer, given observation of actual process behaviour, which combination of componentstatuses best describe that observation. Therefore, solving the fault diagnosis problemcould be reduced to performing inference in a DBN using this approach.A probabilistic fault diagnosis (PD) approach based on Lerner et al. (2000) andLerner (2002) was therefore implemented and investigated in this thesis. A survey ofrecent DBN-based PD approaches was also performed, and it was determined thatrelatively little research had been done on the topic. Furthermore, published resultspresenting fault diagnosis performance for DBN-based PD approaches were typicallyfound to be useless for meaningful comparison with a traditional MSPM approach.In this regard, this thesis aimed to investigate the usefulness of the PD approach incomparison to the MSPM approach, while providing useful fault diagnosis performancemetrics to facilitate comparison with other fault diagnosis approaches.The PD approach tested in this research also extended upon Lerner et al. (2000) and Lerner (2002) by including models for regulatory control systems and recyclestreams based on the work by Yu and Rashid (2013). Additionally, from the samepaper, the concept of abnormality likelihood index (ALI) was implemented in the PDapproach. This enabled the PD approach to function more similarly to the MSPMapproach, facilitating direct comparison.Generally, it was found that the PD approach could provide competitive faultdetection when compared with the MSPM approach. However, this was at the cost ofreal-time fault detection as well as longer detection delay for incipient faults. On theother hand, it was found that the PD approach performed better at root cause analysisthan the MSPM approach. In particular, the PD approach typically provided betterisolation for the root cause of fault conditions.Despite some issues, similar results were observed for the PD approach whenscaling up to larger processes. Nonetheless, these issues may be addressed withadditional research, further improving the capabilities of the PD approach. Therefore,it was concluded that the PD approach is useful for fault diagnosis and should beinvestigated further in future research.

[发布机构] Stellenbosch University