Automating causal analysis for online diagnosis and planning in complex industrial processes

This thesis contributes to the development of decision support systems for large industrial processes that depend on human supervisory control, such as production plants in the chemical, petrochemical and energy industries. Supervision of such systems requires the utilization of process knowledge for the contextualization of plant information and the formulation of viable operating strategies. These tasks demand constant attention from human operators on site and it has been shown that especially demanding situations can result in cognitive overload, which impairs the operators ability to make informed decisions, leading to human error. The contributions in this thesis present methods that contextualize process data and suggest operating strategies automatically or on-demand in order to focus the operators attention and improve human performance. The chosen decision-support framework leverages information contained in causal process models, which encode the causal relations between process variables, operational states and production goals. Following a summary of the state of the art of causal process modeling, extensions related to causal propagation, diagnosis and planning are presented. Propagation algorithms are evaluated based on their utility for specific applications within diagnosis and planning. Ranking strategies for causal diagnosis, specifically root-cause analysis, are extended to incorporate more parameters to improve diagnostic accuracy. Regarding causal model-based planning, existing appraoches are extended by formalizing a semi-quantitative planning framework. The latter enables feedback dependent process re-configuration based on qualitative information about state transitions, which can be generated using causal propagation. The proposed methods for root-cause analysis and planning are evaluated using relevant case-studies. The majority of the presented causal models are build using Multilevel Flow Modeling (MFM), a functional modeling language that facilitates the representation and analysis of complex industrial processes. Two of the presented contributions extend this modeling language; The first introduces the explicit inclusion of operator control actions in MFM and the second proposes the automatic extraction of causal models in the form signed directed graphs. An open-source development and testing framework for causal analysis is detailed in two further contributions that feature a dataset and an improved simulator for the Tennessee Eastman process. The presented framework promises significant improvements for rapid testing and development of causal analysis techniques and generally facilitates research related to fault diagnosis and process re-configuration. The thesis concludes with a discussion of the presented contributions, regarding
academic and industrial relevance, utility and potential extensions.