Projects per year
Causal graphs are models that can be used for plantwide diagnosis of chemical process systems as extensions to existing alarm management. For safety-critical applications, valid and reliable models become crucial for producing trustworthy diagnoses. As part of this dissertation, a method is proposed for validating the causal structure of Multilevel Flow Modelling models. Previous relevant published literature for the Multilevel Flow Modelling (MFM) method is reviewed, and on this basis the first method for quantifying MFM model validity has been proposed. A set of rules are proposed based on the understanding of the model causality. The rules can be used for testing, comparing and revising the models to improve the validity of the structure. The prediction accuracy metric is used for measuring the correspondence of the tested model behaviour to a validation set. The metric can be used for different model parts to identify which parts require revisions to improve the model. A method is proposed for causal analysis based on reviewed methods applied to process systems. Unlike the majority of methods, the proposed method aims to provide a high level of transparency of the results, given the application is safety related. For this reason, the method, unlike the majority of previous approaches, simulates the system with process condition uncertainty, and applies interventions of varying magnitude to observe effects. The method, is to the knowledge of the author the first to use Monte Carlo Simulations for causal analysis of process systems combined with uncertainty and sensitivity analysis for assessing the robustness of the analysed results. The methods for causal analysis and validation were applied to a high-pressure separation system. Based on a causal analysis, the MFM model validity has been quantified. The quantification has been used for revising the model to improve the quantified validity by almost 100 %. The robustness of the analysed causality was evaluated for a defined process window by applying uncertainty and sensitivity analysis methods. The causal analysis method enables verification of the results’ robustness, and the analysis showed that the majority of the results are robust. The analysis further showed that linear sensitivity analysis methods can be insufficient for process systems that exhibit a nonlinear behaviour. In addition, the deterministic system behaviour inferred by Multilevel Flow Modelling was found to oversimplify the behaviour of the physical system. Identifying subsets of the process window where the system exhibits a deterministic behaviour has therefore been proposed, such that a single MFM model does not represent the entire process window, but multiple models represent different subsets. When revising models to improve one model part, the reasoning from other model parts is influenced and suboptimal models are thereby produced. To address the issue of producing suboptimal models, it has been proposed to simulate various failure modes using Morris screening. The risk of each simulation can then be ranked for prioritising certain model behaviour. It was also proposed to evaluate the performance of causal graphs by inferring causes for simulated faults. For this purpose, the simulations of different failure modes should be diagnosed by the MFM model, although this and the ranking has not been proven as part of the dissertation. By emulating a fault on physical pilot plant for produced water treatment, the diagnostic performance was evaluated of an MFM model by assessing whether the correct cause was inferred or not. The study, however, showed that the model may produce more potential causes than the number of alarms used as model input, hereby increasing the amount of information. A subjective method was thus proposed for assessing whether alarms were plausible for each inferred cause. In this way, the actual fault could be identified as the cause by using the model for diagnosis.
|Publisher||Technical University of Denmark|
|Number of pages||231|
|Publication status||Published - 2020|