Document Type

Theses, Ph.D

Disciplines

Computer Sciences, Environmental sciences

Abstract

Alarm systems in process industry control rooms routinely exceed the performance targets set by standards such as EEMUA 191, placing operators under conditions where reliable performance is most difficult to achieve. Predicting how operators respond under such conditions is central to risk management, yet current Human Reliability Assessment (HRA) methods depend on expert judgement that is rarely tested against operational evidence, assume independence among factors known to interact, and do not explicitly represent the cognitive processes through which performance emerges. In Resilience Engineering terms, these methods encode Work-as-Imagined with limited means to assess how far expectations hold when work is actually performed. Data-driven alternatives can capture empirical patterns but offer limited insight into why a prediction is made, a shortcoming where understanding the reasoning behind risk estimates is essential. What is needed is a modelling approach that captures factor interactions, reflects cognitive structure, distinguishes timeliness from effectiveness, confronts theoretical assumptions with empirical data, and remains interpretable to domain experts.

This thesis develops and evaluates Bayesian networks as an explainable AI framework for human performance prediction in alarm management tasks within oil and gas control rooms. A knowledge-based Bayesian network was constructed by integrating the Alarm Initiated Activities model, the Perception–Interpretation–Planning–Execution cognitive framework, and SPAR-H performance shaping factors into a single probabilistic structure whose relationships are explicitly represented and inspectable. A controlled experiment was conducted using a process control room simulator, where 176 participants completed 519 alarm management tasks under varying alarm load and support system configurations. Exploratory analysis revealed substantial variation in success rates across cognitive stages and informed subsequent modelling. Data-driven classifiers established an empirical baseline, and a data-enhanced model was constructed by calibrating the expert-defined parameters with empirical evidence while retaining the network’s inspectable structure.

Comparison across approaches identified where expert assumptions aligned with observed behaviour and where empirical calibration was required, with task complexity emerging as the dominant predictor across all strategies. The data-enhanced model achieved the greatest improvement at later cognitive stages without compromising interpretability. These findings indicate that integration of expert knowledge and empirical evidence within an explainable Bayesian network framework improves predictive accuracy while preserving the transparency required for safety-critical applications, with implications for alarm system design, operator training, and standards implementation.

DOI

https://doi.org/10.21427/83dx-cv04

Funder

European Commission

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Download


Share

COinS