Novel strategies to assess sparse data in human reliability analysis

Morais, Caroline ORCID: 0000-0002-9329-4110
(2021) Novel strategies to assess sparse data in human reliability analysis. PhD thesis, University of Liverpool.

[img] Text
201153759_March2021.pdf - Unspecified

Download (10MB) | Preview


Major industrial accidents are usually attributed to problems in the interaction of human, technological and organisational factors. Although many of these accidents arealmost impossible to predict, some can be predicted and prevented by techniques dealing with human error assessment. This is the reason why it is expected that comprehensive risk analyses use a technique known as human reliability analysis. It usually relies on data elicited from experts with operational knowledge, or empirically collected from simulated scenarios, records of near-misses, and major accident reports. The present research proposes the use of MATA-D (Multi-attribute Technological Accidents Dataset), which is based on major accident investigation reports, thus potentially capturing a more realistic relationship between human erroneous actions and technological and organizational factors that shape human performance. Currently, the most recommended probabilistic tool to model human reliability data is the Bayesian network. However, the assessment of its conditional probability tables (CPTs) requires enough data to describe all possible conditions dictated by the model. However, despite increasing collection efforts of empirical human reliability data, the available databases are still insufficient to fulfil conditional probability tables, especially for models where each variable is conditioned on many others. In these cases, the most common solution relies on the adoption of expert elicitation to fill in the missing combinations. This research has been focused on developing strategies to enable empirical data-driven human reliability analysis, such as precise and imprecise probability tools to tackle epistemic uncertainty inherent to such databases. The used probabilistic tools are Bayesian and credal network, the latter to tackle missing data in conditional probability tables. Using credal networks the prediction analysis depicts results with interval probabilities rather than point values measuring the effect of missing-data variables. As taking decisions is more difficult when comparing intervals then point-values, a decision-making strategy is suggested to unveil the most relevant variables for risk reduction in presence of imprecision. The results support the hypothesis that realistic uncertainty depiction implies less conservative human reliability analysis and improves risk communication between assessors and decision-makers. Finally, a natural language processing technique based on machine-learning has been developed to extract and classify new accident reports in order to collect new data for MATA-D. This aims to decrease the number of missing combinations in CPTs. A constant collection of new data for this dataset aims not only to decrease epistemic uncertainty in human reliability data but also to timely update models, reflecting changes in human behaviour due to evolving technology and organisational arrangements. The automated approach, called the virtual human factors classifier, is able to classify a new report more than one thousand times faster than a human being. Future developments are discussed, such as the strategy to compute reliability analysis with confidence, by using credal networks and c-boxes to tackle different and sometimes very small sample sizes in a database.

Item Type: Thesis (PhD)
Divisions: Faculty of Science and Engineering > School of Engineering
Depositing User: Symplectic Admin
Date Deposited: 08 Feb 2022 16:37
Last Modified: 01 Jan 2024 02:31
DOI: 10.17638/03140945