A classification-based framework for trustworthiness assessment of quantitative risk analysis. (November 2017)

• Record Type: Journal Article
• Authors: Zeng, Zhiguo; Zio, Enrico
• Is Part Of: Safety science. Volume 99: Part B (2017)
• Page Start: 215
• Page End: 226
• Publication Date: 2017-11
• Subjects: Quantitative Risk Analysis (QRA) -- Validity -- Reliability -- Trustworthiness -- Naive Bayes classifer; Industrial accidents -- Periodicals; Accident Prevention -- Periodicals; Safety -- Periodicals; Travail -- Accidents -- Périodiques; 363.11
• DOI: 10.1016/j.ssci.2017.04.001
• Languages: English
• ISSNs: 0925-7535

Abstract

Highlights: An assessment framework is developed for trustworthiness of QRA. A Naive Bayes classifier is developed to assess the trustworthiness of QRA. Consistency of training data is checked by developing a statistical hypothesis testing. Abstract: In this paper, we develop a classification-based method for the assessment of the trustworthiness of Quantitative Risk Analysis (QRA). The QRA trustworthiness is assumed to be determined by the quality of the QRA process. Six quality criteria, i.e., completeness of documentations, understanding of problem settings, coverage of accident scenarios, appropriateness of analysis methods, quality of input data, accuracy of risk calculation, are identified as the factors most influencing the trustworthiness. The assessment is, then, formulated as a classification problem, solved by a Naive Bayes Classifier (NBC) constructed based on a set of training data, whose trustworthiness is given by experts. NBC learns the expert's assessment from the training data: therefore, once constructed, the NBC can be used to assess the trustworthiness of QRAs other than the training data. Leave-one-out cross validation is applied to validate the accuracy of the developed classifier. A stochastic hypothesis testing-based approach is also developed to check the consistency of the training data. The performance of the developed methods is tested for ten artificially generated scenarios. The results demonstrate that the developed framework is able to accurately mimic a variety of expert behaviors in assessing the trustworthiness of QRA.