Enhancing safety and human reliability through data-driven and NLP innovations

Johnson, Karl

Thesis

Enhancing safety and human reliability through data-driven and NLP innovations

Download PDF

Creator

Johnson, Karl

Rights statement

Strathclyde Thesis Copyright

Awarding institution

University of Strathclyde

Date of award

2025

Thesis identifier

T17424

Person Identifier (Local)

202072238

Qualification Level

Doctoral (Postgraduate)

Qualification Name

Doctor of Philosophy (PhD)

Department, School or Faculty

Department of Civil and Environmental Engineering

Abstract

This thesis explores the development and application of innovative algorithmic, data-driven, machine learning, and natural language processing tools designed to enhance human reliability analysis in critical sectors such as nuclear power, aviation, and oil and gas. Motivated by identified challenges and opportunities in these industries, a suite of advanced tools was created to address key aspects of safety analysis and management. Presented in this work are six tools, the first- and second- generation Virtual Human Factors Classifiers, the Human-Centric Summarizer, the High-Potential Violation Trigger Identification tool, the Ambiguity Identifier and finally the Human Factors Causal Relationships tool. The Virtual Human Factors Classifiers were designed to automatically read analyze accident reports to classify the contributing factors. The primary motivation for this development was the expansion of a human reliability analysis database (MATA-D, Multiattribute Technological Accidents Dataset), to provide the additional data necessary to address the issue of missing information and reduce the uncertainty of human error probability models. The tools have also demonstrated additional applications such as aiding assessors in their reviews of accidents and informing the procedure design process. Complimentarily the Human-Centric Summarizer was developed to distill lengthy accident reports into high-quality concise summaries, that emphasize the human role in the incident. The summarizer serves a dual purpose. Firstly, it aids researchers and safety professionals in rapidly grasping each report, as well as any models based on the incident, without delving into the pages of detailed reports. Secondly, it assists in maintaining and updating the MATA-D. The summaries generated provide a quick reference to the key points of each incident, facilitating easier analysis and review of performance shaping factor classification. In high-risk industrial environments, the clarity and accuracy of standard operating procedures are critical for ensuring safety and regulatory compliance. The presence of ambiguities in standard operating procedures can lead to misunderstandings, errors, and increased risks. While violations of procedural directives can significantly contribute to catastrophic outcomes. To address these issues, two additional tools are introduced that leverage both rule-based and machine learning methodologies in natural language processing to evaluate the quality of standard operating procedure documents. The High-Potential Violation Trigger Identification tool identifies directives within procedural guides that when violated pose a high-risk potential. And the Ambiguity Identifier has been designed to detect various types of ambiguities and misleading steps within procedure guides. By addressing these linguistic and procedural discrepancies, the tools aim to enhance the clarity and applicability of standard operating procedures, ultimately improving adherence and reducing risks in complex operational settings. The final tool presented in this work is the Human Factors Causal Relationships Tool. It leverages data collected through the MATA-D to identify causal relationships among performance shaping factors. This tool is designed to reduce reliance on expert judgment in the development of human error models, thereby helping to mitigate concerns related to subjectivity and bias. Case studies are presented for each tool, demonstrating their real-world utility and effectiveness in critical industry contexts. This thesis highlights the potential of data-driven and natural language processing approaches to revolutionize human reliability analysis practices, ultimately enhancing safety across critical industries.

Advisor / supervisor