Thesis

Natural language processing for automatic classification of human and causal factors from maritime safety reports

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2026
Thesis identifier
  • T18210
Person Identifier (Local)
  • 202270657
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Maritime and offshore operations remain vulnerable to accidents and near misses in which human and organisational factors play a central role. Although narrative incident reports contain rich qualitative evidence about these contributory mechanisms, current analytical practices rely heavily on manual review, making large-scale learning slow, inconsistent, and resource intensive. Recent advances in natural language processing (NLP) offer new opportunities to exploit this data, yet applications in the maritime and offshore wind sectors remain fragmented, rarely taxonomy-aligned, and seldom extend to near-miss and Safety-II perspectives. This thesis develops and evaluates a unified NLP framework for extracting human and causal factors from maritime and offshore incident narrative summaries under both low-information (short, unstructured summaries) and high-information (structured, taxonomy-labelled) conditions. The high-information branch introduces SHIELD-NLP, a machine-learning-adapted version of the SHIELD taxonomy, combined with SeaSafeBERT, a domain-adapted transformer fine-tuned on maritime accident reports, a hybrid semantic-lexical feature representation, and a hierarchical, probability-calibrated logic regression classifier. Association rule mining is then used to make interaction patterns between human, organisational, and environmental factors across accident types. The low-information branch applies BERTopic to short narratives, supported by GPT-based topic labelling and expert validation, to derive themes, temporal patterns, and links between incident types and preventive actions. Three case studies demonstrate the framework’s applicability. In the taxonomy-coded maritime accident reports, the pipeline achieves strong performance for several high-value SHIELD factors with F1-scores reaching 1 for visual/auditory detection and personnel/training factors, 0.92 for inadequate risk assessment, and 0.85 for inadequate leadership/supervision and failures in information transmission. This enables reliable automatic or semi-automatic coding of failures in detection, communication, planning, leadership, and risk assessment. Association rules reveal shared vulnerabilities across accident types, including links between adverse operational conditions and decision-making failures; for example, incompatible workspaces were associated with failures in information transmission (confidence =0.61, lift=1.76), while environmental and technological difficulties were consistently associated with incorrect decisions or plans (confidence=0.60-0.62). In the ferry near-miss database (4,360 reports), BERTopic revealed 76 incident themes and 87 prevention themes, extending the operators' taxonomy with 34% of incident-topic clusters represented patterns not captured by existing taxonomy, uncovering seasonal patterns in passenger/traffic and technical risks, and systematically links identified topics to documented prevention actions. This provided empirical evidence of adaptive capacities and resilient practices through which ferry operations are maintained under varying operational conditions, including crew vigilance, teamwork and ship-shore coordination. Analysis of 5,403 offshore incidents from 2017-2021 identified 85 distinct topics using a combination of topic modelling, and association rule mining on pre-structured multi-label outcomes (e.g. high-potential events, dropped objects, ERME). The findings highlight recurring risks in lifting operations, vessel transfers, working at heights, and PPE non-compliance, as well as characteristics of injury patterns and environmental contributors. ARM further identified workplace safety violations as an important risk-related themes, with 94% of occurrences associated with hazard outcomes and 73% associated with high-potential incidents. The theme was also associated with working at heights 72% of the time, highlighting procedural compliance during elevated work as a priority for targeted intervention. A prototype decision-support interface illustrates how these components can be operationalised into a human-in-the-loop safety intelligence tool, providing calibrated factor predictions, causal pattern visualisations, and near-miss learning support. The principal contribution is an integrated and explainable NLP framework that accommodates both taxonomy-rich and information-sparse safety data, thereby supporting conventional accident analysis alongside Safety-II-oriented learning from near misses and successful preventive actions. By transforming large volumes of previously underutilised narrative safety data into structured and interpretable safety intelligence, the framework reduces reliance on exclusively manual analysis and enables safety practitioners to identify recurring human and organisational vulnerabilities, detect emerging or previously uncategorised operational risks, prioritise targeted interventions, and systematically learn from effective preventive actions at scale. This provides a practical pathway for embedding NLP within organisational safety-management processes while retaining expert oversight of automated outputs.
Advisor / supervisor
  • Kurt, R. E.
  • Turan, Osman
Resource Type
DOI

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T18210 2026-10-14 Public Download