Scholarlink

<p>Autonomous and remotely supervised vessels require explainable safety analytics for compliance, transparency, and trust in mixed human–artificial intelligence operations. We present a prediction-oriented framework that operationalizes the Human Factors Analysis and Classification System (HFACS) as auditable case-level labels and combines it with supervised machine learning to predict and explain human-error types in maritime collisions. From 101 Korea Maritime Safety Tribunal (KMST) written judgments (2014–2023), two trained coders extracted HFACS-relevant contributory factors, logged cross-level linkages for qualitative interpretation, and assigned one dominant HFACS level per case as the prediction target. Using structured case features describing vessel/encounter context and International Regulations for Preventing Collisions at Sea (COLREGs) indicators (including violation assessments), we trained random forest (RF), extreme gradient boosting, support vector machine, and naïve Bayes classifiers. Under stratified holdout evaluation, RF achieved 0.86 accuracy (weighted F1-score = 0.85), with the strongest signals from COLREGs-related violations, visibility, and give-way status. The framework enabled interpretable explanations for frontline categories (Unsafe Acts and Preconditions) but showed reduced separability for latent supervisory and organizational categories, consistent with heterogeneous reporting in KMST narratives. Although this dataset does not support stable inferential dependency/path analysis across HFACS levels, the documented linkages motivate future hierarchical or multi-label dependency modeling using larger, multi-source datasets. Overall, the approach strengthens COLREGs-aware, explainable decision support and aligns with the International Maritime Organization’s electronic navigation strategy for future autonomous bridge and vessel traffic service applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable human error prediction for autonomous vessels: combining human factors analysis and classification system with machine learning

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Do-Hoon Kim

摘要

Autonomous and remotely supervised vessels require explainable safety analytics for compliance, transparency, and trust in mixed human–artificial intelligence operations. We present a prediction-oriented framework that operationalizes the Human Factors Analysis and Classification System (HFACS) as auditable case-level labels and combines it with supervised machine learning to predict and explain human-error types in maritime collisions. From 101 Korea Maritime Safety Tribunal (KMST) written judgments (2014–2023), two trained coders extracted HFACS-relevant contributory factors, logged cross-level linkages for qualitative interpretation, and assigned one dominant HFACS level per case as the prediction target. Using structured case features describing vessel/encounter context and International Regulations for Preventing Collisions at Sea (COLREGs) indicators (including violation assessments), we trained random forest (RF), extreme gradient boosting, support vector machine, and naïve Bayes classifiers. Under stratified holdout evaluation, RF achieved 0.86 accuracy (weighted F1-score = 0.85), with the strongest signals from COLREGs-related violations, visibility, and give-way status. The framework enabled interpretable explanations for frontline categories (Unsafe Acts and Preconditions) but showed reduced separability for latent supervisory and organizational categories, consistent with heterogeneous reporting in KMST narratives. Although this dataset does not support stable inferential dependency/path analysis across HFACS levels, the documented linkages motivate future hierarchical or multi-label dependency modeling using larger, multi-source datasets. Overall, the approach strengthens COLREGs-aware, explainable decision support and aligns with the International Maritime Organization’s electronic navigation strategy for future autonomous bridge and vessel traffic service applications.