Hazard classification and risk prioritization in heavy manufacturing industries using natural language processing and MCDM method

Abstract Heavy manufacturing industries are associated with different types of interacting hazards. Conventional hazard identification and prioritization approaches are largely manual, subjective, and difficult to scale. This study proposes an integrated decision support framework that combines Natural Language Processing(NLP) enabled hazard classification with Full Consistency Method (FUCOM)—Risk Priority Number(RPN) based MCDM approach for systematic workplace risk prioritization. Through an extensive literature review, a total of 25 hazards were identified and categorized into five categories: physical, chemical, ergonomic, psychosocial, and environment/workplace hazards. Eight NLP-based classification models were evaluated, including lexical similarity methods, Bidirectional Encoder Representations from Transformers (BERT) based embeddings, and Sentence Transformer architectures. Findings indicate that the use of Sentence-Transformer models paired with Cosine, Euclidean, and Minkowski distance metrics achieved 100% precision, recall, and F1-score, whereas BERT with cosine similarity and Natural Language Toolkit (NLTK) with Jaccard similarity yielded only 28% accuracy, with 0% recall for psychosocial and environment/workplace hazards. For risk prioritization, the FUCOM was applied on the judgement of 20 industry expert, leading to highest weight for chemical hazards (0.284), followed by environment/workplace (0.243) and physical hazards (0.242). RPN analysis identified stress (0.087), toxic gas release (0.065), and fume exposure (0.053) as the most critical individual risks. The integrated FUCOM-RPN ranking further confirmed toxic gas release, fume exposure, and electrocution as the highest-priority hazards. Sensitivity analysis with 1–15% weight variation demonstrated stable rankings for both top- and bottom-ranked hazards, confirming model robustness. The proposed framework offers an effective data-driven method of occupational hazard management, which is quantitative, scalable, and reliable in heavy manufacturing industries.