Skip to content
DSML.KZ
Publications

Talk · IEEE ... International Conference on Data Mining workshops · November 12, 2025

LLM-Enhanced Semantic Clustering for Predictive Industrial Incident Detection: A Zero-Shot Approach to Safety Analytics

Magzhan Kairanbay, Sayed Ahmed Khalaf, Ali Alshehab, Basaveswar Reddy Mora Reddy, Ahmed Bahar

Magzhan Kairanbay
Open via DOI

Abstract

Timely incident detection and prevention in industrial settings, such as aluminum smelters, is critical for safety, minimizing downtime, and reducing operational costs. While prior state-of-the-art methods, notably Kurian et al.'s framework achieving ~90% classification accuracy, focused on reactive post-incident analysis using TF-IDF and rule-based methods requiring extensive manual labeling, this paper presents a predictive analytics system leveraging transformer-based embeddings, large language models (LLMs), and zero-shot learning. Analyzing five years of textual incident data covering 61,593 incidents, our approach eliminates manual annotation through LLM-based attribute extraction, shifts from reactive classification to proactive prediction, and provides deep semantic understanding via BERT embeddings and Claude Sonnet integration. The system combines unsupervised clustering, temporal pattern analysis, and automated insight generation to forecast incidents, prioritizing recall to minimize false negatives. Evaluation achieves an F1 score of 76% and recall of 84%, representing a paradigm shift from post-incident analysis to proactive prevention and demonstrating the transformative potential of modern NLP techniques in industrial safety management while reducing deployment effort across operational contexts.

Topic ModelingText and Document Classification TechnologiesOccupational Health and Safety Researchincident detectionindustrial safetyaluminium smelter

More by Magzhan Kairanbay