A hybrid machine learning framework integrating persian NLP for occupational risk classification in steel manufacturing

Abstract

Human error remains a leading cause of occupational accidents in high-risk industries, yet traditional risk assessment methods often overlook valuable information embedded in unstructured textual data. This study develops and validates a hybrid machine learning framework combining structured risk data with Natural Language Processing (NLP) to improve hazard classification accuracy in industrial settings. Data were collected from 258 work activities across 16 operational units in a steel manufacturing facility over five months, yielding 3,110 validated risk scenarios. Hazard identification employed Job Safety Analysis (JSA), while quantitative risk evaluation used the William Fine method. The dataset comprised structured variables (severity, exposure, probability) and unstructured Persian-language text (hazard descriptions, locations, activities). Text preprocessing included tokenization, stopword removal, stemming, and lemmatization. Features were extracted using TF-IDF vectorization for all models. Random Over-Sampling (ROS) addressed class imbalance. Four traditional machine learning algorithms were evaluated: Logistic Regression (LR), Random Forest (RF), K-Nearest Neighbors (KNN), and Support Vector Machine (SVM). On unbalanced data, SVM achieved the highest accuracy (93%), with F1-scores of 0.96, 0.83, and 0.92 for low, medium, and high-risk categories, respectively. After ROS balancing, performance showed practical improvements across most models, particularly for minority risk classes; however, the difference did not reach conventional statistical significance (p = 0.189), likely due to the small comparison sample (n = 3 models). SVM reached near-perfect classification, whereas KNN remained the weakest performer. The hybrid framework significantly outperforms traditional risk assessment methods by leveraging both structured data and unstructured textual information. The prevalence of low-risk scenarios (66.1%) suggests effective existing safety controls, while improved classification of medium and high-risk cases enables targeted interventions. This approach provides a foundation for real-time AI-driven safety monitoring systems in high-risk industrial sectors.