A Comparative Study of Intelligent Methods for Steel Surface Defect Classification in Industrial Quality Control

Abstract

The intelligent transformation of manufacturing processes increasingly relies on automated visual and parametric inspection of intermediate products, where machine learning models classify surface defects from sensor-derived features. Despite the rapid progress of deep tabular learning architectures, the relative performance of classical ensemble methods, gradient boosting, and modern neural-attention models on small industrial datasets remains insufficiently characterized. The absence of unified comparison protocols hinders informed deployment decisions in industrial quality-control environments. This paper presents a comparative study of five representative intelligent methods — logistic regression, Random Forest, CatBoost, a multilayer perceptron, and TabNet — applied to the classification of steel-plate surface defects across seven fault categories. A unified evaluation protocol is proposed, integrating correlation-based feature selection, synthetic minority oversampling within cross-validation folds, exhaustive grid-search hyperparameter tuning under stratified cross-validation, and held-out test-set evaluation. The protocol prevents data leakage at every stage and treats class imbalance through SMOTENC applied inside the resampling loop. The experimental results show that ensemble methods — Random Forest and CatBoost — achieve closely matched leading performance (test F1-macro 0.81) The multilayer perceptron and TabNet trail by 4-8 F1-macro points. Per-class analysis identifies systematic confusion among visually similar defect categories. Feature-importance analysis attributes the predictive signal primarily to defect-size descriptors and production-line geometry. The proposed protocol provides a reproducible framework for selecting intelligent methods in resource-constrained industrial quality-control deployments.