Decision Support in Industrial Inspection: Systematic Comparison of Machine Learning Algorithms in Defect Classification in Steel Plates

This work investigates the automatic classification of defects in steel plates using supervised machine learning algorithms applied to the Steel Plates Faults (UCI) set. Seven models — Logistic Regression, Decision Trees, Random Forest, Gradient Boosting, SVM and KNN — were evaluated through stratified 10-fold cross-validation. The pre-processing included standardization of the data and balancing with SMOTE, applied within each fold to prevent information leakage. An ablation study looked at the impact of these steps. The results show that SVM and KNN are highly dependent on the scale of the data, while ensemble methods are more robust. Random Forest performed best in the test (F1 Macro = 0.8307), while Gradient Boosting stood out for its stability (F1 = 0.8158). Statistical tests indicated that both outperformed the other models (p < 0.01). The results reinforce the effectiveness of ensemble methods on unbalanced data.