Predictive Quality Control and Defect Detection in Steel Manufacturing Using Statistical Learning and Process Analytics

Abstract

Steel plate manufacturing faces both operational hassles and financial impact caused by surface imperfections; yet, conventional human inspection usually suffers from fatigue, subjectivity, and modern rolling mills' throughput demands. This research aims to bridge three gaps in the field of automated quality control; these gaps highlight the scarcity of methodical, comprehensive cross-paradigm comparisons of algorithms. They also identify the limited operational availability of model interpretability frameworks and the lack of design of easily employable dashboarding tools for floor production personnel. From the UCI Steel Plates Faults Dataset, 1,941 instances of steel plates structured across seven defect labels were used. These instances were characterized by 27 process and geometric attributes. Within structured, reproducibly framed environments, classification algorithms of all biases and variance were both constructed and then evaluated. The five algorithms of this category were Linear Discriminant, Multinomial Logistic, Random Forest, Support Vector, and XGBoost. Class imbalances were corrected through real-sample up-sampling in the training portion, while a 30% held-out testing portion preserved the natural class distribution. The greatest observed testing set accuracy and weighted ROC-AUC were 79.79% and 0.9702, respectively, for XGBoost. Random Forest testing set upholding 78.58%, and AUC 0.9641, followed closely, with both classifiers noted for substantially surpassing the performance of linear classifiers. SHAP and DALEX permutation analyses discerned class-level feature drivers located within spatial parameters of Z_Scratch and Bumps, luminance attributes of K_Scratch and Stains, and the steel type for Dirtiness. The described analyses and attributes offer explainable, spatial process diagnostics of relevance. The research also focuses on the design of proprietary tools, where an interactive, seven-module R Shiny dashboard was both developed and utilized as a deployment method to fulfill real-time defect prediction, along with SHAP layered class suggestions for process personnel who lack subject specialization, thus satisfying the trinity of research ventures. Keywords: steel plate defect detection, predictive quality control, XGBoost, Random Forest, SHAP interpretability, Statistical Process Control, Industry 4.0, R Shiny dashboard