Constraint-aware imbalanced learning for multi-class defect prediction and process-window identification in sand casting

Class imbalance and competing process requirements complicate defect prediction and process optimization in sand casting. This study develops a framework combining constraint-aware imbalanced learning, model interpretation, and joint process-window identification for QT450-10 ductile iron castings. An industrial dataset comprising 6,385 records, 18 process variables, and five quality classes was analyzed. Borderline-PC SMOTE augmented minority classes exclusively within the training partition while restricting synthetic records to prescribed process bounds. The resulting backpropagation neural network achieved an overall accuracy of 93.58% and an all-class macro-F1 score of 0.8488. Compared with conventional BP, balanced accuracy increased from 83.31% to 87.73%, and macro-F1 across the four defect classes increased from 0.8008 to 0.8196. SHAP analysis identified distinct predictor patterns: pouring temperature and carbon were prominent for cold shut; inoculant dosage, new sand fraction, and aluminum for blowhole; new sand fraction and sand temperature for sand inclusion; and magnesium, pouring temperature, carbon, and silicon for shrinkage. Temperature responses differed between cold shut and shrinkage, supporting joint parameter evaluation. A separate probability-calibrated BP surrogate identified a candidate window across all 18 variables. Within this window, 19,852 of 20,000 Sobol-sampled combinations met both the defect-free classification criterion and a calibrated probability threshold of 0.90, yielding a model-based acceptance rate of 99.26%. The framework improves class-balanced defect recognition and translates model responses into candidate operating intervals for independent production validation.