Data-Driven Quality Prediction and Process Optimization for Large Complex Castings

The quality of large complex castings is affected by nonlinear and partially interacting production variables, making shop-floor parameter adjustment difficult when conventional experience-based rules are used alone. This study develops a data-driven framework for qualification-rate prediction, production-record-label-based scrap-cause screening, process-parameter recommendation, and engineering validation. The two production datasets contained 354 records; after numerical completeness checking of the 49 input variables and the qualification-rate target, 350 valid records were retained. Four candidate regression models were compared under five-fold cross-validation, and a 150-tree random forest was selected as the qualification-rate predictor because it achieved the lowest mean RMSE of 3.6865 percentage points. A 49–128–64–13 deep feedforward neural network was retained as an exploratory multi-label screening model for routine production scrap-cause labels, achieving a five-run Micro-F1 of 0.2060 under sparse positive labels. Six adjustable variables were selected by jointly considering model relevance, shop-floor adjustability, and engineering feasibility. The random-forest-based counterfactual evaluation predicted a context-averaged qualification-rate increase of 0.8359 percentage points, while the auxiliary DNN label scores used in the feasibility screening did not increase. In the field implementation, all 5 products in the pilot stage were qualified, followed by 197 qualified products out of 200 across four production batches, corresponding to an overall qualification rate of 98.5%. The field results showed repeated-batch consistency under actual production conditions.