A Unified Framework for Casting Defect Classification: CNN Meets Random Forest

This study offers a new method for casting defect prediction by incorporating a hybrid model made up of Random Forest classifiers and Convolutional Neural Networks (CNNs). Inclusions, Blowholes, Shrinkage Defects, Inclusions, Misruns, Hot Tears, Misalignment, Metal Penetration, Sand Inclusion, and Warpage are among the ten different casting defect types that the model is intended to categorize. With precision, recall, and F1-score metrics for every class, the model's performance is carefully assessed. Its robustness in identifying and classifying casting flaws is demonstrated by its continuously high results, which approach 94%. Support values or proportions validate a forecast distribution that is balanced among classes. The model shows a strong 86.20% accuracy in correctly recognizing casting flaws in a set of 12,981 pictures, providing a strong basis for improving defect detection techniques in practical applications. A stunning 95.93% is reached by the macro-average F1-score, suggesting balanced prediction skill across all classes. When class imbalances are considered, the weighted average F1 score stands out at 95.91%, further supporting the model's adaptability. Micro-average precision, recall, & F1-score combined findings show that the model's overall accuracy is 95.91%. These results confirm the remarkable 99% accuracy rate and reliability of the proposed hybrid model. The research's powerful casting fault prediction tool offers notable gains in quality control inside production processes. The hybrid CNN-Random Forest method's efficacy in upholding and enhancing product quality standards within the domain of casting defect prediction is validated by its noteworthy precision, recall, F1 scores, and balanced support proportions.