Defect Detection Refinement: The Synergy of CNN-RF for Casting Quality Enhancement

This research study offers a full evaluation of the categorization model’s efficacy when recognizing six key casting flaws’ events observed across different industrial production processes. The model performance is tested based on the accuracy, recall, and F1-Score scores considering distinct classes of defects, including Swell casting, Metal penetration, Shrinkage cavity, Cold Shut, Slag inclusion, and Pinholes. The obtained results reveal high values of accuracy calculating from 94.05% to 96.05%, which suggests that the model has a consistent ability to categorize events of every fault. The recall scores reveal values ranging from 94.61% to 96.09% implying the model’s ability to appropriately detect real faults. The F1-Score measure, defining a mean between the two considerations, presents values from 94.33% to 95.82%, which supports the assumption of the model’s competency in the fault categorization. Additionally, the value of support provides insights into the dataset distribution, considering the proportion of each calculation row and emphasizing the fault’s comparative significance. The overall model’s accuracy is presented at 0.98 to 0.99, suggesting a high level of suitability for defect categorization. Furthermore, macro, weighted, and micro-average values help summarize the knowledge concerning each fault class. In such a way, it is possible to receive an aggregate evaluation and a full model’s assessment. These are the results made possible by machine learning methods, ensuring the continuous growth of defect detection systems in industrial production. Diffident types of faults may thus become easier to identify, contributing to increased product quality and overall efficiency.