From Tables to Computer Vision: Transforming HPDC Process Data into Images for CNN-Based Deep Learning

This paper proposes a methodology for leveraging convolutional neural networks (CNNs) in conjunction with advanced data preprocessing to facilitate optimal quality control decision-making in high pressure casting (HPDC) processes. The approach assists in predicting key values of the dependent variable associated with defect occurrence, enabling foundries to enhance product quality, reduce waste, and augment overall production process efficiency. The proposed study is founded on two principal pillars: the transformation of process tabular data (generated using the Conditional Tabular Generative Adversarial Network (CTGAN)), involving the mapping of features onto a fixed grid in a heatmap structure, and the configuration of the CNN algorithm to extract complex patterns in the data that are not readily apparent in the original tabular format. The study utilized a substantial dataset with a total of 61,584 images, and the most effective model attained an impressive Root Mean Square Error (RMSE) of 0.81, underscoring the model's remarkable capacity to accurately detect and predict casting quality issues. The model's efficacy was evaluated through its application to both large and small, differently distributed data sets. Utilizing a combination of statistical pre-processing, intelligent generative models, visual data transformations and deep learning, the methodology offers a comprehensive approach to enhancing production efficiency, ensuring superior process control and improving the quality of HPDC products. This development signifies a significant advancement in the field of intelligent systems for manufacturing process optimization, aligning with the principles of Industry 4.0 and Quality 4.0.