Explainable deep learning for diagnosing the causes of casting defects based on representation of features as images with neighbourhood dependencies and local interpretable model-agnostic explanations methods

Systematic monitoring of defects in the castings produced by high-pressure die casting (HPDC) is essential to ensure the production of high-quality components, particularly given their growing prevalence in safety-critical automotive applications.However, conventional analysis techniques may not fully capture the complex dynamics of the process, resulting in limited effectiveness and an impossibility of continuous improvement.Therefore, this study proposed an innovative methodology that autonomously performs pre-processing by transforming tabular data into images using the REFINED (representation of features as images with neighbourhood dependencies) approach.This transformation enables convolutional neural networks (CNNs) to recognise hidden patterns and dependencies in the data.Additionally, explainable artificial intelligence (XAI) principles were employed to elucidate the influence of process parameters on defect formation, by the local interpretable model-agnostic explanations (LIME) method application.The proposed model achieved high predictive accuracy of 99% and a low RMSE of 0.07, effectively capturing nonlinear interdependencies also revealed 'maximum pressure' as the main significant factor influencing defect formation.By combining proposed methods, the approach enhances process transparency and enables autonomous, data-driven decision-making in HPDC quality control, advancing the vision of intelligent manufacturing and Quality 4.0 within the Industry 4.0 paradigm.