Explainable Artificial Intelligence (XAI) has emerged as a critical enabler for trustworthy machine learning (ML) deployment in industrial settings. In Additive Manufacturing (AM), while predictive models achieve high accuracy for defect detection, their "black-box" nature impedes engineering adoption, as operators cannot translate opaque predictions into corrective process adjustments. This study proposes an Integrated Explainable Defect Detection Framework (IEDDF) that bridges this interpretability gap through multi-level Shapley additive explanations (SHAP) analysis combined with ensemble learning. The framework encompasses four phases: data preprocessing with class imbalance handling, comparative benchmarking of ten ML models, dual-scale explainability analysis (global feature importance and local prediction decomposition), and threshold-based risk boundary identification. Applied to a publicly available manufacturing dataset (3,240 observations, 17 variables), the framework found that ensemble methods achieved top-tier accuracy (0.946), while SHAP analysis revealed a concentrated importance structure dominated by three operational features accounting for 54.5% of total contribution. Cross-method comparison between SHAP and Local Interpretable Model-Agnostic Explanations (LIME) yielded only 20% agreement in local feature ranking, providing empirical evidence that single-method XAI approaches produce incomplete explanations in manufacturing contexts. The proposed framework offers a replicable, model-agnostic methodology for transitioning from black-box prediction toward interpretable, evidence-based quality management in AM environments.
Science Journal 