As manufacturing transitions toward Industry 5.0, the demand for trustworthy, operator-centric predictive maintenance has intensified; yet today’s CNC milling models often fall short on the very qualities that make them deployable: leakage-free evaluation, calibrated performance under imbalance, and explainability. This paper introduces XAI-PdMNet-Bench, a leakage-safe and explainable PdM framework validated on two CNC milling benchmarks: AI4I 2020 ( \(n{=}9{,}973\) ; 3.31 % failures) and PHM 2010 (real cutting signals; tool-independent split). A 25-feature physics-informed matrix is constructed after removing all leaked indicator columns, and minority-class augmentation is performed via statistically audited CTGAN (Frobenius \(= 0.083\) ; all KS tests passed). The proposed XGBoost achieves F1 \(= 0.991\) and PR-AUC \(= 0.994\) under leakage-safe cross-validation, exceeding the best leakage-unsafe baseline (F1 \(= 0.983\) ). TabTransformer achieves F1 \(= 0.987\) , the best deep-learning result. On PHM 2010, the proposed RCNN (Conv1D + BiLSTM) attains F1 \(= 0.921\) and PR-AUC \(= 0.941\) ; staged transfer learning from AI4I 2020 yields F1 \(= 0.894\) . SHAP attribution identifies RPM \(\times \) wear as the dominant fault driver, and a structured operator report supports operator-facing maintenance interpretation. Ablation studies confirm CTGAN contributes the largest gain under leakage-safe balanced cross-validation ( \(\Delta \) F1 \(= +0.160\) over no augmentation; \(\Delta \) F1 \(= +0.140\) on the imbalanced held-out set), followed by physics-informed feature engineering ( \(\Delta \) F1 \(= +0.050\) ).
