Physics-informed, cost-aware fault classification with comparative explainability for predictive maintenance: a SHAP–LIME agreement analysis

Predictive maintenance is a core capability of Industry 4.0, yet most machine learning studies on the widely used AI4I 2020 benchmark rely on raw sensor signals, treat all misclassification errors as equally costly, and adopt explainability tools without verifying their mutual consistency. This study develops and evaluates a multi-class, cost-aware fault-classification framework on AI4I 2020 that combines three components: (i) five physics-informed features (mechanical power, overstrain index, thermal delta, a thermal-speed interaction, and a wear-torque ratio) derived from the documented failure equations; (ii) an illustrative fault-severity cost matrix used both as a class-weighting scheme and as an operational evaluation metric; and (iii) a quantitative agreement analysis between SHAP and LIME feature rankings using Spearman rank correlation. Four classifiers (logistic regression, support vector machine, random forest, and XGBoost) are trained on a scaled, SMOTE-balanced training partition and evaluated with a leakage-safe cross-validation design in which scaling and oversampling are fitted inside each fold. Random forest is the strongest model on the learnable classes (test weighted F1 = 0.987, macro F1 = 0.629, ROC-AUC = 0.992) and the lowest-cost model (total cost = 148). An ablation study shows that the physics-informed features contribute independently: a random forest trained on the five physics features alone (weighted F1 = 0.980, cost = 190) outperforms one trained on the six raw features alone (weighted F1 = 0.976, cost = 290), and the combined set is best (cost = 148); this feature contribution is the central, best-supported result of the study. We emphasise two limits. First, robust six-class classification is not demonstrated: the model misses every tool-wear (TWF) test instance and the random-failure (RNF) class is absent from the test partition, so the method performs well only for the learnable classes (no-failure, HDF, PWF, OSF). Second, because the physics features are derived from the same rules that generate the data, the strong SHAP alignment with the physical triggers is partly expected and should not be read as evidence of transfer to noisy real systems. Consistent with the cost-aware framing, across ten random seeds cost-derived class weighting does not produce a statistically significant reduction in total cost relative to standard training (120.0±13.8 versus 124.6±17.0; Wilcoxon p = 0.078), because the cost is dominated by the data-scarce TWF class. A multi-instance SHAP and LIME analysis (up to 50 correctly classified instances per class) yields a pooled median Spearman correlation of 0.714 (interquartile range 0.618 to 0.830), indicating moderate rather than strong agreement that is sensitive to the LIME sample size and varies across instances and seeds. The findings support physics-informed feature engineering for interpretable predictive maintenance, caution against single-instance LIME explanations, and show that cost-sensitive weighting is ineffective when the dominant cost driver is an unlearnable minority class.