Industry 4.0 plants collect rich IoT data, but many predictive models remain black boxes and give only point estimates. This limits trust and can lead to risky operating decisions. We propose a unified framework that makes high-accuracy gradient boosting both explainable and uncertainty aware, and then converts its outputs into actionable shop-floor settings. We train paired XGBoost models for regression KPIs and a maintenance-risk classifier using group-aware cross validation to avoid unit leakage. We explain each prediction with SHAP values to reveal the key controllable levers. We add uncertainty quantification for both task types. We use conformal prediction intervals for continuous KPIs and Wilson score bounds for calibrated class probabilities. We then run a multi-objective decision optimizer that minimizes energy and completion time while constraining worst-case risk using upper confidence bounds, not mean probabilities. On a distributed manufacturing dataset (3,741 samples, 10 units), the framework achieves stable generalization with about 90% interval coverage across major KPIs. The optimizer reduces completion time and energy with only a negligible downtime trade-off, and the explanations show that control and policy variables dominate over passive sensor flags.

Keywords

Explainable AI Manufacturing, XGBoost Uncertainty Quantification, SHAP Predictive Maintenance, Industry 4.0 Optimization, Risk-Aware Decision Framework