Work-order-adaptive and uncertainty-aware contextual pressure monitoring: an industrial case study in multi-specification aluminum extrusion

In multi-specification aluminum extrusion, extrusion pressure is measured for every billet, but its expected level varies with product specification, observed process conditions, equipment thermal state, and work-order context. Fixed pressure limits can therefore confound normal contextual variation with deviations requiring engineering review. This study developed a work-order-adaptive and uncertainty-aware contextual pressure-monitoring framework that reconstructs billet-level records from SCADA signals, learns specification-aware pressure baselines from historically normal data, calibrates unseen work orders using residuals from billets 2–6 without retraining the baseline model, constructs cross-work-order conformal reference intervals, and separates Tree SHAP explanations from the work-order correction. The industrial dataset comprised 1,741 valid billets, including 1,394 historically normal observations from 48 work orders across four specifications. Under leave-one-work-order-out validation, overall \:R² decreased from 0.745 under random billet-level cross-validation to 0.361, revealing substantial optimism when work-order structure was ignored. Work-order adaptation increased \:R² to 0.689, reduced MAE from 3.394 to 2.585 MPa (23.8%), and reduced RMSE from 4.635 to 3.234 MPa (30.2%); the paired Wilcoxon signed-rank test at the work-order level was significant ( $$\:p=0.026$$ ). Empirical coverage of the 95% and 99% reference intervals was 95.03% and 98.92%, respectively. The framework converts existing pressure measurements into context-aware, uncertainty-calibrated, and interpretable information for engineering review while maintaining a strict distinction between pressure deviation and product quality classification.