<![CDATA[Operational costs in coal getting (CG) and overburden removal (OBR) represent the largest expenditure in coal mining, yet systematic anomaly detection remains limited. This study develops an integrated machine learning pipeline using 1,508 daily observations from a coal mining company in Indonesia (July-December 2025), integrating five operational data sources. Anomaly detection combines Isolation Forest, Local Outlier Factor, ECOD, COPOD, HBOS, and an Autoencoder under a non-circular evaluation design in which detection features and the weak ground-truth label are strictly independent. All six baselines show weak discriminative power against this weak label (AUC 0.47-0.59), with the Autoencoder narrowly the strongest (AUC=0.590). A continuous Ensemble Score, thresholded at the 95th percentile, isolates 76 candidate anomalies (5.0%) for typology classification and full-population expert validation. Validating all 76 candidates plus a 30-observation Normal control sample (106 total, blind labeling) yields Precision=0.82, Recall=1.00, F1=0.90, and Cohen's Kappa=0.71 (substantial agreement); critically, AUC against this expert label reaches 0.91, confirming that the Ensemble Score's ranking ability is genuine even though its AUC against the weak proxy label is not. Root-cause analysis identifies breakdown hours, downtime ratio, and lost time - not unit cost - as the dominant discriminators of anomalous days. Hauling costs dominate at 65.3% of total operational expenditure. A material-split regression model reduces prediction error (MAPE) from 23.80% to a weighted 16.77%, with SHAP analysis identifying operational efficiency as the primary cost driver. Seven evidence-based managerial recommendations are proposed for predictive-preventive cost management.]]>
Copyrights © 2026
