Trustworthy Mining Digital Twins for Autonomous Operations through Calibrated Self-Awareness: Detecting Overconfidence and Model Misperception

Abstract

Trustworthy mining autonomy requires more than accurate state estimation: a digital twin should also indicate when the uncertainty attached to its state representation is unreliable. We formulate this requirement for Eco-Cognitive Mining Systems (ECMS) as a Calibrated Self-Awareness Index (C-SAI). A(t) measures tolerance-normalized state alignment, C(t) measures local central-interval calibration, CDI(t) identifies the direction of miscalibration, and OPS(t) reports predictive sharpness separately. C-SAI combines A(t) and C(t) and is reference-conditioned because both components require trusted reference outcomes. Accordingly, C-SAI is evaluated here as a construct-validation and benchmarking measure when such reference data are available, while NIS is retained as a reference-free runtime consistency diagnostic. Construct validation used 2,100 Monte Carlo runs spanning seven process, sensing, model, and combined challenge classes at three severity levels. Systematic unmodelled degradation produced strong severity-dependent reductions in minimum C-SAI (Spearman ρ = -0.904 to -0.939; all Holm-adjusted p < 0.001), whereas recognised measurement noise and short sensor dropout produced substantially smaller or non-monotonic losses. In a non-overlapping 40-run holdout, point-estimation trajectories were kept identical while predictive intervals alone were compressed. Median C(t) decreased from 0.699 to 0.244; the median paired decrease was 0.453 (95% bootstrap CI 0.444-0.463; Wilcoxon p<0.001; rank-biserial=1.00), while CDI shifted from -0.060 to -0.750. Thus, the calibration layer detected overconfidence that an alignment-only score cannot distinguish. The principal severity result remained stable across changes in the C-SAI mixing coefficient, calibration-window length, engineering tolerances, and state weights. The evidence remains simulation-based and does not establish superiority over NIS/NEES or full-distribution probabilistic diagnostics; comparative detection tests and independent plant validation are still required.