Digital-Twin-Conditioned Compact Ordinal Recognition of Aircraft Electrical Power-System Health States

A component health indicator does not directly establish whether an aircraft electrical power system (EPS) can sustain its required functions. This study evaluates a digital-twin-conditioned compact ordinal recognizer (DT-COR) for four ordered service states in a normalized EPS. Measured battery, insulated-gate bipolar-transistor, and bearing trajectories are synchronized through a declared network twin; system labels are produced by its priority-service and N − 1 rules. Testing holds out source assets across 32 episodes and 4800 operating points. The DT-COR emission head contains 108 learned coefficients (3 × 36); three intercepts, three validation offsets, one temperature term, and the temporal-mixing parameter are reported separately, while gate parameters, priors, unusable-zone values, and transition matrices are external configuration. DT-COR attains 90.8% four-state accuracy, 0.107 ordinal mean absolute error, and 0.024 expected calibration error on the reported asset-disjoint test set. Collapsing S2 and S3 yields 93.0% three-state accuracy from the reported confusion counts, while only 37 of 157 S2 points are identified as S2 in the four-state task and 103 are assigned to S3. The directional constraint applies only to directly constrained pre-filter emissions under fixed context; it does not guarantee monotonicity of the temporally filtered posterior. Under the reported 40% random masking protocol, DT-COR has an 11.8% false-safe proportion, but the baselines were not trained with identical masking augmentation, so this result is retained only as a descriptive unequal-training stress test. A separate state-contract sensitivity study over 27 prespecified threshold perturbations gives DT-COR severe-state AUROC in the range 0.988–0.990 and AUPRC in the range 0.811–0.825. The study therefore supports a compact research interface under twin-generated labels, not an aircraft-qualified recognizer.