Symmetry-Aware Neural Evidential Reasoning for Aero-Engine Health-State Assessment Under Operating-Condition and Fault-Mode Asymmetries

Symmetry and asymmetry appear together in aero-engine prognostics and health management. A monitoring system should preserve a common decision structure across operating scenarios, while sensor evidence becomes asymmetric under changing operating conditions and fault modes. This study formulates the National Aeronautics and Space Administration (NASA) Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) turbofan benchmark as a four-level health-state assessment problem using remaining useful life (RUL) thresholds of 100, 50 and 15 cycles, and proposes a neural evidential reasoning (ER) framework with decision-structure symmetry. FD001–FD004 share the same health-state space, threshold map, reliability-discounted ER operator and final argmax rule, whereas feature selection, evidence transformation and reliability parameters are estimated independently for each subset. Sliding statistical descriptors and training-fold-only minimum-redundancy maximum-relevance (mRMR) selection retain six compact sensor-derived features. FD001 and FD003 therefore use six inputs; FD002 and FD004 additionally retain the same three operating-setting variables and use nine inputs. Boundary soft labels and auxiliary ordinal/RUL supervision exploit ordered degradation information near adjacent-state boundaries. Evaluation uses engine-disjoint validation, three random seeds, identical subset-specific inputs for every comparator, ensemble/neural/ordinal baselines, component ablation and calibration/error-detection analyses. Across the four subsets, the proposed model achieves the highest mean three-metric average (Avg3), 0.664 (standard deviation 0.007), with Accuracy 0.816 (0.006), Macro-F1 0.582 (0.010) and Balanced Accuracy 0.594 (0.005). It is statistically comparable to histogram gradient boosting, random forest and a plain multilayer perceptron, and significantly exceeds the tested compact-input cumulative ordinal, Feature Transformer and correlation-graph attention controls after Holm correction. Fixed-probe diagnostics identify the five-cycle window as the best overall short-history setting. The mRMR top-6 interface reduces the sensor-feature dimension by 97.1% while retaining 98.2% and 97.0% of full-pool Avg3 on FD002 and FD004, respectively. Unknown mass ranks erroneous predictions above correct predictions on every subset (area under the receiver operating characteristic curve (ROC-AUC) 0.749–0.810), including the highest value on the most heterogeneous FD004 subset. These results support a compact and auditable evidence interface that combines competitive classification with source-wise reliability and uncertainty information.