Reliability-Aware Diagnostic Status Formation for Marine Diesel Engine Fault Diagnosis Under Noisy and Incomplete Evidence

A forced fault label does not establish that diagnostic evidence supports an engineering conclusion. This study evaluates a post-prediction D0–D5 layer combining calibrated class probabilities, predicted severity, pressure–torsion feature-group consistency, and evidence completeness. The simulated 3500-DEFault benchmark at 2500 rpm and 0, 15, 30, and 60 dB was partitioned by scenario into training, calibration, validation, and test sets. Operating parameters were selected on validation data and locked before final-test evaluation. Across five split seeds, the calibrated classifier had error 0.0462 ± 0.0032. The status layer achieved coverage 0.9791 ± 0.0015 and four-class selective risk 0.0342 ± 0.0039. The status false-safe rate was 0.00208 ± 0.00165, compared with the forced false-safe rate of 0.0094 ± 0.0038. Among true G4 cases, 0.9795 ± 0.0024 received D3/D4. Approximately matched entropy rejection had slightly lower selective risk (0.0329 ± 0.0037); superiority in rejection efficiency is not claimed. Partial torsional-feature loss exposed substantial severe-case underwarning without universal refusal. The contribution is an interpretable separation of normal, intermediate, high-confidence fault, severe-fault, and refusal indications. Repeated splits quantify split sensitivity, not independent engine replication. Results are limited to the simulated fixed-speed benchmark and do not independently validate shipboard maintenance decisions.