Explainable Remaining-Useful-Life Prediction for Turbofan Engines with SHAP and an Audited Multi-Agent Literature Cross-Check

Models that predict the remaining useful life (RUL) of turbofan engines are now accurate on public benchmarks, but they rarely tell a maintenance engineer whether the sensors behind a prediction make physical sense. This paper builds and critically evaluates a pipeline that pairs a two-layer LSTM with SHAP sensor attributions and a four-stage language-model agent chain (Literature, Synthesis, Hypothesis and Critique) that compares each engine’s explanation with published research and labels it Confirmed, Contradicted or Novel. We evaluate on NASA C-MAPSS FD001, FD002 and FD004, and the FD002 and FD004 criteria were recorded in a dated file before any result existed. Against RUL capped at 125 cycles the model reached an RMSE of 13.58, 14.99 and 15.08. Against the raw RUL files the RMSE was 14.90, 28.38 and 27.28, because 22% of FD002 and 27% of FD004 test engines live longer than the cap. Normalising each sensor within its operating regime, and keeping three sensors that the FD001 rule would have dropped, changed what the model learned: in FD004 sensor 6, a bypass-duct pressure, was the top SHAP driver for 98 of 248 engines, consistent with that subset’s fan fault. The top driver stayed the same across five SHAP reruns for 237 of 259 FD002 engines but only 126 of 248 FD004 engines. The agents returned no Confirmed verdicts on FD002 (60 Contradicted, 199 Novel) and 55 Confirmed, 24 Contradicted and 169 Novel on FD004. Most Novel verdicts reflect gaps in the literature knowledge base, not new findings. The Critique Agent reviewed the FD004 engines without stability evidence and agreed with 4 of 248; on 122 FD002 engines re-reviewed with that evidence it agreed with 102, where its earlier review had agreed with none, and we state that this re-run is partial. An audit of the language-model benchmark table found fabricated and mislabelled rows, and we report which components were not evaluated on their own.