Auditing Operating-Condition Normalization in C-MAPSS Remaining Useful Life Evaluation

Abstract

NASA C-MAPSS is a canonical benchmark for turbofan remaining useful life (RUL) prediction, yet conclusions can depend on preprocessing and evaluation protocol in addition to model architecture. We present a protocol and reproducibility audit using compact MLP, temporal CNN, and LSTM baselines as controlled probes. A leakage-safe factorial grid varies the C-MAPSS subset, RUL cap, window length, normalization policy, architecture, and random seed, while enforcing and recording a fixed train/validation/final preprocessing fit policy. All 1080 planned canonical-grid runs completed, with no engine-split overlap or non-finite metrics. On single-regime FD001/FD003, global and per-condition normalization are intentionally equivalent in the harness and are used only as an implementation unit test; on multi-regime FD002/FD004, per-condition normalization lowers RMSE in every matched cell, with mean signed RMSE deltas of -4.39 and -7.40, respectively. A focused 720-run FD002/FD004 ablation further shows that jointly refitting the final model and normalizer on train+validation engines changes reported RMSE by -1.119 to -0.206 on average relative to train-only final fitting, depending on subset and normalization. Across the full grid, subset identity explains 60.8% of cell-mean RMSE variation; within multi-regime subsets, RUL cap and normalization are descriptively large among the evaluated protocol factors, with a seed-balanced sensitivity giving the same qualitative conclusion. These results extend prior PHM studies of operating-condition-specific standardization by quantifying matched protocol effects under a frozen audit contract and by making split unit, fit scope, regime assignment, score convention, and reproducibility records explicit.