Fidelity-Utility Dissociation in Synthetic Data Augmentation for RUL Prediction of Window-based Tabular Transformation

Purpose: Remaining useful life (RUL) prediction is central to the prognostics and health management (PHM) of aero-engines, but run-to-failure data are scarce, leading to a small-sample problem. This study reconstructs the variable-length multivariate time series of the NASA N-CMAPSS dataset into a fixed tabular form and evaluates, along two separate axes, whether synthetic data augmentation improves RUL prediction.<br/>Methods: The HDF5-based time series is aggregated at the unit-cycle level and transformed into a 192-dimensional tabular representation through sliding windows (size 30, stride 5) and six statistics per variable. CTGAN, TVAE, and GaussianCopula are then applied at 0.5x/1.0x/2.0x ratios, and the effect is assessed by distributional Fidelity (KS statistic, correlation Frobenius norm) and predictive Utility (test RMSE) using RandomForest, GBM, and Ridge.<br/>Results: TVAE achieved the most stable distributional fidelity (Frobenius norm 1.67), whereas the lowest test RMSE was obtained by GaussianCopula augmentation at a 1.0x ratio with RandomForest (RMSE 9.15, a 15.1% improvement over the baseline of 10.78). CTGAN showed the weakest distributional fidelity and unstable predictive utility across augmentation ratios.<br/>Conclusion: Distributional fidelity of synthetic data does not guarantee predictive utility; therefore task-oriented utility evaluation must accompany fidelity metrics in PHM data augmentation.