Remaining useful life (RUL) prediction is more useful for maintenance decision-making when the basis of a warning can also be inspected. This study develops an explainable RUL prediction framework for rolling bearings using the open-source NASA IMS 1st_test run-to-failure dataset. Each one-second vibration record is divided into 1024-sample windows, from which RMS, Peak-to-Peak, Kurtosis, and Crest Factor are extracted and aggregated as a four-feature observation. Ten consecutive observations form the input to a lightweight gated recurrent unit (GRU) regressor. To avoid imposing an arbitrary degradation-onset boundary on the prediction target, normalized RUL is defined directly from elapsed life and decreases continuously from 1 to 0. Using the first 70% of the run for training and the final 30% for chronological testing, five random initializations give an MAE of 0.329 ± 0.084 and an RMSE of 0.351 ± 0.098 for the GRU model, compared with MAE/RMSE values of 0.348 ± 0.128/0.393 ± 0.149 for the 1D-CNN and 0.363 ± 0.011/0.386 ± 0.013 for the LSTM. Gradient sensitivity, Integrated Gradients, SHAP, and permutation analysis consistently identify Kurtosis as the most influential late-life feature, while a feature-deletion test produces the largest average MAE increase when Kurtosis is removed. The study is intentionally limited to one fixed-speed, fixed-load IMS run and should therefore be interpreted as controlled benchmark validation rather than vehicle-level validation.
