INTERPRETABLE REMAINING USEFUL LIFE PREDICTION OF ROLLING BEARINGS FROM SINGLE-CHANNEL VIBRATION USING RANDOM FOREST, XGBOOST AND LSTM

Abstract

Interpretable data-driven remaining useful life (RUL) prediction is crucial for the practical prognostics of rolling bearings in industrial environments. This paper presents a two-stage framework in which Random Forest classifier first identifies anomalous operating regimes, followed by XGBoost and a Long Short-Term Memory (LSTM) regressor for RUL estimation on degradation segments. Features are extracted from fixed-length windows of single-channel vibration data and include a limited set of time-domain statistics and frequency-domain descriptors based on dominant spectral components. The proposed approach is validated using the NASA IMS bearing run-to-failure dataset. The Random Forest classifier achieves 98.6% accuracy on a held-out test set, with high precision and recall for the fault class. A hybrid RUL estimator combining XGBoost and LSTM predictions reduces MAE and RMSE by approximately 15–20% compared to individual models. Model interpretability is enhanced using SHAP, identifying RMS, crest factor, and dominant spectral features as key degradation indicators.