A Sequentially Optimized Stacked LSTM Framework for Residual-Based Bearing Fault Detection

Abstract

This study presents a healthy-only bearing fault detection framework in which a stacked long short-term memory (LSTM) predictor learns normal vibration dynamics through multi-step forecasting, and deviations between predicted and observed vibration sequences are used for residual-based anomaly detection. The methodological contribution lies in the controlled stage-wise development of the complete prediction-to-decision pipeline, including temporal configuration, predictor selection, residual scoring, and decision-threshold calibration, together with strict bearing-level separation between framework development and final evaluation. The framework was evaluated on the Paderborn bearing dataset using independent healthy bearings for training, validation, calibration, and testing, artificially damaged bearings for detector development, and previously unseen real-damage bearings for final fault evaluation. The finalized detector achieved an ROC–AUC of 0.8385 and a PR–AUC of 0.9200, with a healthy false-positive rate of 0.13% and precision of 0.9971. However, recall at the selected operating threshold was limited to 22.57%, showing that strong anomaly-score discrimination did not translate into equally high fault sensitivity under the conservative global threshold. Performance remained consistent across independent model initializations, while evaluation on a fully reserved healthy bearing produced a 17.95% false-positive rate, highlighting threshold calibration and healthy-domain generalization as the principal limitations of the current framework.