A Hybrid LSTM–XGBoost Framework for Vibration-Based Predictive Maintenance of Rotating Machinery

Abstract

The unexpected failure of an induced draft fan in a cement plant highlighted the need for more reliable predictive maintenance strategies. This study proposes a hybrid LSTM-XGBoost framework for one-hour-ahead vibration prediction. Trained on 18496 hourly sensor readings from that fan, the framework predicts vibration levels one hour ahead. The LSTM learns how vibration behaves over time; XGBoost corrects what the LSTM misses. Feature selection used Pearson correlation analysis, dropping bearing temperatures — noisy and redundant under cement mill conditions — and retaining speed, airflow, and winding temperature. The hybrid framework achieves an RMSE of 0.34, an MAE of 0.19, a MAPE of 1.33%, and an R 2 of 0.90, outperforming every standalone model tested. Vibration at time t often reflects thermal or mechanical conditions that began hours earlier. These results highlight the importance of modelling temporal dependencies when predicting vibration behaviour in industrial rotating equipment. Future directions include federated learning for cross-plant generalisation and explainable AI to make the model's decisions legible to the maintenance engineers who act on them.