Time-Series Anomaly Detection and Imbalanced Classification for Predictive Maintenance

Predictive maintenance plays a vital role in reducing equipment failure rates and optimizing operational costs in the Industry 4.0 era. However, time-series sensor data face two major challenges: severe class imbalance between normal and fault samples, and complex nonlinear temporal dependencies that hinder traditional modeling. This paper proposes a hybrid deep learning framework integrating unsupervised anomaly detection via LSTM-Autoencoder and imbalanced classification using SMOTE-Tomek with XGBoost for rotating machinery fault prediction. Time-domain features are extracted through sliding-window segmentation, and LSTM-Autoencoder learns normal patterns for anomaly detection. The SMOTE-Tomek strategy balances the data and enhances fault classification performance. Experiments on real industrial datasets show that the anomaly detection module achieves an F1-score of 0.857 (8.4% higher than Isolation Forest), while the classification model attains macro-and weighted-average F1-scores of 0.865 and 0.966, respectively, with minority-class recall improved from 0.412 to 0.834 (102.4% increase). This framework provides a reliable, interpretable, and efficient solution for industrial predictive maintenance.