Cyber-attack resilient data-driven approaches for early fault prediction system for wind turbines

Predictive maintenance (PdM) technologies, facilitated by smart sensors and artificial intelligence, are increasingly adopted in smart grids to ensure reliable operation and prevent financial losses and physical damage. However, these data-driven systems remain susceptible to cyber-attacks. This study examines the resilience of machine learning and deep learning models in predicting wind turbine (WT) faults 30 min in advance under False Data Injection Attacks (FDIAs). Random Forest (RF) and eXtreme Gradient Boosting (XGBoost) outperformed Decision Tree Classifier (DTC) and Gradient Boosting Machine (GBM) as fault predictors under normal conditions. Consequently, a comprehensive evaluation was performed using RF, XGBoost, and Long Short-Term Memory (LSTM) models under cyber-attacks, revealing that such attacks can reduce prediction accuracy by up to 15% and recall by 28%. Both autoencoder-enabled LSTM and adversarial training were implemented as defense mechanisms. While the autoencoder improved stability, adversarial training achieved superior robustness, making XGBoost the most resilient model, maintaining reliable fault prediction under cyber threats with almost no loss. Unlike existing studies focusing solely on fault prediction with clean SCADA data, this work integrates cyber-attack resilience assessment, adversarial defense, and structured data processing into a unified framework, bridging the gap between reliability, security, and data quality for trustworthy WT PdM.