Physics-informed anomaly and fault detection for wind energy systems using deep CNN and adaptive elite PSO-XGBoost

Abstract

Wind energy systems require fault diagnosis that identifies faults despite data inconsistencies. This study addresses challenges in supervisory control and data acquisition (SCADA) systems for monitoring wind turbine conditions from imbalanced data representation and error vulnerability. It examines the efficacy of adaptive elite-particle swarm optimization (AEPSO)-tuned extreme gradient boosting (XGBoost) on an imbalanced SCADA dataset for wind turbine fault classification. The methodology integrates the resampled dataset with t-distributed stochastic neighbour embedding represented deep learning features. Employing AEPSO-XGBoost classifier trained on merged SCADA and deep learning data from a physics-informed deep convolutional neural network forms the basis of the fault (alarm) classification model. The AEPSO-XGBoost regressor is validated across three distinct rear bearing temperature datasets, facilitating parameter optimization and model robustness. Also, this study explores supervised and unsupervised anomaly detection models using PDCNN and AEPSO-XGBoost with rear-bearing temperature data. Findings exhibit substantial fault classification and prediction enhancements by merging resampled SCADA data with deep learning features. Moreover, results show that applying AEPSO-XGBoost can significantly improve anomaly detection metrics. Through AEPSO-XGBoost's efficacy in enhancing fault prediction within imbalanced SCADA datasets, the study proposes an integrated framework for fault classification and anomaly detection as an innovative predictive maintenance system for wind energy systems.