Wind Turbine Blade Bolt Anomaly Detection Based on WInfoGAN-IJSGBDT

As the proportion of wind power generation continues to increase, blade bolt failures in large wind turbines have become increasingly prominent. To address the issues of class imbalance in operational data from wind turbine blade bolts and the poor detection accuracy of existing models, this study proposes an anomaly detection method based on a fusion model combining a mutual information generative adversarial network and a gradient boosting decision tree, termed the Wasserstein Information Maximizing Generative Adversarial Network – Improved Jellyfish Search optimized Gradient Boosting Decision Tree. The proposed model employs a targeted data augmentation approach using the WInfoGAN framework to generate synthetic data samples with desired anomaly characteristics, thereby balancing the number of normal and abnormal samples. Additionally, the model incorporates an IJSGBDT module optimized by an enhanced jellyfish search algorithm to achieve anomaly detection for wind turbine bolts. This module utilizes adaptive step sizes and an elite retention mechanism to holistically optimize key hyperparameters, thereby enhancing parameter optimization capability and convergence speed. Analysis based on SCADA data from doubly-fed induction generators in an actual wind farm demonstrates that the WInfoGAN-IJSGBDT anomaly detection model achieves a false negative rate of 0.364% and a false positive rate of 0.519%.