Title: Trans GAN-WT: An Interactive Learning and Feature Extraction Framework for Anomaly Detection in Wind Turbine Time Series Data

Abstract

As the scale and quantity of wind farms continue to expand, the daily operation and maintenance expenses associated with wind turbines are rising. To mitigate these costs and improve the reliability of operational data prior to the onset of catastrophic failures, it is essential to monitor equipment status closely and identify faults at an early stage. Leveraging operational data to assess anomalies and enable real-time monitoring of wind turbine conditions holds significant practical value. However, current anomaly detection techniques struggle to effectively model relationships within data containing substantial redundant information and fail to adequately utilize valuable anomaly data.

To address these challenges, this study introduces a novel anomaly detection model that integrates a Transformer architecture with a generative adversarial network. The proposed approach first minimizes the leakage detection rate for minor deviation anomalies by amplifying reconstruction errors. It then employs autoregressive inference to extract multimodal features, thereby improving the stability and generalization capabilities of the training process. Additionally, a temporal feature extraction module is developed to facilitate interactive learning among features across different time scales, effectively reducing temporal redundancy.

Experiments conducted on real-world wind turbine generator (WTG) datasets demonstrate that TransGAN-WT achieves an average F1 score of 96.10% across various wind turbine datasets. This performance surpasses several state-of-the-art baseline methods by 5.84% and 2.89%. Furthermore, the model maintains a false positive rate (FPR) of just 0.06%. Statistical validation via the Wilcoxon signed-rank test confirms that the performance improvement over existing state-of-the-art baselines is significant, thereby ensuring the stable operation of wind turbines.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC