Leveraging generative adversarial networks for data augmentation to improve fault detection in wind turbines with imbalanced data

In the realm of modern power systems, wind turbines (WTs) have gained significant environmental advantages, making them pivotal in the expanding landscape of wind power generation. The paramount task of ensuring the effective monitoring and fault classification of WTs is indispensable for the stability of wind farm systems. Nevertheless, a formidable challenge arises from the need for more fault information gleaned from small fault data samples, resulting in data insufficiency and imbalance issues. The consequential impact of such imbalances on fault detection accuracy underscores the critical nature of the concern in wind turbine fault diagnostics. Notably, the disparity in dataset sizes among faults and the uneven distribution of fault classes pose a substantial obstacle in the realm of fault detection for wind turbines. This challenge is compounded by the time-intensive process required to amass sufficient fault data, setting it apart from more typical scenarios in the domain of wind turbines. This paper introduces an innovative approach employing Generative Adversarial Networks (GAN) for synthetic fault data generation to address these challenges effectively. This paper utilizes a Wasserstein Conditional Generative Adversarial Network (WC-GAN), which replaces the KL divergence in CGAN with the Wasserstein distance to rectify data imbalances by generating synthetic fault samples for wind turbine fault classification. This strategic manoeuvre balances the class distribution and enhances fault classification accuracy. Incorporating conditional data generation contributes to training stability and sample quality while utilising wasserstein distance ensures a faster convergence rate. Experimental validation conducted on Supervisory Control and Data Acquisition (SCADA) data for fault classification of wind turbines underscores the superiority of our method over other approaches, primarily attributed to the quality of conditionally generated samples. In essence, the proposed approach adeptly tackles the challenge of imbalanced samples by generating high-quality synthetic fault data, thereby elevating the efficacy of wind turbine fault classification.