Data Augmentation by Generative Adversarial Networks for Fault Detection in Power Production System

We consider the problem of fault detection (FD) in power production systems (PPSs) working under variable conditions and the issue that, in practice, data from operating conditions are not available. To address this issue, a Generative Adversarial Network (GAN) is developed for generating synthetic data of normal system operation, leading to the augmentation of the available dataset. The augmented dataset is, then, processed by an Autoencoder (AE) for the reconstruction of the signal values expected in normal system operation to enable FD. The proposed modelling framework is applied to a highfidelity model of a 4.8 MW wind turbine. Satisfactory FD performances are obtained under all considered wind regimes, including rare high-wind conditions. Specifically, the Area Under the Curve (AUC) increases from 0.84 to 0.86 when GAN generated data are added to the training set, and from 0.78 to 0.83 on high wind conditions.

Keywords: Power Production System, Wind Turbines, Fault Detection, Data Augmentation, GAN.