A GAN Based Fault Detection of Wind Turbines Gearbox

Abstract

Wind turbine gearbox condition monitoring is an essential part of the modern wind energy industry. Condition monitoring system can detect gearbox failures in time to avoid further damage. Data-driven condition monitoring methods have been widely used to ensure operation safety and reduce the downtime of wind turbines. Fault detection of wind turbine gearbox is usually based on vibration signals. However, faulty samples are not easy to collect in practice. Hence, it can be regarded as an imbalance data classification problem. In this paper, a generative adversarial networks (GAN) based method is proposed for the detection of gearbox faults. In the proposed method, the vibration signals are first pre-processed using short-term Fourier transformations (STFT). Then a GAN model is trained with healthy samples. A low score is obtained while inputting healthy samples into the neural networks of trained GAN. If the samples are collected in faulty scenarios, the trained GAN model will give a high score. Therefore, the faulty samples can be discriminated from healthy samples. The effectiveness of the proposed method is demonstrated with wind turbine gearbox condition monitoring vibration datasets by comparisons with other methods.