A high‐accuracy and robust diagnostic tool for gearbox faults in wind turbines

Abstract

Faulty gears are a common cause of wind turbine failures. For this sake, this work was developed as a reliable diagnostic tool for wind turbines to improve wind power stability accordingly. A convolutional extension neural network (CENN) was proposed to identify vibration and audio signals captured from a gearbox. According to the status of the contained faulty gears, a gearbox was categorised as one of the three types: (i) broken, (ii) rusty and (iii) a combination of (i) and (ii). It was further assigned one of the three severity levels: mild, moderate and severe. Therefore, there were a total of nine combinations for identification. Captured raw vibration and audio signals were applied to a chaotic synchronisation detector by which 3D chaotic error scatter feature images were generated to train and test the CENN. The recognition rate provided by CENN and the majority rule reached 99.6%, and then slightly fell to 97.4% in a noise robustness test, and consequently CENN outperformed counterparts in terms of the recognition rate and the robustness against noise. Accordingly, multiple gearbox faults can be well diagnosed for the first time in the literature. Finally, this paper concludes with a simplified version of the original proposal.