Wind turbine gearbox failure and remaining useful life prediction using machine learning techniques

TLDR

The study leverages one of the world's largest wind turbine operational databases, providing thousands of gearbox failure cases and comprehensive SCADA and vibration data, enabling unprecedented analysis. The study aims to predict gearbox failure and remaining useful life in modern multi‑megawatt wind turbines. The authors employ machine learning models trained on extensive labelled SCADA and vibration data, analyzing data trends and input weighting to predict failure and RUL. Artificial neural networks provide the most accurate predictions, correctly identifying failures 72.5–75% of the time with SCADA data and 100% with vibration data, and enabling failure forecasts up to one month ahead using SCADA and 5–6 months ahead using vibration; multi‑class networks further improve performance when trained on vibration data.

Abstract

Abstract This research investigates the prediction of failure and remaining useful life (RUL) of gearboxes for modern multi‐megawatt wind turbines. Failure and RUL are predicted through the use of machine learning techniques and large amounts of labelled wind turbine supervisory control and data acquisition (SCADA) and vibration data. The novelty of this work stems from unprecedented access to one of the world's largest wind turbine operational and reliability databases, containing thousands of turbine gearbox failure examples and complete SCADA and vibration data in the build up to those failures. Through access to that data, this paper is unique in having enough failure examples and data to draw the conclusions detailed in the remainder of this abstract. This paper shows that artificial neural networks provide the most accurate failure and RUL prediction out of three machine learning techniques trialled. This work also demonstrates that SCADA data can be used to predict failure up to a month before it occurs, and high frequency vibration data can be used to extend that accurate prediction capability to 5 to 6 months before failure. This paper demonstrates that two class neural networks can correctly predict gearbox failures between 72.5% and 75% of the time depending on the failure mode when trained with SCADA data and 100% of the time when trained with vibration data. Data trends in the build up to failure and weighting of the SCADA data inputs are also provided. Lastly, this work shows how multi‐class neural networks demonstrate more potential in predicting gearbox failure when trained with vibration data as opposed to training with SCADA data.