Machine-Learning Based Fault Detection for Wind Turbines
We develop intelligent and scalable deep learning algorithms for fault detection in critical components of wind turbines based on readily available SCADA data.
Nispera is a third-party energy forecasting and performance monitoring solution provider for renewable energy assets. In this project we develop a new software module for condition-based and predictive maintenance for the main components of wind turbines and integrate it in the existing platform of Nispera. For this purpose we develop state of the art deep learning algorithms for early fault detection and isolation in critical turbine components. Early detection enables wind farm owners to schedule maintenance in a planned manner before a complete stoppage of the turbine, thus avoiding long downtimes and the related high costs. The cost effectiveness of this software solution is due to the fact that the required data is recorded and stored by the Supervisory Control And Data Acquisition (SCADA) system which is already present on all wind farms and thus does not require any additional investment by the Owner. Our algorithms are based on multi-output Convolutional Neural Networks (CNNs) and allow for early fault detection of multiple critical components of the turbines with scalable training times. For scalable implementation in all operational wind farms we developed a fully automatic data pipeline, from data selection, through pre-processing, training, validation, up to alarm setting. Moreover, we extended our algorithms to enable fault detection also in cases of newly installed turbines (or farms) with little training data.