Fault Detection in Wind Turbines using Deep Learning

Abstract

Institutions have been redirecting investments away from fossil fuels, creating a path for clean energy generation. The wind industry has seen an exponential increase in recent years. Early fault detection creates an alternative for operation and maintenance (OM), allowing costs to be avoided before they reach a catastrophic stage, and improving turbine reliability. Predictive maintenance was the solution that presented itself for this problem, in which faults are detected before they occur and fixed accordingly. LSTM-Autoencoder and time-series data collected from SCADA sensors installed in wind turbines are used to detect anomalies in several components of the wind turbines that insinuate a major fault might occur. The dataset is collected from a wind farm in the West African Gulf of Guinea in 2016. Results have shown how PCA can be productive in identifying the features with the most influence on the prediction process, with the ability to predict faults 17.5 days prior on average.