Wind turbine pitch systems play an important role in wind energy extraction. Such systems are known to suffer high failure rates resulting in long downtime maintenance. Consequently, industries wish to utilize failure prognostic techniques to schedule effective maintenance in order to reduce losses. A cost-effective method that analyses existing data collected from the built-in sensors in the wind turbine (WT), such as the supervisory control and data acquisition (SCADA) data is presented in this paper. Information such as wind speed, power generation and subsystem measurements are collected for every 1 second resulting in large datasets. By combining a SOM clustering technique and radial basis function neural network (RBFNN), patterns are revealed while reducing redundant information. The performance of the WT failure prediction system was evaluated and tested using 37 sensors from SCADA data for a single WT in Levenmouth, Scotland. A failure state classification accuracy of between 96%-99% was observed.
Keywords wind energy, self-organizing map, k-means, radial basis function neural network, clustering, classification