Anomaly detection indicators of a wind turbine gearbox based on feature extraction from its vibration performance

This paper proposes a method for obtaining several health condition indicators for wind turbines based on vibration data driven from two similar experimental turbines (damaged and healthy). These indicators are able to capture the bearing and gear condition of the gearbox in the wind turbines. Signal processing and feature extraction were carried out –on both the time and frequency domains– from raw data in order to generate datasets for each shaft of power of the wind turbines. Based on good health condition data, a data mining approach was used to build two reference models for the indicators, one using Self-Organizing Maps (SOM) and another one using Gaussian Mixture Models (GMM). These reference patterns for the indicators were tested with a dataset coming from a damaged wind turbine and the results obtained confirmed the adequacy of these indicators to detect anomalies in the health condition of a wind turbine.