Residual Signature Analysis for Wind Turbine Failure Prognosis

In operation and maintenance of wind turbines, predictive maintenance is widely used to reduce the downtime and low operating efficiency due to failure of a wind turbine, in which the condition of a wind turbine is continuously monitored and measured by attaching a number of sensors, sign of near-future failure or degradation is detected early before the actual failure occurs, and faulty parts are replaced or maintenance is performed. In this context, automatic prognosis of failure by analyzing the time-series data of the attached sensors is very important. Recently, various machine learning techniques are proposed where sign of failure is predicted based on the residual errors between the actual condition and predicted condition of a wind turbine with a prediction model built by using the time-series data. However, when the residual errors fluctuate due to external factors, this conventional approach may generate many false alarms and/or overlook potential points of failure prognosis. In this paper, we propose Residual Signature Analysis (ReSA) that generates anomaly scores by applying signal processing and clustering techniques on the segmented residual errors, which are generated by applying a prediction method on the time-series data. To show the effectiveness of the proposed method, we apply ReSA to three wind turbines of two publicly available datasets and show the experimental results. Our experimental results suggest that our proposed method can detect sign of failure much earlier than actual time of failure with a very low rate (<0.75%) of false alarms compared to conventional residual analysis-based approach.