Anomaly Prediction of Wind Turbines using Metric Learning with Vibration Data

Machine learning methods have attracted interest in automatically detecting and predicting anomalies in wind turbines by exploiting condition monitoring system data, such as vibration data, and supervisory control and data acquisition data. These techniques are expected to improve the profitability of a wind power generation business by reducing maintenance costs and downtime. For the anomaly detection and prediction of wind turbines, an unsupervised learning method is preferable because typical anomaly patterns are difficult to obtain. Autoencoder-based (AE) techniques, which do not necessitate anomaly patterns, have also been studied as an unsupervised learning method for detecting and predicting malfunctions in wind turbines. In particular, vibration data are considered practical because they reflect the changes of the values in the early stages of a malfunction. Although AE-based techniques can achieve better prediction performance, it is difficult to treat vibration data obtained from multiple sensors mounted on a wind turbine because of their different sampling timings and rates. The use of the multiple vibration data is expected to contribute to a better prediction performance. In this study, we propose a metric-learning-based method to cope with multiple data to improve prediction performance. We demonstrate the performance improvement in comparison with an AE-based method using vibration data captured from a real wind turbine.