Control chart-integrated machine learning models for incipient fault detection in wind turbine main bearing

Wind farm operators traditionally rely on SCADA temperature alarms for early signs of main bearing degradation. However, these alarms are sometimes delayed due to slow propagation of temperature in the main bearing. This study proposes the use of turbine multi-sensor vibration data as a viable alternative. Anomaly detection models including One-class support vector machine (OCSVM) and isolation forest (IF) models are first employed to compute anomaly scores from the dataset. An exponentially weighted moving average (EWMA) control chart receives the anomaly scores for fault prognosis. The study developed a methodology for identifying the optimal contamination fraction of the anomaly detection models and an index called the “anomaly model evaluation index (AMEI)” for evaluating the performance of the anomaly detection models. The optimal contamination fraction for the anomaly detection models was 4%. The IF model outperformed the OCSVM model, with an AMEI index of 5.84, in contrast to the OCSVM model’s score of 5.32. However, the OCSVM computed 18 times faster than the IF model. Furthermore, EWMA-IF achieved a higher True Positive Rate of 79.8% compared to 59.94% for EWMA-OCSVM, indicating a better ability to correctly identify abnormal observations. The EWMA-IF model alerted for an approaching main bearing fault six hours earlier than the EWMA-OCSVM control chart. The persistence of the anomaly scores above the threshold of the control charts provides evidence to suggest that a potential main bearing failure is imminent.