Data analytics for fault prediction and diagnosis in wind turbines
University College Cork
As feed-in tariffs for wind energy are gradually being replaced by market driven auction-based systems, the need for cost savings at every stage of a wind energy project is more apparent than ever. A proven and effective way of reducing maintenance costs is through a condition-based maintenance (CBM) strategy. By using supervisory control and data acquisition (SCADA) system data instead of retrofitting a dedicated condition monitoring (CM) system, CM functionality can be gained at a fraction of the cost. This thesis investigates using SCADA system data for various levels of CM: fault detection, diagnosis and prediction. First, a case study is presented on using classification techniques for CM using SCADA data. Various methods for dealing with the massive class imbalance seen in fault data are evaluated. It was found that all three levels of CM are possible using classification techniques, though with a high number of false positives. Adding a class-weight to the minority class or undersampling the majority class were found to be the best ways of dealing with class imbalance. Sources of accurate failure data can be difficult to obtain for wind turbines. The second part of this thesis presents a novel way of building a historical failure database using alarm system and availability data. This was shown to produce an accurate database of unplanned stoppages related to assembly-level failures, scheduled maintenance, or grid, noise or shadow-related events. Next, common issues with some of the classification approaches present in the literature are addressed, as well as the lack of demonstration of how these approaches would perform in the field. A formalised framework with a prescribed list of steps following best practice guidelines is presented for performing CM using classification techniques on turbine SCADA data. A case study is performed which uses a sliding window metric to evaluate field performance, showing that such a system is effective at flagging faults in advance, but more data is needed to reduce the false positive rate. It is noted throughout the thesis that turbine alarm systems have some consistent shortcomings, and do not live up to their full potential. Hence, a novel methodology is presented which uses clustering techniques to identify similar sequences of alarms as they occurred during unplanned stoppages. A case study applying the methodology showed that just under half of the 456 stoppages could be sorted into one of fifteen distinct types of alarm sequence.
Wind energy , Condition monitoring , Scada data , Machine learning
Leahy, K. 2018. Data analytics for fault prediction and diagnosis in wind turbines. PhD Thesis, University College Cork.