Unsupervised Ranking of Outliers in Wind Turbines via Isolation Forest with Dictionary Learning

Abstract

Predictive maintenance strategies for the detection of faults in wind turbines require approaches that consider the limited human resources available responsible for the final assessment of the machine. Here, we present an unsupervised framework for the ranking of wind turbines (assets) in a wind farm (fleet) based on the detection of faults in the monitored machine elements, which will help experts determine the turbine that has higher priority in further machine diagnostics. Previous work has shown that the use of sparse coding with dictionary learning enables the identification of faults in rolling element bearings. However, it has not shown how the information of the identified faults can be used in an unsupervised strategy that enables a detector to provide some sort of recommendation on how to proceed with the information of the detected faults. We describe how features derived from the sparse coding with dictionary learning method are used together with the isolation forest outlier detection algorithm to create a score for the ranking of the monitored assets. We consider scenarios where all the turbines are evaluated together and each of them individually in the creation of the ranking and we compare these results with a condition where features taken from the time-domain are considered. Sparse coding with dictionary learning together with isolation forest produces an anomaly score that can be used to rank wind turbines by their need for a maintenance action given the presence of faults in their systems resulting in an unsupervised warning system that can support the work of maintenance experts.