An Approach for Utilizing Correlation among Sensors for Unsupervised Anomaly Detection of Wind Turbine System

An anomaly detection model needs to be built by considering the characteristics of a target system. In this paper, we study unsupervised anomaly detection for a wind turbine system. A wind turbine system has many components and sensors of the same components are highly correlated. Considering these characteristics of a wind turbine system, our approach for unsupervised anomaly detection is to utilize correlation among sensors. We first generate a standard correlation matrix (i.e., the most representative normal correlation matrix) from the given normal data. Then, we measure a distance between the standard matrix and a correlation matrix of a target test data using mean squared error as an anomaly score. Through experiments using real wind turbine system data, we show that our unsupervised approach achieves a maximum of 0.928 ROC AUC (0.894 on average),