Abstract
This paper presents autoencoder (AE)/Gaussian mixture model (GMM) tandem connectionist anomaly detection for an efficient deployment of wind turbine anomaly detection systems. In this method, robust features are extracted using AE that is trained with a large variety of normal state data and taken as inputs to an anomaly identifier based on a GMM for a specific target machine. Since the AE is trained with data from various machines, the feature representations obtained using this AE can be robust against the difference in the machine types and operation conditions of wind turbines. Experimental comparisons conducted using vibration signals from wind turbines demonstrated that the proposed method achieved ideal anomaly detection, which detects anomalies without miss detections and with few false alarms. Even when the training data were small, the proposed system gave better performance than existing systems, showing its effectiveness in early operation.
