Learning deep representation of imbalanced SCADA data for fault detection of wind turbines

Numerous intelligent fault diagnosis models have been developed on supervisory control and data acquisition (SCADA) systems of wind turbines, so as to process massive SCADA data effectively and accurately. However, there is a problem ignored among these studies. That is, SCADA data distribution is imbalanced and the anomalous data mining is not sufficient. The amount of normal data is much more than that of abnormal data, which makes these models tend to be biased toward majority class, i.e., normal data, while accuracy of diagnosing fault is poor. Aimed at overcoming this problem, a novel intelligent fault diagnosis methodology is proposed based on exquisitely designed deep neural networks. The between-classes imbalance problem is handled by learning deep representation that can preserve within-class information and between-classes information based on triplet loss. The proposed method encourages a pair of data belonging to same class to be projected onto points as close as possible in new embedding space. It tries to enforce a margin between different class data. The effectiveness and generalization of the proposed method are validated on the SCADA data of two wind turbines containing blades icing accretion fault. The result demonstrates the proposed method outperforms the traditional normal behavior modeling method.