Graphical temporal semi-supervised deep learning–based principal fault localization in wind turbine systems

Abstract

Principal fault localization of the faults chain, as a branch of fault diagnosis in wind turbine system, has been an essential problem to ensure the reliability and security in the real wind farms recently. It can be solved by machine learning techniques with historical data labeled with principal faults. However, most real data are unlabeled, since the labeled is expensive to obtain, which increases the difficulty to localize the principal fault if just using unlabeled data and few labeled data. So, in this article, a novel approach using unlabeled data is proposed for principal fault localization of the faults chain in wind turbine systems. First, a deep learning model, stacked sparse autoencoders, is introduced to learn and extract high-level features from data. Then, we present a graphical temporal semi-supervised learning algorithm to develop the pseudo-labeled data set with an unlabeled data set. Considering the temporal correlation of wind power data, we add a time weight vector and apply the cosine-similarity in the proposed algorithm. Finally, based on the pseudo-labeled data set, a classifier model is built and trained for the principal fault localization of the faults chain. The proposed approach is verified by the real buffer data set collected from two wind farms in China, and the experimental results show its effectiveness in practice.