Graph-based semisupervised learning for icing fault detection of wind turbine blade

Abstract

The detection of blade icing faults in wind farms is an important task in improving the reliability and safety of wind power systems. Detection is primarily achieved through supervised learning, using labeled samples. However, labeling a large volume of data often consumes lots of manpower, which means that semisupervised learning is usually required. This study proposes a graph-based semisupervised learning system to achieve efficient fault detection in wind turbines (WTs). The proposed method detects abnormal samples by optimizing a small amount of labeled and a large amount of unlabeled data. Firstly, feature selection based on mutual information is performed on supervisory control and data acquisition data, so as to eliminate redundant variables. Secondly, a graph is constructed, using the limited labeled data to indicate the connections between samples. Finally, the initial class information is propagated to the entire data set by iteratively updating the input of the graph model. Experimental results based on a selection of benchmark data sets and real WT data demonstrate the feasibility of the proposed method.