A novel multi-task transfer model to realize unsupervised fault diagnosis of newly constructed wind turbines under variable conditions

Abstract

Fault diagnosis can effectively improve the power generation of the wind turbines. Deep learning has promoted the intelligent development of wind turbine fault diagnosis. Traditional deep learning usually requires a sufficient amount of labeled data. However, for newly constructed wind turbines, there are problems such as insufficient samples, limited labels, variable operating conditions. Transfer learning provides a new way to solve these problems. Establishing appropriate models to reduce the distribution differences between existing and newly built units is the key to improving unsupervised fault diagnosis accuracy for newly built units. To address these challenges, a novel multi-task transfer model based on improved model agnostic meta learning (MT-TL-MAML) was proposed to realize unsupervised fault diagnosis of newly constructed wind turbines. The proposed model utilizes the advantages of MAML in generalization of new tasks and small sample diagnosis, and can adapt to randomly changing working conditions. By means of iterative learning, the gap between source domain and target domain is further narrowed, and the classifier can realize more accurate diagnosis of target domain data. This article takes the SCADA and CMS data of two wind farms as case studies to conduct unsupervised fault diagnosis and compare it with other literatures. The results validate the advantages of the proposed model in unsupervised fault diagnosis of newly constructed wind turbines.