Semi-Supervised Prototype Networks With Similarity Information Selection for Fault Diagnosis of Wind Turbine Gearboxes

Deep learning has been widely researched for fault diagnosis of wind turbine (WT) gearboxes. However, a major challenge is the inability to obtain a large number of labeled samples to train deep learning models. To deal with this problem, a semi-supervised prototype network (ProNet) with similar information is proposed for gearbox fault diagnosis in this article. First, the ProNet is trained with limited labeled samples to obtain the metric space of features. Then, according to the sample distribution relationship in the metric space, a new pseudo-labeling strategy based on similarity information selection is proposed to obtain samples of entropy reduction as pseudo-labeling samples. Meanwhile, the coefficient of variation weighted classification loss is adopted to reduce the jumps phenomenon. Finally, extensive experiments on experimental datasets and WT gearbox datasets verify that the proposed method has a better identification ability under limited labeled samples.