Semi-Supervised Temporal Meta-Learning Framework for Wind Turbine Bearing Fault Diagnosis Under Limited Annotation Data

Recently, deep learning has made brilliant achievements in wind turbine bearing fault diagnosis field. However, there are two problems that cannot be ignored: 1) the fault data is so scarce that it is time-consuming to acquire a well-behaved deep learning model; 2) much unlabeled data cannot be adequately utilized to explore useful fault information without prior. Therefore, a novel semi-supervised temporal meta learning method (SSTML) is proposed, which can not only probe representative deep features from massive raw unlabeled vibration data adequately, but also make the best of small annotation data to complete fault identification tasks. Transplanting meta learning ideas into semi-supervised learning, a novel deep learning framework - SeMeF is proposed. The proposed SeMeF is capable of drawing on the advantages of two mechanisms to exert efficiency beyond themselves. Furthermore, a temporal convolutional module is proposed to relieve overfitting due to the depth of the model, which can fully excavate temporal features along the depth of the network. The superiority of the proposed method is demonstrated on the wind turbine bearing dataset. Experimental results indicate that the model proposed can reach high diagnostic accuracy with limited annotation data, which outperforms many advanced deep learning models.