A meta-transfer learning with natural language supervised pretraining for few-shot fault diagnosis of bearings

The intelligent deep learning algorithms have achieved a great success in bearing fault diagnosis. However, most of them rely on data-driven optimization or prior experience without considering physical fault characteristics. Moreover, it requires a large amount of labeled data with expensive computational resources for model training, which is not always available in insufficient training samples. To address these, this article proposes a meta-transfer learning method with nature language supervised pretraining for few-shot fault diagnosis of bearings. First, a physics-guided feature extraction backbone network is designed, which incorporates cepstrum and entropy to leverage periodic features of vibration signals. Furthermore, a meta-transfer learning on scaled sampling layers is achieved under limited training samples. Finally, to overcome the drawback of one-hot encoding for meta-transfer learning, a nature language supervised pretraining is devised to enhance the complex feature representation in the source domain. Several experiments are performed to verify the effectiveness of the proposed method. On the local dataset for meta-learning and meta-test, an accuracy of 95.9% is achieved in the 5-way-5-shot task mode.