A meta-learning method based on meta-feature enhancement for bearing fault identification under few-sample conditions

Recently, deep learning has achieved remarkable success in the field of rolling bearing fault diagnosis. However, two issues cannot be ignored: 1) Deep learning models typically require a large amount of labeled data for training, yet fault data is extremely scarce; 2) The decision- making process of the models lacks interpretability. In this paper, a novel meta-learning method based on meta-feature enhancement is proposed and applied to few-shot bearing fault identification across different working conditions and test rigs, which is called meta-feature enhancement meta-learning (MFEML). Within this method, a meta-feature enhancement module and an adaptive squaring module are proposed, which respectively enhance the convolutional network model's ability to recognize fault features in complex signals and improve its adaptability to varying signal lengths. In addition, through a dual iterative optimization process, the initial parameters of the base model are adjusted, enabling it to learn meta-knowledge from sparse samples across different tasks. Finally, the proposed MFEML method is experimentally proved through two datasets from different labs. Noise is introduced to mimic real industrial conditions, further validating the effectiveness and practicality of MFEML. Additionally, an interpretability analysis is also performed on the model's outputs.