Few-Shot Cross-Domain Fault Diagnosis of Transportation Motor Bearings Using MAML-GA
Huimin Zhao, Chao Liu, Xiangjun Dang, Junjie Xu, Wu Deng
Abstract
The dynamic working conditions and scarcity of fault samples in train motors pose significant challenges to the generalization capability of diagnostic models. To address this issue, this paper proposes a cross-domain fault diagnosis method with strong generalization capability under few-shot conditions, termed MAML-GA. In the meta-task construction phase, a dynamic meta-task augmentation mechanism is introduced within the meta-learning framework to alleviate the problem of limited data by expanding the sample space and generating virtual meta-tasks. During the training phase, a parameter update operator is designed to regulate the inner-loop gradient descent process, thereby optimizing the training strategy and preventing the model from overfitting to noisy or irrelevant features. These two mechanisms significantly enhance the cross-domain generalization ability of the diagnostic model and reduce the risk of overfitting under few-shot scenarios. Furthermore, a lightweight key feature extraction network is developed as the backbone of the proposed method. This network integrates a multi-dimensional spatial-channel attention module, which improves the extraction of discriminative fault features under both limited data and dynamic working conditions. Experimental validations conducted on both a metro transmission system dataset and a laboratory dataset demonstrate that the proposed method achieves diagnostic accuracies of 99.46% and 97.42%, respectively. Compared with mainstream methods such as ProtoNet and RelationNet, the proposed method shows improvements of 2–6% and 6–15%, respectively, thereby confirming its effectiveness in handling variable rail transit conditions and data scarcity.