A novel meta-learning network with adversarial domain-adaptation and attention mechanism for cross-domain for train bearing fault diagnosis

Abstract Traction motor bearings, serving as a critical component in trains, have a significant impact on ensuring the safety of train operations. However, there is a scarcity of sample data for bearing failures during train operations, and the complex and variable operating conditions of train bearings result in significant differences in domain distribution. Traditional cross-domain fault diagnosis methods are no longer adequate for addressing train bearing faults. Therefore, this study proposes a novel adversarial domain-adaptation meta-learning network (NADMN) for the purpose of diagnosing train bearing faults. Firstly, a deep convolutional neural network is proposed, which enhances the model’s feature extraction capability by incorporating attention mechanisms. Moreover, by employing domain adversarial adaptation learning strategy, it effectively extracts domain-invariant features from both source and target domains, thereby achieving generalization across different domains. Three experiments of bearing fault diagnosis are carried out, and the superiority of NADMN is proved by charts, confusion matrix and visualization techniques. Compared with the other five methods, NADMN showed obvious advantages in diagnostic scenarios characterized by significant changes in domain distribution.