A novel meta-learning method based on relation network for train bearing fault diagnosis

Bearing, being an essential element in trains, assumes a central position in guaranteeing the safety of train operations. However, due to the scarcity of fault samples and significant differences in data distribution, existing fault diagnosis methods often fail to provide accurate and reliable results. This highlights the urgent need for innovative approaches that can address these challenges, particularly under conditions of limited labeled data and domain shifts. To tackle this problem, we propose a approach called novel relation network meta learning (NRNML) specifically designed for train bearing fault diagnosis. Firstly, we introduce a dual-scale residual neural network that aims to enhance the ability of the model to extract features at different scales. Additionally, we propose a novel semi-supervised relation calibration method that leverages unlabeled data to refine the generated class prototypes and rectify their relation score, thereby augmenting the model's performance in fault diagnosis. Based on this foundation, we incorporate the Ranger optimizer to improve convergence speed and diagnostic performance. Our experimental results demonstrate the effectiveness of NRNML. Through ablation experiments and fault diagnosis cases, the proposed approach outperforms other methods. NRNML demonstrates higher classification accuracy, strong feature extraction, and clustering advantages, confirming its potential for practical train bearing fault diagnosis