Kailin Bi

Abstract Traction motor bearings are crucial for guaranteeing the safe operation of metro vehicles. However, in the metro traction motor bearing fault diagnosis, there are usually problems of small samples and missing fault samples, leading to inaccurate results. Therefore, a novel bearing fault diagnosis method utilizing a track-vehicle-bearing coupled dynamic model and the improved deep convolutional generative adversarial network-multiscale convolutional neural network with mixed attention (IDCG-MAMCNN) model is proposed in this paper. The IDCG-MAMCNN model combines an improved deep convolutional generative adversarial network (IDCGAN) with a multi-scale convolutional neural network with mixed attention (MA-MCNN). Specifically, simulation data is first provided by the coupled dynamic model to supplement missing fault samples. Secondly, the IDCGAN, along with a training method that involves pre-training models with simulation samples and fine-tuning models with experimental samples, is introduced to generate high-quality samples and augment experimental samples under small samples. Lastly, the MA-MCNN serves as the classification model, trained with the augmented dataset comprising experimental, simulation, and generated samples. The fault diagnosis performance of the proposed method is evaluated on the experimental samples of two bearing datasets under small samples and various conditions of missing fault samples. It has been demonstrated by the experimental results that the proposed method exhibits robust fault diagnosis performance and generates high-quality samples under small samples and missing fault samples. Furthermore, the proposed method showcases its adaptability to different operation speeds.

Abstract The compound problem of few-shot and data imbalance remains a major obstacle in rolling bearing fault diagnosis. To address this issue, this paper proposes IMSSiT-SSL, a self-supervised learning framework that integrates prior knowledge with simulation data-driven approaches. Unlabeled simulation datasets are used for self-supervised pre-training through two physics-informed pretext tasks: a kurtosis-guided masked reconstruction task that selectively masks high-kurtosis segments to force the model to capture intrinsic fault features, and a characteristic indicator-based pseudo-label prediction task that explicitly embeds domain expert knowledge into the representation learning process. This dual-task pre-training strategy alleviates the few-shot challenge while directing the model’s attention toward fault-relevant impact segments, improving the physical interpretability of feature extraction. Subsequently, a small amount of labeled data is used for fine-tuning, where an adaptive boundary-regularized dual focal loss is introduced to encourage the model to focus on discriminating minority class samples, thus mitigating the negative effects of imbalanced data. Experimental results demonstrate that the proposed method achieves G -mean improvements of 3.07%–20.62% over competing methods across various imbalance ratios, and maintains a diagnostic accuracy of 98.78% under cross-speed few-shot conditions with only 30 labeled samples per class.