A robust ATUB-Net for bearing fault diagnosis under unbalanced sample scenarios

In order to achieve outstanding fault diagnosis classification performance, most deep learning-based intelligent fault diagnosis methods require a large number of labeled samples for training. However, collecting sufficient labeled fault samples in practice is challenging because the process is time-consuming and labor-intensive, resulting in an imbalanced dataset. Aiming at the problem of degraded fault diagnosis performance due to severely skewed data distribution in industrial scenarios, this paper proposes an anti-bias fault diagnosis network, ATUB-Net, to achieve high-precision classification under unbalanced data through a multilevel feature reconstruction and attention mechanism. By designing the anti-imbalance mode, separating the normal and faulty sample feature streams with multi-branch feature cuts, enhancing the extraction ability of local high-frequency mutation signals by combining 3 × 3 group convolution, and dynamically reinforcing the key fault regions and suppressing the noise interference through the compression while fusion unit integration channel and spatial attention mechanism. Experiments show that in two different extreme imbalance datasets, the accuracy and G-Means metrics of ATUB-Net significantly outperform those of other mainstream models, and the Grad-CAM visualization shows that its thermal distribution accurately focuses on the faulty time-frequency mutation region, and the T-SNE feature clustering boundary is clear.