Bearing fault diagnosis method based on improved meta-ResNet and sample weighting under noise label

Traditional methods often require the presetting of a weight function based on the dataset to address the issue of noisy labels. However, these methods often encounter challenges related to poor generalization capability. To overcome this obstacle, we propose an improved meta-residual network and sample weighting (SWMeta-IResNet) approach for bearing fault diagnosis. This method leverages singular value decomposition (SVD) matrix decomposition technology to design a global SVD pooling layer. By replacing the max-pooling layer in the original ResNet, this layer effectively reduces the parameter count and enhances the correspondence between feature maps and categories. This method trains the network through alternating input of a small number of unbiased, cleanly labeled samples (meta-samples) and noisy labeled samples. By automatically learning the weight function mapping relationship between the training loss and sample weight from the data, it adaptively learns weights from the meta-samples to improve accuracy. Experimental results on three datasets demonstrate that, even with a noise label rate of 40%, SWMeta-IResNet achieves significant improvements in average accuracy compared to the original ResNet model. Specifically, it enhances the average accuracy by 14.5%, 11.94%, and 6.38%, respectively, yielding accuracy rates of 86.48%, 82.23%, and 94.35%. Moreover, in the bearing failure task with noisy labels, this method exhibits substantial improvements in accuracy and showcases excellent generalization performance across different datasets. As a result, SWMeta-IResNet proves to be highly applicable and effective in addressing the challenges posed by noisy labels in diverse scenarios.