Improving the performance of fault diagnosis network via interpretable projection information

The performance of fault diagnosis networks for rotating machine mainly depends on the classifiers and feature extractors. To improve the performance of the two components, a novel regularization and a modified normalization module are proposed based on the interpretable projection information. Specifically, the classifier is firstly studied from the perspective of projection mechanism rather than the traditional base representation. It is found that a negative correlation is beneficial to classifiers, and a corresponding regularization is proposed for improving its ability. Meanwhile, by analyzing the convolution operations in convolution networks from the projection perspective, we find that the projection information will be affected by the traditional batch normalization block followed with a rectified linear unit, if the batch size is huge and the number of classes is small. To solve this problem, a novel normalization module, which is designed based on absolute value operation, is proposed to fully retain the projection information while effectively suppressing the noises in features. Finally, the proposed method is used to improve several typical fault diagnostic networks, and the fault diagnosis experiments on rolling bearings and planetary gearboxes demonstrate that the proposed method can effectively improve the performance of diagnosis networks.