The deep-learning (DL) techniques are now pervasive in the realm of fault diagnosis for rolling bearings, more specifically, fault type classification. However, DL models often confront two problems in real applications: generalizability and interpretability. In this paper, a regularization term is proposed inspired by the phenomenon that the saliency analysis results of training samples within the same fault types are actually quite different, which are thought to be similar according to the fault-pattern knowledge. Then, a DL diagnostic framework containing the proposed regularization term and a simple convolutional neural network is established. The framework is compared with other advanced DL models in a fault diagnosis experiment for bearings. It is found that the DL framework does not underperform other models overall if the test samples are not added with additional noise. However, if the test samples are added with additional noise, the test accuracy of the proposed framework is higher than other models, suggesting a better anti-noise ability brought by the proposed regularization term, which is realized in a physics-informed and interpretable way.