State-of-the-art deep learning methods have demonstrated impressive performance in the intelligent fault diagnosis of rolling element bearings. However, in previous studies, critical issues such as domain discrepancy and the inability to interpret a classification decision made it difficult to apply deep learning in real industrial scenarios. Although domain adaptation and domain generalization-based methods have been investigated to solve domain discrepancy, collecting labeled data for various domains (especially continuous and non-stationary working conditions) is extremely difficult in an engineering application. Furthermore, since the classification decision cannot be physically explained, serious reliability problems arise for unseen working conditions (i.e., target domain with domain discrepancy). This study proposes the single domain generalizable and physically interpretable (SDGPI) framework. The proposed model embeds prior knowledge into the neural network combined with signal-preprocessing, which simultaneously enables single source domain generalization and domain interpretation with physical guarantees. Comprehensive case studies demonstrate that domain generalizable representation leads to 1) superior performance and robustness compared with existing methods for various untrained working conditions, as well as 2) efficient data inference even with small data size. Finally, the diagnosis results could be physically understood by displaying the classification decision in terms of the theoretical characteristic fault frequency (i.e., the characteristic fault order), indicating that SDGPI is a versatile and reliable diagnostic tool for unseen working conditions.