Fault diagnosis of rotating machinery is essential for condition monitoring and predictive maintenance. Although deep learning methods have achieved high classification accuracy in fault recognition from rotating machinery signals, their diagnostic processes are often weakly linked to physical fault mechanisms, which limits their credibility and verifiability in engineering applications. To address this issue, this paper proposes an interpretable model based on a physics-informed Transformer and Koopman operator for rotating machinery fault diagnosis. Instead of relying solely on black-box representations, the proposed method introduces common physical diagnostic evidence, including time-domain impulsiveness, frequency-domain distribution, shaft-frequency harmonics, and envelope modulation, into the observable representation space. A multi-head observable mechanism is further designed to explicitly model different sources of diagnostic evidence. In addition, a Koopman-based dynamic consistency constraint is introduced to enforce coherent evolution between adjacent signal segments in the observable space, thereby improving the stability and physical consistency of latent representations. Unlike conventional post-hoc explanation methods, the proposed model embeds interpretability into its architecture, enabling diagnostic evidence to be analyzed from overall, class-level, and samplelevel perspectives through observable-head weights. Experiments on the Southeast University (SEU) drivetrain fault diagnosis dataset show that the proposed method achieves high fault recognition performance and improved noise robustness under moderate noise conditions. The observable-head weighting patterns align well with the diagnostic importance of corresponding physical evidence, thereby providing physically meaningful support for different fault classes.