Rotating machinery fault diagnosis is important for equipment condition monitoring and predictive maintenance. Although deep learning methods have achieved high recognition accuracy, their diagnostic processes often lack physical interpretability, which limits their credibility in engineering applications. To address this issue, this paper proposes a physics-informed Transformer-Koopman interpretable model for rotating machinery fault diagnosis. A six-channel signal representation is first constructed by integrating normalized waveform, first- and second-order differences, Hilbert envelope, local root mean square, and spectral magnitude mapping, which respectively capture impulsive components, signal trend variations, amplitude modulation, local energy fluctuations, and frequency distribution. A Transformer encoder is then used to extract global temporal features. Physics-informed Koopman observable heads are designed to decouple deep features into five diagnostic evidence subspaces, including time-domain statistics, frequency-domain structures, shaft-frequency harmonics, envelope modulation, and data-driven evidence. A Koopman dynamic consistency constraint is further introduced to regularize the evolution of adjacent signal segments in the observable space. Experimental results on the Southeast University (SEU) drivetrain fault diagnosis dataset show that the proposed method achieves high diagnostic accuracy and good robustness under moderate noise conditions. Observable head analysis further indicates that the learned evidence allocation is consistent with rotating machinery fault mechanisms, providing interpretable decision bases for model predictions.