Abstract: To address the limited accuracy of excitation-system fault diagnosis under complex operating conditions and the scarcity of labeled real fault samples, a fault diagnosis model based on physics-informed guidance and cross-domain representation transfer is proposed. The model uses a pre-trained time-series foundation model as the backbone and employs low-rank adaptation for parameter-efficient domain adaptation. On this basis, a multi-task learning architecture integrating physical-variable evolution constraints, fault classification, and signal-level attribution is constructed, and a two-stage transfer strategy consisting of simulation-domain adaptation and few-shot real-domain calibration is adopted to reduce the distribution discrepancy between simulated and real data. Simulation experiments on the IEEE 39-bus system show that the proposed model achieves an AUROC of 0.947 and a macro-F1 of 0.827, with an average action latency of 97.1 ms. Transfer experiments on real data show that only about 18% of the labeled samples are required to reach the predefined performance threshold for engineering application. The results show that the proposed method can reduce the demand for labeled real fault samples while improving the detection and classification performance of complex excitation faults and maintaining online diagnostic response at the hundred-millisecond scale.