For unsupervised anomaly detection in aircraft engines, traditional methods predominantly focus on mining temporal features, often neglecting complex interdependencies among multiple variables. Although existing graph neural networks can model spatial relationships, their purely data-driven graph construction tends to overlook system thermodynamic mechanisms, thereby introducing spurious correlations. To address this, this paper proposes a Physics-Guided Spatio-Temporal Attention Graph Network (PG-STAG). This method fuses aerothermodynamic priors with health-data correlation information to construct a knowledge-data collaborative graph topology with physical consistency. Furthermore, a dual spatio-temporal attention mechanism is utilized to extract multi-dimensional spatio-temporal features, and anomaly identification is achieved based on reconstruction errors. Experiments on the N-CMAPSS dataset and real-world civil aviation QAR data demonstrate that PG-STAG achieves the highest F1-score and ROC-AUC among the compared methods, exhibiting strong robustness and physical interpretability.