Anomaly detection in unmanned aerial vehicle (UAV) flight data remains challenging because purely data-driven models cannot fully capture the temporal dependencies and physical couplings inherent in flight dynamics. This paper proposes a Physics-Informed Dual-Residual Transformer (PI-DRFormer) that integrates temporal modeling, physics-informed variable attention, and gray-box dynamic constraints. A dual-branch architecture captures temporal dependencies and physically related intervariable interactions, while physical relations regularize nominal prediction. During inference, prediction and physical residuals are combined into a unified anomaly score, and the detection threshold is calibrated exclusively from anomaly-free validation data. Experiments on real-world UAV flight data show that PI-DRFormer achieves consistently high detection accuracy and low false-positive rates, outperforming representative baseline methods

UAV flight data, anomaly detection, physics-informed learning, dual-residual detection