FCLs-PcaTr: Flight control-link signals detection and classification via physics-consistent augmentation and tri-view representation

Reliable identification of unmanned aerial vehicles (UAVs) from wireless signals is a challenging intelligent signal analysis task, particularly when video transmission signals (VTS) are intermittent, weak, or unavailable. This paper presents a deep-learning framework for flight control-link signals (FCLs) detection and classification in time-frequency spectra, entitled Flight Control-Link signals Detection and Classification via Physics-Consistent Augmentation and Tri-View Representation. First, a tri-view representation, composed of a grayscale base map, a logarithmic compression map, and a local Z-score normalized map, enhances weak FCLs patterns from the perspectives of absolute energy, dynamic range, and local anomaly. Second, a physics-consistent combinatorial augmentation strategy synthesizes weak-signal and multi-class coexistence samples through target residual extraction and local background matching, thereby improving data efficiency and generalization. Third, a directional context harmonized backbone with single-scale aggregation explicitly models temporal, spectral, local, and dilated contextual structures for joint detection and classification. On the public DroneRFb-Spectra dataset, the proposed method achieves a precision of 0.974 and a recall of 0.959. Under multi-UAV testing on DroneRF2025 dataset, it achieves a precision of 0.973 and a recall of 0.935. The model requires only 0.721 million parameters and 1.313 billion floating-point operations, showing that the proposed intelligent system achieves a favorable balance between recognition performance and computational efficiency.