PeirceTriadicNet: A Peircean Triadic Three-Branch Fusion Network for Non-Cooperative Signal Recognition in UAV Communications

Non-cooperative signal recognition is essential for spectrum awareness and interference management in unmanned aerial vehicle (UAV) communications, where unknown emitters, jamming threats, and low signal-to-noise ratio (SNR) make modulation classification and parameter estimation particularly challenging. Existing approaches often treat modulation classification, parameter estimation, and semantic reasoning as isolated tasks, making it difficult to exploit the complementarity of cross-domain information. This paper proposes PeirceTriadicNet (PTN), a Peircean triadic-relation network for this problem. The Representamen branch processes observable signal representations (IQ waveforms, time–frequency maps, and power spectral densities), the Object branch embeds physical source parameters (modulation type, symbol rate, carrier-frequency offset, roll-off, bandwidth, and SNR), and the Interpretant branch incorporates modulation-family and SNR priors for semantic reasoning. The final model, PTN-ResIcon, fuses the three branch embeddings with fixed equal weights (αR=αO=αI=1/3). On RML-S-Full-Std, a UAV-oriented synthetic dataset that re-generates the RML2016.10a modulation set under representative UAV link conditions (Rician fading, Doppler shift, and barrage, partial-band, follower, and deceptive jamming), with 220,000 IQ frames of 11 modulation formats over −20 to +18 dB, per-frame physical-parameter and semantic annotations, and a released open-data generation script—the covered PSK/QAM OFDM video downlinks and FSK/CPM frequency-hopping control links mirror the signal families observed in real commercial drone RF links—PTN-ResIcon achieves an overall modulation-recognition accuracy of 80.67%, which is 17.6 percentage points higher than the Representamen branch alone and 24.0, 20.1, and 20.2 percentage points higher than CNN2, CNN-LSTM, and ResNet1D, respectively; the +18 dB accuracy is 91.50%. On the real DroneRFa corpus (USRP-acquired recordings of four DJI devices at three flight distances), the complete triadic configuration reaches 72.50% device-identification accuracy, demonstrating that the framework transfers from synthetic to genuine UAV signals. An SNR-biased learnable-attention variant, PTN-ResIcon-Learned, dynamically adjusts the triadic weights from the oracle SNR and branch confidences; it reaches 76.86%, showing that fixed equal-weight triadic fusion itself provides the main gain and that overly aggressive SNR gating slightly degrades generalization. In the UAV-swarm scenario of RML-S-Full-Std, PTN-ResIcon maintains an accuracy of 81.44%, demonstrating its applicability to UAV communication environments and to spectrum-awareness tasks in emergency-response missions.