Radio Frequency Signal Recognition of Unmanned Aerial Vehicle Based on Complex-Valued Convolutional Neural Network

Area of Science:

• Artificial Intelligence
• Signal Processing
• Aerospace Engineering

Background:

• Unmanned aerial vehicle (UAV) recognition is crucial due to rapid technological advancements.
• Conventional real-valued CNNs (RV-CNNs) struggle with radio frequency (RF) spectrograms, especially in low signal-to-noise ratio (SNR) environments, due to discarded phase information.
• Existing methods often exhibit degraded performance under noisy conditions.

Purpose of the Study:

• To develop a robust RF-based recognition method for UAVs that overcomes the limitations of RV-CNNs in low SNR conditions.
• To propose and evaluate a complex-valued CNN (CV-CNN) that effectively utilizes both magnitude and phase information from RF signals.
• To investigate the impact of architectural choices on CV-CNN performance and identify stability constraints.

Main Methods:

• A complex-valued CNN (CV-CNN) was designed, utilizing a complex representation of RF spectrograms (logarithmic PSD as real part, Sobel edge detection as imaginary part).
• Genuine complex convolutions were employed to fuse magnitude and structural cues, enhancing noise resilience.
• Comprehensive ablation experiments were conducted to analyze hyperparameter sensitivity and network depth limitations.

Main Results:

• The CV-CNN achieved 100.00% accuracy on the 25-class DroneRFa dataset under noise-free conditions.
• In low SNR regimes, the CV-CNN demonstrated superior robustness: achieving 15.58% accuracy at -20 dB SNR (over seven times higher than a dual-channel RV-CNN) and 45.86% at -15 dB SNR.
• Performance degradation was observed beyond 4-5 network depths, indicating critical stability constraints.

Conclusions:

• The proposed CV-CNN offers significantly improved robustness and interference resistance for RF-based UAV recognition compared to real-valued counterparts.
• The complex-valued approach effectively leverages RF signal characteristics, maintaining high accuracy even in challenging low-SNR environments.
• The findings highlight the potential of CV-CNNs for reliable UAV identification in diverse operational conditions.