Method for Classification of UAV Flight Control RF Signals Based on Multi-scale Divergence Entropy and Optimized Neural Networks

To address the insufficient classification accuracy of UAV flight control radio frequency (RF) signals in complex electromagnetic environments, this paper proposes a classification method based on Multiscale Divergence Entropy (MDE) feature fusion and BP neural network optimized by the Artificial Lemming Algorithm (ALA). Firstly, MDE extracts dynamic features of RF signals across multiple time scales, and grid search optimizes multi-scale factors to construct a 12-dimensional feature matrix with high discriminability. Secondly, the Artificial Lemming Algorithm globally optimizes the weights, biases, and hidden layer nodes of the BP neural network, dynamically balancing exploration and exploitation to avoid local optima. Experiments were conducted on the DroneRFa dataset covering RF signals from six mainstream UAV models. Results demonstrate that the proposed method (ALA-BP) outperforms traditional optimization algorithms (GA-BP, PSO-BP, etc.) in classification accuracy (97.2%), noise robustness (90% accuracy at SNR=0 dB), and convergence speed (reaching 90% accuracy in 65 iterations). ROC curve analysis (AUC=0.97) further confirms the model’s robustness, providing an efficient technical solution for airspace security supervision and low-altitude traffic management.