DWEL-Net: Dynamic Weighted Ensemble Learning for Robust Modulation Classification in UAV Networks

Automatic modulation classification (AMC) is critical in non-cooperative communication systems, particularly in unmanned aerial vehicle (UAV) networks, where dynamic channel conditions and heterogeneous noise pose substantial challenges, leading to significant degradation in classification accuracy. To address these challenges, this paper proposes a Dynamic Weighted Ensemble Learning (DWEL) framework that robustly enhances modulation recognition through adaptive weight fusion. Specifically, DWEL employs a sigmoid-based accuracy model to capture the nonlinear dependency between signal-to-noise ratio (SNR) levels and classification reliability, dynamically adjusting node weights in real time. Additionally, a hybrid fusion mechanism that strategically integrates probabilistic averaging with majority voting is proposed to mitigate transient decision errors under Rayleigh fading conditions. Extensive simulations demonstrate that the DWEL framework achieves an accuracy exceeding 95% at 10 dB SNR in a 15-node cooperative network scenario of UAVs. Moreover, in a representative 5-node setup, DWEL surpasses state-of-the-art methods such as weighted averaging (WA) and Bayesian fusion by approximately $15 \%$ to $50 \%$ at 0 dB. The core algorithm code is publicly accessible at the following link: https://github.com/ZHR-HEU/Dynamic-Weighted-Ensemble-Learning-for-Robust-Modulation-Classification-in-UAV-Networks.