Abstract
Adaptive interference suppression in complex electromagnetic environments has become a key challenge to ensure communication reliability. Traditional methods rely on prior knowledge of interference and have insufficient adaptability in dynamic scenarios. The standard CycleGAN model suffers from limited feature extraction capability, insufficient receptive field of the discriminator, and loss function not optimized for communication performance when directly processing one-dimensional radio frequency signals. This paper proposes an improved CycleGAN model for low-altitude communication interference suppression. The model introduces residual dense blocks and cross-layer identity mapping in the generator to enhance feature reuse and gradient flow, and adopts large receptive field initialization to capture global context; multi-scale dilated convolutions and a refined PatchGAN structure are integrated in the discriminator to expand the receptive field and achieve accurate local discrimination; meanwhile, signal-to-interference-plus-noise ratio loss and signal reconstruction loss are incorporated into the loss function to directly optimize communication quality and signal fidelity. Experiments on the public drone radio frequency dataset show that the proposed method outperforms the comparison algorithms (LMS, DnCNN and original CycleGAN) significantly in signal-to-interference-plus-noise ratio gain (12.4 dB), bit error rate (4.8×10⁻⁴), peak signal-to-noise ratio (36.9 dB) and structural similarity (0.982). Robustness tests further verify that the method can maintain stable performance under various interference types (narrowband, pulse, co-frequency) and different signal-to-interference ratio conditions. This work provides an effective deep learning solution for robust and adaptive interference suppression in complex low-altitude electromagnetic environments.