A Signal Denoising Model Based on an Optimized GAN Network Using U-Net
ID：51
Submission ID：56 View Protection：ATTENDEE
Updated Time：2025-11-03 12:31:36
Hits：92
Oral Presentation
Start Time：2025-11-09 09:45 （Asia/Shanghai）
Duration：15min
Session：[S5] 5.AI-driven technology » [S5] 5.AI-driven technology
Presentation File
Tips: The file permissions under this presentation are only for participants. You have not logged in yet and cannot view it temporarily.
Abstract
Signal denoising in complex electromagnetic environments is a key challenge in 5G communication, autonomous driving, and other fields. Traditional methods (wavelet transform, empirical mode decomposition) suffer from feature degradation in non-stationary noise processing, and existing generative adversarial networks (GANs) have the defect of losing high-frequency details in one-dimensional signal denoising. To address these issues, this paper proposes a U-Net-based generative adversarial network (U-NetGAN) for signal denoising. The model combines the skip connection structure of U-Net with residual modules through a multi-scale feature fusion mechanism, achieving collaborative optimization of signal global trends and local transient features. A piecewise weighted local mean square error loss function is designed to enhance the detail reconstruction capability in critical regions. Experiments show that on a composite linear frequency modulation signal dataset, under 0 dB noise conditions, U-NetGAN achieves a signal-to-noise ratio increment (ΔSNR) of 22.97 dB, a 21.47% improvement over the best traditional method, and a structural similarity (SSIM) of 0.746. Additionally, the parameter count (3.1 million) is reduced by 35% compared to similar GAN models. Ablation experiments confirm that the skip connection mechanism contributes a ΔSNR gain of 1.77 dB, and the local weighted loss improves SSIM by 1.8%. This study provides new insights for high-precision signal processing in dynamic noise environments and has potential applications in radar target recognition and medical monitoring devices.
Keywords
signal denoising,generative adversarial networks,U-Net,Encoder-Decoder structure
Speaker
Comment submit