Abstract
Next-generation wireless communication systems operate under increasingly dynamic channel conditions, where thermal noise, multipath fading, co-channel interference, hardware impairments, and mobility can significantly distort received signals. Conventional denoising techniques depend strongly on predefined statistical assumptions and fixed signal-processing parameters, which limits their adaptability across diverse wireless environments. Recent studies demonstrate that deep learning can learn complex signal characteristics directly from data and support intelligent physical-layer processing. Existing deep-learning denoising approaches may experience performance degradation when the signal-to-noise ratio, modulation condition, channel characteristics, or interference pattern changes from those represented in training data. This distribution-shift problem restricts their reliability in practical next-generation networks. This study proposes a novel Adaptive Residual Attention Denoising Network (ARAD-Net) for wireless signal restoration. ARAD-Net combines residual convolutional feature extraction, channel-attention weighting, and an SNR-aware adaptive gating mechanism to identify informative signal components while suppressing dynamically varying noise. The network jointly learns clean-signal reconstruction and noise-residual estimation, enabling adaptive denoising across heterogeneous channel conditions. Experimental evaluation across -10 to 20 dB SNR shows that ARAD-Net achieves MSE values from 0.132 to 0.007, RMSE values from 0.363 to 0.084, PSNR values from 20.78 to 32.36 dB, SNR improvements from 10.2 to 5.4 dB, and SSIM values from 0.861 to 0.977. Compared with the evaluated existing methods, ARAD-Net provides consistently lower reconstruction errors and higher signal-quality measures across all tested SNR conditions.
Authors
M. Senthil Vadivu
Sona College of Technology, India
Keywords
Deep Learning, Adaptive Signal Denoising, Wireless Communication, Residual Attention Network, Next-Generation Networks