A Deep Learning Model Based on Spectrogram Estimation and Wavelet Denoising for Spectrum Sensing in Cognitive Radio Systems

Cognitive Radio (CR) has materialized as a pivotal technology in wireless communications, addressing spectrum scarcity by enabling dynamic spectrum access. A fundamental component of CR at base transceiver stations is Spectrum Sensing (SS), which detects available frequencies within the licensed spectrum of Primary Users (PUs) to minimize channel interference. This study introduces an advanced SS model that integrates two complementary techniques: Convolutional Neural Networks (CNNs) and Wavelet Denoising (WD). The proposed WD-based CNN model (WD-CNN) is designed to enhance detection accuracy under low Signal-to-Noise Ratio (SNR) conditions. In the WD-CNN framework, an initial CNN model is employed as a feature extractor, leveraging spectrogram images of received signals to differentiate signal components from noise. The network is trained on diverse signal representations at low SNR to improve robustness. Unlike conventional CNN-based SS approaches, the proposed model further integrates WD with soft thresholding as a secondary denoising mechanism, refining the extracted features to accurately identify PU signals. Performance evaluation demonstrates that WD-CNN significantly outperforms both a CNN model (First CNN model) and transfer learning-based approaches, such as AlexNet. Specifically, at SNR levels of −10 dB and −20 dB, WD-CNN improves detection accuracy by 10.94% and 25.1%, respectively, compared to the First CNN Model. Furthermore, the proposed model surpasses AlexNet by 13.28% at −10 dB and 23.44% at −20 dB. Additionally, WD-CNN achieves a 30.06% accuracy improvement at −15 dB and a substantial 53.44% gain at −20 dB over conventional deep-learning-based SS techniques, while also exhibiting reduced computational complexity. Results demonstrate that the proposed WD-CNN model significantly enhances SS performance under challenging SNR conditions, providing a more robust and computationally efficient alternative to conventional CNN-based detection methods.