Intelligent Spectrum Sensing of Consumer IoT Based on GAN-GRU-YOLO

• Published in: IEEE transactions on consumer electronics Vol. 70; no. 3; pp. 6140 - 6148
• Main Authors: Gao, Zhihe, Li, Yufang, Chen, Zhe, Asif, Muhammad, Xu, Lingwei, Li, Xingwang, Gulliver, T. Aaron
• Format: Journal Article
• Language: English
• Published: IEEE 01.08.2024
• Subjects: Consumer electronics; Consumer Internet of Things; Continuous wavelet transforms; CWT; Feature extraction; Frequency-domain analysis; GAN; GRU; Internet of Things; Sensors; spectrum sensing; YOLO; YOLOv5
• ISSN: 0098-3063; 1558-4127
• DOI: 10.1109/TCE.2024.3418103

In the swift evolution of 5G cellular communication technology and Internet of Things (IoT), the consumer electronics market is booming. Consumer IoT has become an emerging industry. However, the development of the consumer IoT is subject to limited spectrum resources. Hence, this study suggests a smart spectrum sensing approach for consumer IoT based on GAN-GRU-YOLO. First, a Continuous Wavelet Transform (CWT) is used to capture frequency domain information from the received signals. A frequency domain feature matrix is constructed and then converted to a signal spectrogram to improve data diversity and enhance sensing. GAN is used to learn the signal spectrogram to generate more realistic synthetic data to achieve data enhancement and improve the classification performance of the overall model. Then, a two-branch GRU-YOLO network is employed to learn the signal characteristics in the time and frequency domains. The upper branch captures local feature information in the frequency domain and the YOLOv5 network captures high-level features. A combination of GRU and CNN in the lower branch extracts features from the data time series to ensure information continuity. Finally, the branch outputs are fused for further processing. The GAN-GRU-YOLO network has high generalization ability and efficiency. Compared with other methods, the proposed approach has a lower false alarm probability <inline-formula> <tex-math notation="LaTeX">(P_{f}) </tex-math></inline-formula> and a higher detection probability <inline-formula> <tex-math notation="LaTeX">(P_{d}) </tex-math></inline-formula>. At a signal-to-noise (SNR) ratio of -15 dB, the <inline-formula> <tex-math notation="LaTeX">P_{d} </tex-math></inline-formula> is 11% to 65% higher and the <inline-formula> <tex-math notation="LaTeX">P_{f} </tex-math></inline-formula> is 25% to 61% lower than the ResNet, MobileNet, Transformer and YOLOv6 algorithms.