Adaptive Compressed Spectrum Sensing Using Neural Networks in Cognitive Radio Networks

Abstract

In today's era of Internet of Things (IoT) applications, Cognitive Radio (CR) is recognized as the most promising technology to access the unused spectrum. Such CR networks can be easily built using Wide-band spectrum sensing. However, the processing of wide-band signals involves a high sampling rate. In such scenario, Compressed Spectrum Sensing (CSS) overcomes the challenges of real-time signal recovery and sampling. CSS uses Sparse signals which are widely used in many applications as they aid in processing large data. Beyond sparsity, all the real-world signals will have special structures (like Restricted Isometric property, Null space property). Moreover, it is very hard to determine the domain in which the signal is sparse. To find this we need a random measurement matrix which plays an important role in extracting the sparse coefficients of the signal. In this paper, a customized neural network is employed to identify the peculiar structures of sparse signals for efficient recovery in real time. The neural network learns and trains an adaptive measurement matrix from the sparse signals to reduce the sensing overhead. The transmitter and receiver systems are configured using Universal Software Radio Peripheral (USRP) boards with LabVIEW® and MATLAB® extensions for peer to peer communications. The implementation results depict the superior performance of neural network based recovery in assimilating the additional structures of real-time signals.