Wireless Channel Recognition by CNN

Abstract

This paper proposes a wireless channel recognition algorithm based on convolutional neural network (CNN) combined with multi-domain feature extraction. Channel State Information (CSI) of signals usually involves the information of the environment in which the receiver is located. In this paper, the Digital Terrestrial Multimedia Broadcast (DTMB) signal is used to extract the CSI features. The cross-correlation between PN420 sequences in frame header and the local PN sequences in the receiver is employed as the time-domain features. Moreover, the time-domain features are transferred into time-frequency images by Wigner-Ville Distribution (WVD). Then, CNN is adopted to recognize the wireless channels. The performances of the combined WVD time-frequency domain and time-domain feature extraction are compared with the WVD time-frequency domain feature extraction. Simulation results show that the proposed channel recognition method based on time-frequency domain has good performance; and its accuracy can reach more than 96% when SNR is −2dB; the performances of the jointed multi-domain feature extraction outperform those of the time-frequency domain feature extraction method. The accuracy can reach more than 99% when SNR is −10dB.