Classification of Radio Signals in Multipath Fading Channel Using Neural Network

Knowledge of signal type is essential for cognitive radio systems. For this reason, a lot of papers have been published recently that explore different signal recognition algorithms. Most often, the papers consider signals passed through an AWGN channel. In this work, we use algorithms based on deep learning in the multipath Rayleigh fading channel. The channel parameters were selected in accordance with the ITU-R recommendation M.1225. We have considered some of the most popular architectures of neural networks that are used for modulation recognition in channel with AWGN, including convolutional and recurrent neural networks architectures. To solve the problem, a recurrent neural network was used with signal samples and higher order cumulants as inputs to achieve an accuracy of 0.98 at signal-to-noise ratio of 2 dB for BPSK, 0.79 for QPSK, 0.91 for 16QAM and 0.83 for 16PSK. At 6 dB SNR, accuracy greater than 0.97 was achieved for all modulation types.