Classification of Analog Modulated Signals Using Convolutional Neural Networks

In this paper, we propose two methods to classify analog modulated signals using convolutional neural networks (CNNs). The first method uses time domain inphase (I) and quadrature phase (Q) sequences as CNN inputs. In the second method, the CNN input is a spectrogram image obtained through fast Fourier transform of complex baseband I and Q signals. The neural network structures of both classification methods are almost identical except for the input image size. Eight modulation signals are considered for classification, including six amplitude modulation (AM) signals, frequency modulation signal and other digitally modulated signals. The six types of AM are double sideband with and without a carrier, upper sideband with and without a carrier, and lower sideband with and without a carrier. The classification accuracy of the two methods are obtained and compared based on sequence length and the number of filters of CNN. Simulation results show that the accuracy of both methods is over 99.9% for sequence lengths 256 to 1024, and the frequencydomain method is slightly better than the time-domain method.