Convolutional Neural Networks with Feature Fusion Method for Automatic Modulation Classification

The analogy and application of Automatic modulation classification (AMC) detects the modulation type of received signals. Henceforth, the received signals can be correctly demodulated and, consequently, the transmitted message can be recovered. In Deep Learning (DL) based modulation classification, one major category, challenge is to pre-processing a received signal and representing it in a proper format-manner, before passing the desired-signal into the neural network. However, most existing modulation classification algorithms are neglecting the fact of mixing features between different representations, and the importance of features fusion method. This paper, however, attempted a Feature fusion scheme, for AMC, using convolutional neural networks (CNN). The approach was taken, in order to attempt fuse features extracted from In-phase & Quadrature (IQ) sequences, as well as, the Amplitude & Phase (AP) Sequences and Constellation Diagram images. Finally, simulation results show that fusing features from different representations can incorporate and leads to the best accuracy figures, achieved from each representation separately. Furthermore, our model achieved a classification accuracy of 84.68% at 0dB and 75.29% at -2dB and over 90% accuracy for high SNRs with a maximum accuracy of 94.65%, were available.