DFENet: A diverse feature extraction neural network for improving automatic modulation classification accuracy in wireless communication systems.

Abstract

In this paper, a convolutional neural network (CNN) model, named DFENet, is composed of multi-branch blocks for diverse feature extraction (DFE) to improve the accuracy of automatic modulation classification (AMC) in wireless communication systems. The DFE blocks primarily involve advanced processing sub-blocks designed to extract signal features from IQ (In-phase and Quadrature-phase) data using filters at multiple scales. By extracting diverse intrinsic features, the convolutional layers with different filter sizes can prevent overfitting and gradient vanishing problems, thus enhancing the AMC accuracy while maintaining reasonable computational complexity. Experimental results with the HisarMod2019 dataset demonstrate that the DFENet model with 4 DFE blocks and 128 filters in the convolutional layers achieved an average AMC accuracy of 82.76%, excelling particularly at low SNR levels, with an accuracy exceeding 60% at SNR = -20 dB and greater than 90% at SNR = -2 dB. For the RadioML2018.01A dataset, the proposed model yielded an accuracy of AMC greater than 93% for SNR >6 dB. In comparison, our model outperforms other state-of-the-art models in terms of accuracy while maintaining reasonable computational complexity and fast execution time.