Deep-Learning-Based Classifier With Custom Feature-Extraction Layers for Digitally Modulated Signals

Abstract

The paper presents a novel deep-learning (DL) based classifier for digitally modulated signals that uses a capsule network (CAP) with custom-designed feature extraction layers. The classifier takes the in-phase/quadrature (I/Q) components of the digitally modulated signal as input, and the feature extraction layers are inspired by cyclostationary signal processing (CSP) techniques, which extract the cyclic cumulant (CC) features that are employed by conventional CSP-based approaches to blind modulation classification and signal identification. Specifically, the feature extraction layers implement a proxy of the mathematical functions used in the calculation of the CC features and include a squaring layer, a raise-to-the-power-of-three layer, and a fast-Fourier-transform (FFT) layer, along with additional normalization and warping layers to ensure that the relative signal powers are retained and to prevent the trainable neural network (NN) layers from diverging in the training process. The classification performance and the generalization abilities of the proposed CAP are tested using two distinct datasets that contain similar classes of digitally modulated signals but that have been generated independently, and numerical results obtained reveal that the proposed CAP with novel feature extraction layers achieves high classification accuracy while also outperforming alternative DL-based approaches for signal classification in terms of both classification accuracy and generalization abilities.