Filtered Multicarrier Waveforms Classification: A Deep Learning-Based Approach

Abstract

Automatic signal recognition (ASR) plays an important role in various applications such as dynamic spectrum access and cognitive radio, hence it will be a key enabler for beyond 5G communications. Recently, many research works have been exploring deep learning (DL) based ASR, where it has been shown that simple convolutional neural networks (CNN) can outperform expert features based techniques. However, such works have been primarily focusing on single-carrier signals. With the advent of spectrally efficient filtered multicarrier waveforms, we propose in this paper, to revisit the DL based ASR to account for the variety and complexity of these new transmission schemes. Specifically, we design two types of classification algorithms. The first one relies on the cyclostationarity characteristics of the investigated waveforms combined with a support vector machine (SVM) classifier; while the second one explores the use of a four-layer CNN which performs both features extraction and classification. The proposed approaches do not require any a priori knowledge of the received signal parameters, and their performance is evaluated in a multipath channel through simulations for a signal-to-noise ratio (SNR) ranging from −8 to 20 dB. The simulation results show that, despite cyclostationary characteristics being highly discriminative, the CNN outperforms the cyclostationary based classification especially for short time received signals, and low SNR levels.