Automatic modulation classification is an essential and challenging topic in the development of cognitive radios, and it is the cornerstone of adaptive modulation and demodulation abilities to sense and learn surrounding environments and make corresponding decisions. In this paper, we propose a spectrum interference-based two-level data augmentation method in deep learning for automatic modulation classification. Since the frequency variation over time is the most important distinction between radio signals with various modulation schemes, we plan to expand samples by introducing different intensities of interference to the spectrum of radio signals. The original signal is first transformed into the frequency domain by using short-time Fourier transform, and the interference to the spectrum can be realized by bidirectional noise masks that satisfy the specific distribution. The augmented signals can be reconstructed through inverse Fourier transform based on the interfered spectrum, and then, the original and augmented signals are fed into the network. Finally, data augmentation at both training and testing stages can be used to improve the generalization performance of deep neural network. To the best of our knowledge, this is the first time that radio signals are augmented to help modulation classification by considering the frequency domain information. Moreover, we have proved that data augmentation at the test stage can be interpreted as model ensemble. By comparing with a variety of data augmentation techniques and state-of-the-art modulation classification methods on the public dataset RadioML 2016.10a, experimental results illustrate the effectiveness and advancement of proposed method.

Spectrum interference-based two-level data augmentation method in deep learning for automatic modulation classification

Automatic modulation recognition (AMR) plays a vital role in modern communication systems. This letter proposes a novel three-stream deep learning framework to extract the features from individual and combined in-phase/quadrature (I/Q) symbols of the modulated data. The proposed framework integrates one-dimensional (1D) convolutional, two-dimensional (2D) convolutional and long short-term memory (LSTM) layers to extract features more effectively from a time and space perspective. Experiments on the benchmark dataset show the proposed framework has efficient convergence speed and achieves improved recognition accuracy, especially for the signals modulated by higher dimensional schemes such as 16 quadrature amplitude modulation (16-QAM) and 64-QAM.

A Spatiotemporal Multi-Channel Learning Framework for Automatic Modulation Recognition

In this paper, we review a variety of deep learning algorithms and models for modulation recognition and classification of wireless communication signals. Specifically, deep learning (DL) has shown overwhelming advantages in computer vision, robotics, and voice recognition. Recently, DL has been proposed to apply to wireless communications for signal detection and classification in order to better learn the active users for electromagnetic spectrum sharing purposes. Therefore, we aim to provide a survey on the most recent techniques which use DL for recognizing and classifying a wireless signal. We focus on the most widely used DL models, emphasize the advantages and limitations, and discuss the challenges as well as future directions. In addition, we also apply a DL algorithm, convolutional neural network (CNN), to demonstrate the feasibility of using CNN to recognize and classify the over-the-air wireless signals using Mathworks DL toolbox with PlutoSDR and Universal Software Radio Peripheral (USRP), respectively.

Deep Learning for Modulation Recognition: A Survey With a Demonstration

Automatic determination of digital modulation types with different noises using Convolutional Neural Network based on time-frequency information

This paper proposes a convolutional neural network (CNN), called SCGNet, for low-complexity and robust modulation recognition in intelligent communication receivers. Principally, the network combines two types of sparse convolutional layers–depthwise and regular grouped in an architecture to achieve high recognition accuracy while keeping the network more lightweight. The network architecture leverages sparsely connected convolutional layers in three principal modules: speed-accuracy tradeoff (SAT), deep feature extraction and processing (DFEP), and generic feature extraction (GFE) data pre-processing module. For a good tradeoff between complexity and accuracy, SAT deploys depthwise convolutional layers to enrich the relevant features outputted by the former GFE module. In addition to SAT, DFEP employs a cascade of regular grouped convolutional layers for mining more discriminative features from SAT via a multilayer transformation module. This cascade structure aims to prevent a loss of essential details of the signal as the network becomes deeper. Additionally, skip connections are deployed between sub-blocks within SAT and DFEP to allow inter-module feature sharing and to handle inter-block features loss. Experimental results on the RadioML2018.01A dataset indicate that SCGNet achieves an overall recognition accuracy of around 94.39% at a signal-to-noise ratio of +20 dB.

Sparsely Connected CNN for Efficient Automatic Modulation Recognition

Recent convolutional neural networks (CNNs)-based image processing methods have proven that CNNs are good at extracting features of spatial data. In this letter, we present a CNN-based modulation recognition framework for the detection of radio signals in communication systems. Since the frequency variation with time is the most important distinction among radio signals with different modulation types, we transform 1-D radio signals into spectrogram images using the short-time discrete Fourier transform. Furthermore, we analyze statistical features of the radio signals and use a Gaussian filter to reduce noise. We compare the proposed CNN framework with two existing methods from literature in terms of recognition accuracy and computational complexity. The experiments show that the proposed CNN architecture with spectrogram images as signal representation achieves better recognition accuracy than existing deep learning-based methods.

Fei Han

Papers

Spectrum Analysis and Convolutional Neural Network for Automatic Modulation Recognition