LSTM Framework for Classification of Radar and Communications Signals Victoria Clerico * (Student Member, IEEE), Jorge Gonz´ alez-L´ opez * , Gady Agam † , Jes´ us Grajal * (Senior Member, IEEE) * Information Processing and Telecommunications Center, Universidad Polit´ ecnica de Madrid. E.T.S.I. Telecomunicaci´ on, Av. Complutense 30, 28040 Madrid, Spain. Corresponding author: Victoria Clerico (e-mail: mclerico@ieee.org) † Department of Computer Science, IIT, Chicago, USA Abstract—Although radar and communications signal classifi- cation are usually treated separately, they share similar charac- teristics, and methods applied in one domain can be potentially applied in the other. We propose a simple and unified scheme for the classification of radar and communications signals using Long Short-Term Memory (LSTM) neural networks. This proposal provides an improvement of the state of the art on radar signals where LSTM models are starting to be applied within schemes of higher complexity. To date, there is no standard public dataset for radar signals. Therefore, we propose DeepRadar2022 1 ,a radar dataset used in our systematic evaluations that is available publicly and will facilitate a standard comparison between methods. Index Terms—Communications signals, radar signals, auto- matic modulation classifier, neural networks, long short-term memory networks. I. I NTRODUCTION Automatic modulation classification (AMC) consists in automatic determination of the modulation of a series of collected samples. It is the step that follows the detection of the signal and that is needed for data demodulation, therefore, it plays an important role in many civilian and military receivers [1]. Taking into account the classical approaches, AMC algo- rithms can be classified into two categories: those based on the likelihood function (LB, ‘Likelihood-Based’) and those based on feature extraction (FB, ‘Feature Based’) [2]. The first offers the optimal solution by minimizing the probability of false classification through the assumption that the probability density function (PDF) contains all the information needed for a specific waveform. Therefore, classification is performed by comparing the PDF likelihood ratio with a decision threshold [2]. The problem lies in its high computational complexity, which means that this method could not be suitable for real working environments. In contrast, FB algorithms extract rep- resentative features of each type of signal for their subsequent classification. These algorithms are suboptimal but are often preferred because they are easy to implement and suitable for real-time applications [2]. Despite that, FB classifiers rely This work was supported by project PID2020-113979RB-C21 founded by MCIN/AEI/10.13039/501100011033. 1 Available for download in https://www.kaggle.com/datasets/khilian/deepradar heavily on expert knowledge, so even though they are a good approximation on specific environments, they are highly complex and require a lot of time for development. While these algorithms have been successfully implemented to develop AMCs, Machine Learning (ML) and Deep Learning (DL) are considered good alternatives to develop high-performance and accurate AMCs without the need of time-consuming classical approaches. Algorithms such as K-Nearest Neighbors [3], Support Vector Machine [3], [4], Multilayer Perceptron (MLP) [5], Recurrent Neural Networks (RNN) [6], [7] and Convolutional Neural Networks (CNN) [8]–[11] have recently been used for this purpose. Previous literature shows the success of LSTM networks in processing and classifying sequences. Thus, our objective is to provide a robust and simplified AMC based on LSTM networks for both communications and radar signals. To do so, the public RadioML 2018.01A communications dataset [9] was used. Since there is no public radar signal dataset, we have created and published DeepRadar2022 1 , a radar dataset with continuous and pulsed signals of 23 classes. Furthermore, for comparison purposes, a dataset of eight types of radar signals was reproduced, which was proposed in the current state-of- the-art literature on radar signal classification [7]. The remainder of the paper is distributed as follows. The most relevant proposals and studies on AMCs with DL and ML are outlined in Section II. The signal model and datasets are introduced in Section III. Afterwards, our neural network architecture, metrics, and experimentation are described in Section IV, and the results are presented in Section V. Finally, the main conclusions of this work are presented in Section VI. II. RELATED WORK Despite the usefulness of the LB and FB classification algo- rithms, the appearance of artificial intelligence has revolution- ized many areas of interest. ML and DL tools have been used in the past to build modulation classifiers for communications and radar signals but have been treated separately. When classifying communications signals, the most typical signal representation is time series, and the most widely used method for its classification has been one-dimensional CNN arXiv:2305.03192v1 [eess.SP] 4 May 2023