A new Dataset of Wideband Radar Signals for Training Deep Neural Networks on Classification and Detection Tasks

Abstract

in the deep learning field, the availability of datasets is a very important requirement for developing deep neural network models and benchmarking. This paper introduces a new dataset of wideband radar signals (WBR-DS-1) that is essential for training, developing, and benchmarking deep neural network models for the classification and detection of wideband radar signals. Typical ESM receiver parameters, propagation channels, and environmental parameters are simulated to guarantee the dataset’s usability. The Electronic Support Measures (ESM) systems are responsible for the interception and characterization of the different radar signals. In this paper, the ESM sensor is assumed to be a ground-based one. Two scenarios are proposed to describe the geometric relationship between the ground-based ESM sensor and both the airborne and ground-based radar systems. The air-to-ground scenario is corresponding to airborne radars in front of the ground-based ESM sensor while the ground-to-ground scenario is corresponding to ground radars in front of ground-based ESM. One of the most important impairments that signals are subjected to during propagation is multipath fading. The multipath fading causes random variance in features and parameters of the radar signals. Both Rayleigh and Rician multipath channels with typical path losses and Doppler shifts are applied to simulate the environment. To verify that the dataset is suitable for training deep neural network models, a convolutional neural network (CNN) model has been trained and tested for classification of radar signals and detection of frequency modulated continuous wave (FMCW) radars.