CBRS band experimental waveform dataset for testing radar detection machine learning models

Author Info: Alex Lackpour, [e-mail]<br> Affiliation: Drexel University<br> Drexel Wireless Systems Lab website: https://research.coe.drexel.edu/ece/dwsl/<br> Date: April 5th, 2021<br> Version: 1.0, public release This radar RF waveform dataset (Group6_data) was collected for a tutorial paper that was originally completed in January 2021. The data contained herein was used to demonstrate that the detection accuracy of a spectrogram-based Convolutional Neural Network (CNN) radar detector model is not negatively impacted by the RF hardware impairments experienced due to sending and receiving the data with USRP N210 software defined radios. The dataset contains 900 examples of CBRS band radar activity and 900 examples of simulated random noise. The IQ data samples were sent/received at 10MSps and are each 80ms in duration. The 900 examples containing radar activity have a random SNR that varies between 10 and 20 dB in 2dB steps. All 9 of the original (simulated) MATLAB workspaces were generated using the NIST Simulated Radar Waveform Generator that is available here for download:<br> https://github.com/usnistgov/SimulatedRadarWaveformGenerator A MATLAB script (included in a separate tar archive file in this dataset) is used to send, or transmit, each of the 9 original simulated waveform batches stored in each MATLAB workspace using a USRP radio at 1.5 MHz over a 2 meter long RF coaxial cable and a 30 dB RF attenuator. A separate receiving MATLAB script (included) is used to receive the entire sent dataset and store it in an "experimental" dataset on the receiving computer. With these two versions of the radar waveform samples, it is possible to compare the accuracy of a radar detector for both the source and experimental datasets and determine whether the RF hardware impairments imparted by the USRP radio hardware degrades the detector's binary classification accuracy. The following code repository repo contains a set of baseline deep learning radar waveform detection models that were evaluated and documented in our tutorial paper:<br> https://github.com/usnistgov/BaselineDeepLearningRadarDetectors作者信息：亚历克斯·拉克普尔（Alex Lackpour），电子邮箱：[e-mail]<br>隶属机构：德雷塞尔大学（Drexel University）<br>德雷塞尔无线系统实验室官网：https://research.coe.drexel.edu/ece/dwsl/<br>发布日期：2021年4月5日<br>版本：1.0，公开版本<br>本雷达射频（Radio Frequency, RF）波形数据集（Group6_data）是为一篇于2021年1月初步完成的教程论文采集构建的。本数据集所载数据用于验证：基于频谱图的卷积神经网络（Convolutional Neural Network, CNN）雷达检测模型的检测精度，不会因使用USRP N210软件定义无线电收发数据所产生的射频硬件损伤而出现劣化。<br>本数据集包含900条CBRS频段（Citizens Broadband Radio Service）雷达活动样本，以及900条模拟随机噪声样本。同相正交（In-phase and Quadrature, IQ）数据采样率为10MSps，单条样本时长均为80ms。其中900条雷达活动样本的信噪比（Signal-to-Noise Ratio, SNR）随机取值于10dB至20dB之间，步长为2dB。<br>全部9份初始（模拟）MATLAB工作区均通过美国国家标准与技术研究院（National Institute of Standards and Technology, NIST）的模拟雷达波形生成器生成，该生成器可从以下地址下载：https://github.com/usnistgov/SimulatedRadarWaveformGenerator<br>本数据集附带一份MATLAB脚本（存放于独立的tar归档文件中），用于通过USRP无线电设备，以1.5MHz的带宽，经由2米长射频同轴电缆与30dB射频衰减器，发送每份MATLAB工作区中存储的9组初始模拟波形批次。另有一份配套接收MATLAB脚本（随附于数据集内），用于接收全部发送的数据集，并将其存储于接收计算机的“实验性”数据集当中。<br>通过上述两类雷达波形样本，可对比源数据集与实验数据集的雷达检测器检测精度，进而判断USRP无线电硬件引入的射频损伤是否会降低检测器的二分类精度。<br>本教程论文中评估并记录了一套基线深度学习雷达波形检测模型，相关代码仓库可访问：https://github.com/usnistgov/BaselineDeepLearningRadarDetectors