基于深度学习的CBRS环境感知能力传感器设计

for commercial usage, by the secondary users (SU)1, when it is notbeing used by the incumbent users (IU). Department of Defense(DoD) naval radars and Fixed Satellite Service (FSS) earth stationsare the incumbent users of the CBRS band. The onus of protectingthe IUs from interference from SUs is on the spectrum access sys-tem (SAS), the central component in the CBRS ecosystem. Sincethe FSS transceivers are static, incumbent protection for them canbe implemented using geolocation databases [10], as done for spec-trum sharing in the TV white space [21]. In contrast, incumbentprotection for DoD radars is much more challenging because theradar transceivers, mounted on naval ships, are mobile. The FCChas proposed deploying Environmental Sensing Capability (ESC)[19], with multiple ESC sensors placed in coastal areas [31] to detectnaval radar signals and report those to the SAS in a timely fashion.Based on the reported radar information from the ESC and theESC’s location, the SAS frees up the bands in which active CBRSdevices (CBSDs) might interfere with the radar signals. The SASalso attempts to relocate these CBSDs to other bands. Given thespectrum sharing opportunity enabled by CBRS, there is a lot ofinterest in building ESC systems.In this paper, we present DeepRadar, a novel deep-learning-basedsystem for detecting radar signals and estimating their spectraloccupancy. DeepRadar makes decisions in real-time, maintainingcontinuous operability by adapting its computations based on theavailable computing resources. We implement DeepRadar usinga commercial-off-the-shelf (COTS) software-defined radio (SDR).Our implementation meets all the ESC certification requirements[39]. DeepRadar also addresses several practical constraints thatare not part of the certification, but essential for deployability.Challenges in building DeepRadar: The design and imple-mentation of an ESC pose several challenges. First, the naval radartransmitter emits low duty cycle (ratio of pulse ON time to OFFtime) pulses, while being mobile and its antenna rotating 360 de-grees around its axis. Furthermore, the ESC cannot track/localizethe transmitter as storing information about the radar transmit-ter’s movement/position is not allowed [10]. Therefore, an ESC isilluminated only for a short duration by the narrow radar pulses,at unknown times. Hence, the ESCs must continuously monitorthe 3.5 GHz band and simultaneously analyze the captured signalswith a high temporal resolution. This is especially challenging asthe ESC must sample at a high RF sampling rate [24] due to the100 MHz wide bandwidth. The radar signals can appear anywhereon this band, and the ESC must detect them with 99% accuracy foradherence to the ESC certification criteria. Second, the ESC mustaccurately estimate the spectrum occupied by the radar signals,which is essential for incumbent protection and efficient spectrum560DeepRadar：基于深度学习的CBRS环境感知能力传感器设计0Shamik Sarkar 1 , Milind Buddhikot 2 , Aniqua Baset 1 , and Sneha Kumar Kasera 10Shamik.Sarkar@utah.edu,milind.buddhikot@nokia.com,aniqua@cs.utah.edu,kasera@cs.utah.edu 1 University of Utah 2 Nokia0摘要0我们提出了DeepRadar，一种基于深度学习的新型环境感知能力系统，用于检测雷达信号并估计其频谱占用。DeepRadar通过根据可用的计算资源调整计算来实时做出决策并保持连续可操作性。我们使用各种不同信干比（SIR）水平的测试数据对DeepRadar进行了全面评估。我们的评估结果显示，在20dB的峰值-平均SIR下，DeepRadar以99%的准确率检测到雷达信号，并且平均只丢失不到2MHz的频谱占用。我们使用商用软件定义无线电实现的DeepRadar也达到了同样高的检测准确率。0CCS概念0• 网络 → 认知无线电；网络实验；网络监测；传感器网络；•计算方法论 → 机器学习。0关键词0ESC，CBRS，雷达检测，带宽估计，深度学习，目标检测，频谱感知，嵌入式边缘计算0ACM参考格式：Shamik Sarkar 1 , Milind Buddhikot 2 , Aniqua Baset 1 ,and Sneha Kumar Kasera 1 . 2021.DeepRadar：基于深度学习的CBRS环境感知能力传感器设计。在第27届国际移动计算和网络会议（ACM MobiCom'21）上，2021年10月25日至29日，美国路易斯安那州新奥尔良。ACM，美国纽约，13页。https://doi.org/10.1145/3447993.344863201 引言0允许个人或课堂使用者制作本作品全部或部分的数字或硬拷贝，无需支付费用，但不得为了盈利或商业优势而制作或分发拷贝，并且拷贝必须带有本声明和第一页的完整引用。必须尊重ACM以外的其他组成部分的版权。允许带有署名的摘要。复制或重新发布，发布到服务器或重新分发到列表，需要事先获得特定的许可和/或费用。请向permissions@acm.org申请权限。ACM MobiCom '21，2021年10月25日至29日，美国路易斯安那州新奥尔良 ©2021年计算机协会。ACM ISBN 978-1-4503-8342-4/21/10...$15.00https://doi.org/10.1145/3447993.344863201CBRS有三类用户：（i）现有用户（ii）优先访问许可证和（iii）一般授权访问。我们将类别（ii）和（iii）的用户称为次级用户。570ACM MobiCom '21，2021年10月25日至29日，美国路易斯安那州新奥尔良，S. Sarkar，M. Buddhikot，A. Baset和S. Kasera0感知数据0从训练中学到的模型雷达检测决策0无雷达 无操作0使用跳过技术进行自适应频谱图创建的信息SAS0使用跳过技术进行自适应频谱图创建0确定跳过技术参数0在训练阶段探测ESC的计算资源 跳过参数0在SIL中学习目标检测模型0使用SIL进行目标检测0具有雷达和非雷达信号的数据0图1：DeepRadar的训练阶段和在线阶段概述0共享。在3.5 GHz频段中，有五种不同的雷达类型，带宽在1-100MHz范围内[39]。虽然不同的雷达类型具有不同的频谱特性，但是ESC必须在不知道现有雷达类型的情况下估计未知雷达的频谱占用情况。第三，ESC必须检测雷达信号并估计其频谱占用情况，不仅要在接收机噪声存在的情况下，还要在LTEeNodeB的干扰存在的情况下进行，这些干扰是在CBRS频段上机会性地运行的。最后，ESC必须区分带内雷达（3.55-3.65GHz）和带外（OoB）雷达信号。在3.55-3.65GHz频段之外可能存在高功率雷达信号，在3.5GHz频段中具有相当大的旁瓣功率。ESC不应将这些带外雷达检测为带内雷达，否则可能导致CBRS频谱的利用不足。DeepRadar概述：我们提出了一种称为光谱图像学习（SIL）的深度学习方法，用于解决在各种干扰场景中检测雷达信号以及估计其频谱占用的问题。基于能量检测和匹配滤波的简单方法无法处理未知的动态干扰。相反，深度学习方法可以学习不同的雷达特定特征，并在存在不可预测干扰的情况下进行检测。SIL基于“You Only LookOnce”（YOLO）算法[36]。图1描述了DeepRadar的训练和在线阶段。在训练阶段，SIL使用频谱图学习对象检测模型，基于雷达和非雷达数据。频谱图是二维图像，频率和时间分别沿图像的宽度和高度。我们发现，ESC认证要求在每兆赫的20dB峰均信干比（SIR）下的99％检测准确率对频谱图的宽度施加了约束（见第3.2节）。这个约束使得频谱图变窄而高，雷达信息仅在频谱图的几行中。我们建议在训练和在线阶段跳过计算一些非雷达行，如图1所示。跳过频谱图的某些行使DeepRadar能够稳健地适应ESC的可用计算资源，并确保在线阶段的及时预测。在训练过程中，我们还确定了LTE信号在ESC处的非确定性聚合如何影响SIL中的频谱图，并相应地训练SIL以抵御部署中ESC可能遇到的干扰。在在线阶段，我们首先根据感知数据形成频谱图。在形成频谱图时，我们根据训练阶段确定的参数跳过计算某些行。然后，我们将频谱图输入到SIL中，根据学习的模型进行预测。SIL非常快速，因为它通过单个卷积神经网络（CNN）一次通过输入频谱图同时检测雷达信号并估计其频谱占用。SIL和我们的跳过策略确保了DeepRadar的连续实时可操作性。0具有高时间和频谱分辨率。在检测到雷达信号时，如果其估计带宽的大部分超出了3.5GHz频段，我们将其视为超出频带的雷达；否则，视为频带内的雷达。为了能够在DeepRadar中进行这种区分，我们监测一个以3.6GHz为中心的更大的125MHz频段。我们使用不同SIR水平的各种测试数据对DeepRadar进行了全面评估。我们的评估结果显示，在20dB的峰值与平均SIR下，DeepRadar以每MHz的精度检测雷达信号，并且平均只错过不到2MHz的频谱占用。我们的评估还证明了DeepRadar相对于与3.5GHz频段雷达检测相关的几项工作的卓越性能。最后，我们提出了基于DeepRadar的嵌入式ESC实现。我们根据可用的计算资源分析了不同的COTS SDR，选择AIR-T[1]作为我们实现的SDR。我们证明了我们的实现满足ESC认证的性能要求，同时保持连续可操作性，即及时提供预测，无内存溢出，并跟上ESC的采样率。尽管一些现有的工作已经研究了CBRS的雷达检测[16, 17, 27, 40,44]，但没有一种方法能够准确估计雷达的带宽。DeepRadar相对于现有工作的重要创新和优势在于，它可以同时高效准确地执行雷达检测和带宽估计的两项任务。与现有方法相比，通过准确估计雷达带宽，DeepRadar提供了比10MHz块更精细的带宽分配可能性。总之，我们的工作具有以下贡献：•开发了一种新颖的深度学习方法，用于检测雷达信号并估计其带宽。•确定确保高雷达检测准确性所需的谱图宽度。•根据可用的计算资源调整DeepRadar中的计算。•开发一种用于处理ESC中未知干扰的数据增强方法。•使用ESC可能遇到的不同类型的测试数据对DeepRadar进行全面评估。•在COTSSDR上实现DeepRadar。我们构建了一个符合ESC认证标准的现实解决方案，可以灵活部署，并在不同的干扰场景下稳健工作。02. 3.5 GHz频段的雷达操作0在本节中，我们介绍雷达收发器的操作背景，并描述ESC认证标准。雷达操作：雷达发射机，具有方位狭窄、仰角广泛的天线辐射图案，发射高功率的窄脉冲。表1显示了五种不同雷达的信号特性3DESIGN OF DEEPRADARDeepRadar comprises of four constituent elements, shown in Fig-ure 3, that collectively address the challenges identified in Section 1.The input to DeepRadar are the in-phase (𝐼) and phase-quadrature(𝑄) values, obtained by sampling the captured RF signals. Deep-Radar divides these (𝐼,𝑄) values in different contiguous observationTable 1: Radar signal characteristicsRadarPulsePRRChirpPulsesBursttypewidth (𝜇s)(Hz)width (MHz)per burstlength (𝑚𝑠)10.5-2.5900-1100NA15 - 4013-44213-52300-3000NA5 - 201-6633-5300-300050-1008 - 242-80410-30300-30001-102 - 80.6-26550-100300-300050-1008 - 242-80�, � values from RF samplingPartition in windows of �� ����Divide �� in �slots of �� ��Use skip technique to select � out of � time slotsCompute PSD for chosen � slotsStack up PSDs to form image Feed image to CNNOoB vs. in-band radardecision����…Bandwidth estimationSelection of spectrogram dimensions Adaptive computing����Spectrogram Image LearningRadar bandwidth estimationFigure 3: Flow diagram of DeepRadarwindows, each of duration 𝑡𝑜, and makes a prediction for each ofthem. As noted in Section 1, our ESC receiver’s bandwidth is 125MHz. Hence, the sampling (quadrature sampling [15]) rate of ourESC is 𝑆 = 125 × 106 samples/sec, where each sample is an (𝐼,𝑄)tuple. We describe the four constituent elements of DeepRadar indetail in the following four subsections. We start with SIL, which isthe key element of DeepRadar.3.1Spectrogram image learning (SIL)In SIL, we use spectrograms, that contain both time and frequencydomain characteristics of the captured signals, for simultaneouslydetecting radar signals and estimating their bandwidth. These spec-trograms, used as features in SIL, are formed, as shown in Figure 3.First, we subdivide an observation window in 𝑁 time slots, each ofduration 𝑡𝑠, such that 𝑁 × 𝑡𝑠 = 𝑡𝑜. Next, we compute the PSD foreach of the time slots (using the (𝐼,𝑄) values of the correspondingtime slots), and stack up the PSDs vertically to produce a spectro-gram. Thus, a spectrogram is a matrix of size 𝑁 × 𝑀, whose rowsand columns correspond to different time slots and frequency bins,respectively. Here, 𝑀 = 𝑆 × 𝑡𝑠 is the number of (𝐼,𝑄) samples ina time slot. Accordingly, 𝑀 is also the number of frequency binsin a PSD. In this subsection, we assume that the skip is not used,i.e., PSDs are computed for all 𝑁 time slots. The effect of using skipis explained later in Section 3.3. Given that a spectrogram is animage whose pixel values represent signal power across differentfrequency bins and time slots, the task of spectrogram based radardetection can be considered an image classification problem. Notsurprisingly, some existing works [27, 40] have addressed the spec-trogram based radar detection problem using CNN. However, theseworks are inadequate for accurate radar bandwidth estimation.To address this inadequacy, we frame SIL as a regression problemsuch that the regressor can simultaneously detect the radar signalsand estimate their bandwidth. Using training data, we learn a model,𝑚𝑠𝑖𝑙 : 𝑅𝑁×𝑀 → {R(𝑓𝐿, 𝑓𝐻 ), ¯R}, where R and ¯R denote the pres-ence and absence of radar signals, respectively. When 𝑚𝑠𝑖𝑙 detects580DeepRadar：一种基于深度学习的CBRS环境感知能力传感器设计 ACM MobiCom '21，2021年10月25日至29日，美国路易斯安那州新奥尔良0(a) 3.563 GHz的1型雷达 (b) 3.554 GHz的2型雷达 (c) 3.590 GHz的3型雷达 (d) 3.566 GHz的4型雷达 (e) 3.596 GHz的5型雷达0图2：不同雷达类型的脉冲频谱特性。y轴上的刻度是相对的，而不是绝对的。0五种不同雷达的特性[39]如表1所示。表1的第二列列出了脉冲宽度的范围。在发送脉冲后立即，雷达收发器作为接收器监听被物体反射的脉冲。在1/PRR的时间间隔后，其中PRR是每秒脉冲重复率，雷达发射机发射下一个脉冲。表1的第三列显示了PRR的范围。对于1型和2型雷达，发送信号是脉冲调制的，脉冲宽度为1MHz，而对于3-5型雷达，发送信号是在脉冲内进行频率调制的。图2显示了不同雷达类型的一些样本脉冲的功率谱密度（PSD）[24]。表1的第四列列出了3-5型雷达的频率调制范围。雷达发射机必须用最少数量的脉冲（称为脉冲突发）击中物体，以高置信度地检测到该物体。表1的第五列指定了脉冲突发的数量范围。表1的最后一列显示了脉冲突发的持续时间，包括脉冲传输和监听间隔的持续时间。在发送完脉冲突发后，收发器水平旋转以覆盖所有方向。ESC认证标准：当雷达脉冲的峰值功率为-89dBm/MHz（或更高）且聚合干扰的平均功率为-109dBm/MHz（或更低）时，ESC必须以至少99%的准确率检测到所有五种雷达类型的脉冲突发。准确率是指检测雷达脉冲突发而不是单个脉冲的能力。在认证测试期间，一旦将脉冲突发输入ESC，ESC必须在5秒内检测到它，而不知道脉冲突发被输入的确切时间。雷达脉冲参数在雷达脉冲突发之间是不同的，从表1中随机选择，但对于同一脉冲突发中的所有脉冲是相同的。在测试期间，整个100MHz的CBRS频段上应用了代表聚合CBSD干扰的加性白高斯噪声（AWGN），这在[39]中得到了证明。ACM MobiCom ’21, October 25–29, 2021, New Orleans, LA, USAS. Sarkar, M. Buddhikot, A. Baset, and S. Kasera����FrequencyFigure 4: Object detection in SIL.a radar, it also produces a tuple, (𝑓𝐿, 𝑓𝐻 ), estimating the lower andhigher frequencies of the detected radar signal’s 3 dB bandwidth.We observe from Figure 4 that the estimated bandwidth of the radarsignal can be represented as a rectangle, that is completely definedby (𝑓𝐿, 𝑓𝐻 ). This observation motivates us to use object detection al-gorithms for solving the problem in SIL. In computer vision, objectdetection is the problem of learning a model that can detect differ-ent objects in an image, classify the detected objects, and estimatetheir locations in the image. If we consider all the bright spots in aspectrogram, produced by the radar pulses as shown in Figure 4, asa single object, which we call a radar object, then a rectangle, asin Figure 4, defines the radar object’s location in the image. Ourpurpose of radar detection and bandwidth estimation will be servedif we can detect and localize these radar objects. Thus, the learningproblem in SIL can be converted to an object detection problem, ifthe bright spots are visible on the spectrograms.We use YOLO [36] (and the idea of batch normalization fromYOLOv2 [37]) for object detection in SIL. YOLO detects objects fasterin comparison to other detection algorithms because it requiresonly one pass of the input image through a single neural network.Since our object detection task is simpler than the general objectdetection problem in computer vision, we develop a simpler versionof YOLO in SIL. Simplifying the object detection task reduces SIL’sprediction time further, which is an important consideration forthe ESC; discussed later in Section 3.3. We make the followingsimplifications: i) Our radar objects always have the same height asthe image. The radar objects’ height, which represents the durationof a pulse burst in an observation window, is not important asthe ESC only cares about the occurrence of pulse burst, not itsduration or the exact time of occurrence. ii) We do not anticipatethe occurrence of simultaneous radar signals with spectral overlap;an impractical scenario from the viewpoint of radar operation.Hence, we do not use anchor boxes, an idea introduced in YOLOv2,for localizing multiple overlapping objects of different shapes. iii)Furthermore, we do not expect multiple simultaneous radar signalsto be in spectral proximity. Hence, we use one bounding box percell (defined later). YOLO uses multiple bounding boxes per cell todetect multiple objects that are in proximity.Object detection in SIL: We train the CNN, shown in Figure 5,that produces an output ˆy of size 𝑛𝐶 × 5, as explained below. Wesubdivide every image in 𝑛𝐶 cells, each of width 𝑀/𝑛𝐶 and height𝑁. If a radar object is present in an image, we associate a singlecell, one that contains the radar signal’s center frequency, with thatobject. We use 𝑥 and 𝑤 to denote the location of the radar object’scenter, within the cell, and the radar object’s width, respectively.For each cell, we use confidence 𝑐 to specify the probability of anobject being present in that cell. Finally, for each cell, we use condi-tional probabilities 𝑝(R) and 𝑝( ¯R), to specify the probability of an590卷积块，F = 320卷积块，F = 640全连接层，units=128，activation=ReLu0卷积层，F个过滤器，大小为(3,3)，激活函数为ReLu0批量归一化0卷积块0全连接层，units=128，activation=ReLu0输入频谱图0��0展平0卷积块，F = 32 卷积块，0池化块0最大池化层，池化大小为(2,2)0Dropout，rate=0.250池化块0池化块 池化块0全连接层，units= �� × (3 +2)，activation=sigmoid0图5：SIL中使用的CNN架构。0给定存在目标的情况下，属于R和¯R的对象，因此ˆy包含长度为五的向量[ˆ��, ˆ��, ˆ��, ˆ��(R), ˆ��(¯R)]，对于每个单元格�∈[1, 2,...,��]，在一幅图像中。一旦����产生ˆy，我们计算一个置信度得分�� = ˆ�� ×ˆ��(R)，对于每个单元格，它指定了单元格中存在雷达目标的概率。如果任何单元格的��超过通过交叉验证[41]获得的预定义阈值，则我们宣布存在雷达，并估计其带宽为：0(��, ��) = �� × �−1 × �/�� + ˆ�� × �/�� ± ˆ��/2 + �00其中�� = �/�和�0 = 3537.5MHz分别是频率分辨率和PSD的起始频率。在训练过程中，我们使用一个�� × 5的矩阵y，其中包含[��,��,��,��(R),��(¯R)]，对于每个单元格�∈[1, 2,...,��]，作为训练图像的目标输出。对于没有雷达目标的单元格，��是单元格中心，�� = �/��，�� = 0，��(R) = 0，��(¯R) =1。使用训练数据和Adam优化器[26]，我们最小化以下损失函数来学习����：0L = ����0��0� = 101��0� ∈ {�,�} (�� − ˆ��)2 + ���0��0� = 101���(�� − ˆ��)20+ ����0��0� = 101����(�� − ˆ��0� = 1 1��0� ∈ 0��(�) − ˆ��(�)20其中O = {R, ¯R}。����，���和����是我们通过交叉验证确定的标量。1���如果单元格�中存在雷达目标，则为1，否则为0。1����是1���的补集。在学习过程中，L中的第一项和第四项分别惩罚具有雷达目标的单元格的目标定位误差和错误分类。第二项惩罚如果存在目标则置信度较低，第三项惩罚如果不存在目标则置信度较高，对于所有的单元格。我们通过交叉验证确定我们CNN的架构。我们使用Keras[18]开发SIL。图5中的层名称遵循Keras的约定。03.2 选择频谱图的尺寸0在SIL中，我们将观测窗口划分为持续时间为��的时间槽。我们必须仔细选择��，因为SIL的准确性取决于它。从图4中我们可以观察到，当雷达脉冲的频谱分量的功率高于PSD中的其他频谱分量时，我们在频谱图中得到一个亮点。SIL将频谱图中所有的亮点集合检测为一个雷达目标。因此，SIL的成功取决于亮点相对于频谱图背景的亮度。然而，亮点的亮度取决于��，如下所述。ESC认证标准在第2节中规定，雷达脉冲必须以每兆赫兹20dB峰值-平均信噪比进行检测。DeepRadar: A Deep-Learning-based Environmental Sensing Capability Sensor Design f