MLWiNS: Quality Versus Quantity in Spectrum Sensing with Distributed Sensors

显示无线频谱使用情况的精确地图可用于通知频谱政策，确定使用模式和态势感知，并评估室内和室外机会性频谱接入方法的可行性，特别是在拥挤的6 GHz以下频谱中。然而，创建这样一个地图的成功取决于测量频谱活动的传感器的质量和数量，以及它们的测量是如何由中央服务器处理的。通常，由于传感器工作在未知或时变的环境中，通常需要许多分布式传感器来补偿这种不确定性。因此，传感器的成本、功耗和尺寸必须非常低，才能进行大规模部署。该项目研究了大规模分布式测量系统的传感器质量与数量权衡，该系统可以推断整个区域的无线活动。由于缺乏真正大规模的精确实时测量，对频谱活动的了解受到阻碍，并且部署大量传感器受到成本、功率和复杂性的阻碍。该研究解决了如何将硬件商品化并将成本和功耗降低一个数量级的重要方面，从而有可能为大型频谱数据集铺平道路。这项工作将用于在圣母大学现有的本科生研究经验（REU）网站上播种项目。研究结果将向研究界发布，以鼓励讨论如何以低成本和可扩展的方式可靠地测量频谱占用。频谱传感器的核心是感兴趣频段的功率检测器。传统上，这样的传感器被评估为无线电，因为它通常对其输入进行调谐和下变频到中频或基带信号，数字化，滤波，然后分析功率谱。无线电的标准性能指标包括灵敏度、线性度、带宽、抗像性、频率稳定性、相位噪声和动态范围。满足这些规范的严格要求会直接降低成本和功耗。然而，频谱映射系统的性能也取决于集中式服务器从广泛分布的一组传感器推断整个地区活动的能力。本研究的重点是放宽许多传感器的要求，并分析当传感器数量增加以补偿每个传感器能力降低时，使用机器学习的实时频谱映射系统的性能。这项工作旨在表明，数量比质量更有利——在大规模部署中，从额外的传感器中学到的信息很容易补偿它们减少的单个能力，特别是在考虑成本和功耗的情况下。还考虑了计算和网络负载要求以及传感器尺寸的影响。该研究成果是一种创新的全系统方法，用于确定分布式频谱测量的质量-数量轮廓。这项工作将：(i)为具有分布式传感器的机器学习无线电频谱传感系统的性能建立模型、指标和保真度标准；（ii）研究大幅降低传感器成本和功耗的机会；（iii）建立广泛的传感器测试和测量平台，以验证假设、分析、方法和结论。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

An accurate map showing usage of the wireless spectrum can be used to inform spectrum policy, determine usage patterns and situational awareness, and evaluate the feasibility of opportunistic spectrum access methods indoors and outdoors, especially in the crowded sub-6 GHz spectrum. However, the success of creating such a map depends on the quality and quantity of sensors measuring spectrum activity, and how their measurements are processed by a centralized server. In general, because the sensors operate in an unknown or time-varying environment, many distributed sensors are generally needed to compensate for this uncertainty. As a result, the cost, power consumption, and size of the sensors need to be very low to allow wide-scale deployment. This project examines the sensor quality versus quantity tradeoff for a large-scale distributed measurement system that infers wireless activity throughout a region. Knowledge of the activity of the spectrum is hampered by the lack of accurate real-time measurements on a truly large scale and deploying a large number of sensors is hampered by cost, power, and complexity. The research addresses the important aspect of how to commoditize the hardware and drive the cost and power down by an order of magnitude, thus potentially paving the way to large spectrum data sets. The work will be used to seed projects at an existing Research Experience for Undergraduates (REU) site at the University of Notre Dame. The results will be disseminated to the research community to encourage discussions about how spectral occupancy can be measured reliably in a low-cost and scalable manner.A spectrum sensor is, at its heart, a power detector for frequency bands of interest. Traditionally, such a sensor is evaluated as a radio since it generally tunes and down-converts its input to an intermediate-frequency or baseband signal, digitizes, filters, and then analyzes the power spectrum. The standard performance specifications of a radio include sensitivity, linearity, bandwidth, image-rejection, frequency stability, phase noise, and dynamic range. Meeting stringent requirements on these specifications contributes directly to cost and power consumption. However, spectrum mapping system performance is also determined by the ability of the centralized server to infer activities over a whole region from a widely distributed set of sensors. This research focuses on relaxing many sensor requirements, and analyzing how a real-time spectrum mapping system that uses machine-learning performs when the number of sensors is increased to compensate for the reduced capability per sensor. The effort seeks to show that there is an advantageous trade-off of quantity over quality---the information learned from additional sensors in a large-scale deployment easily compensates for their reduced individual capabilities, especially when accounting for cost and power consumption. Implications for computational and network load requirements and sensor size are also considered. This research effort is an innovative whole-system approach to determining the quality-quantity contour for distributed spectrum measurements. The effort will: (i) establish models, metrics, and fidelity criteria for the performance of a machine-learning radio-frequency spectrum sensing system with distributed sensors; (ii) examine opportunities to dramatically reduce sensor cost and power consumption; (iii) build an extensive test and measurement platform of sensors to validate the assumptions, analyses, methods and conclusions.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.