Advanced Electronic Surveillance DSP and ML Technique

电磁环境(EME)变得越来越拥挤和竞争。雷达和通信系统的设计者都在开发更复杂、越来越难以检测的方法。因此，电子监视(ES)面临着越来越复杂的运行环境。随着雷达和通信系统的适应性发展，对专家系统的要求也大大提高。最近，人们试图将机器学习的进展引入到对电磁环境的理解中。一些工作已将最大似然法应用于通信中的调制识别，而另一些工作则试图根据单个雷达发射机的射频发射(RF)对其进行分类。ES可以利用这些技术为操作员提供对他们周围复杂的EME的更多了解。针对低信噪比、拥挤和竞争环境下的实际应用问题，学术界研究甚少，最大似然估计在电磁环境中的应用还存在许多有待解决的研究问题。这些问题包括：-如何在有限的先验信息或标记能力的情况下收集训练数据。o标记总的来说是一项重大的挑战，而且是一项不容易完成的任务。这个PHD为这一领域的最佳实践开发了方法。-如何对RF数据进行预处理以从ML算法中获得最佳性能，例如，处理变化的功率电平的最佳归一化技术以及使用时频变换来提供更多信息。-如何处理特定的ES挑战，例如，只有几个信号/类别的例子可用于ML技术。o错误的数据标记和不完整的数据集如此拥挤的RF场景干扰和特定于ES问题的多路径问题。-如何跨多个分布式传感器聚合ML技术-对照传统方法的ML技术的基准并定义何时任一方法中断。-如何处理多种信号类型-一个模型或多个模型。尖端解决方案可能实现分层建模解决方案，可在此研究计划内进行研究。-将算法移植到硬件上时，现实世界存在哪些限制这些问题存在于ML可以解决的ES内的许多问题中。因此，PHD将专注于一个特定的研究领域；使用ML检测、计数和分离拥堵的EME中的信号。通过这项工作，博士生将能够利用以前在伦敦大学学院开发的时间-频率变换进行预处理，以及当前伦敦大学学院RFSoC的工作，以创建用于测试的数据集，并在其上部署算法。

The electromagnetic environment (EME) is becoming increasingly congested and contested. Designers of both radar and communications systems are developing methods that are both more complicated and increasingly harder to detect. Electronic Surveillance (ES) is therefore facing an increasingly complex environment to operate within. As radar and communications systems develop in their adaptability the requirements on ES systems increase significantly.Recently there have been attempts to bring the advances made in machine learning (ML) to the understanding of the EME. Several works have applied ML to modulation recognition in communications whilst others have attempted to classify individual radar transmitters from their radio frequency emissions (RF). ES could leverage these techniques to provide the operator with greater understanding of the complicated EME around them. Little academic research has been completed on real world applied problems considering low SNR, congested and contested environments.Many open research questions exist in the application of ML to the Electromagnetic Environment (EME). These include:- How to collect the training data with limited a priori information or ability to label it.o Labelling is a significant challenge in general and a non-trivial task to undertake. This PhD develop methods for best practice in this area.- How to pre-process RF data to get best performance from ML algorithms, e.g., best normalisation techniques to cope with varying power levels and the use of time-frequency transforms to provide more information.- How to deal with specific ES challenges such aso Fleeting signals, e.g., only a few examples of a signal/class are available to ML techniques.o Erroneous data labelling and incomplete datasetso Congested RF scenarioso Interference and multi-path problems specific to ES problems.- How to aggregate ML techniques across multiple distributed sensors- Benchmarking of ML techniques against traditional methods and defining when either methods break.- How to deal with multiple signal types - one model or multiple models. Cutting edge solutions may implement hierarchical modelling solutions which can be investigated within this research programme.- What real world limitations exist when porting algorithms onto hardwareThese questions are present across many of the problems within ES that ML could address. The PhD will therefore focus on a particular research area; detecting, counting and separating signals in a congested EME using ML. Through this the PhD student will be able to leverage pervious time-frequency transforms developed at UCL for pre-processing along with current UCL RFSoC work to create datasets for testing and to deploy algorithms on.