基于机器学习的海杂波特性感知和目标特征检测方法研究

The multi-features-based detector plays a more and more important role for detecting the weak slow-moving target on sea surface, along with the realization of radar’s refined observation. Two core procedures of multi-features-based detector are the accurate sensing of sea clutter characteristics and the design of the criterion of the multi-features-based detector. Nowadays, there are two bottlenecks for the performance improvement of multi-features-based detector. The accurate sensing of sea clutter characteristics confronts the problem of the limited identically distributed samples, and the design of multi-features-based detector confronts the problems about “class-imbalance” and the various types of targets for detected. For these problems, this project will combine the theories and methods of signal processing and machine learning, to develop the increment Bayesian sensing methods for the characteristics of sea clutter, to investigate feature extraction method based on the combination of multiple deep neural networks in the class-imbalance situation and to design the multi-features-based detection framework for the different types of targets based on anomaly detection theory and ensemble learning. Our researches can improve the detection performance of shore-based radar for weak slow-moving target. The achievements will play important roles in island defense, surface monitoring, maritime search and rescue, etc.

随着雷达精细化观测的实现，多特征检测器在海面慢速微弱小目标检测领域发挥着越来越重要的作用。多特征检测器的两个核心关键是海杂波特性的精确感知和多特征检测准则的设计。目前，制约多特征检测器性能进一步提升的两个瓶颈问题在于：海杂波特性精确感知阶段存在着海杂波中同分布样本不足的问题，而多特征检测设计阶段面临着类别非平衡以及待检测的海上目标类型多的困难。针对这些问题，本项目将结合信号处理和机器学习等领域的理论和方法，重点研究空时变海杂波特性增量Bayesian学习理论和方法，并实现类别非平衡情况下基于多种深度神经网络融合的特征提取，最终基于异常检测理论和集成学习实现多目标类型覆盖的多特征检测器设计。本课题的研究将提高岸基对海雷达的慢速微弱小目标的检测性能。研究成果可应用在岛礁防御、海面监视、海上搜救等方面。

复杂海杂波背景下的海面低速、漂浮小目标的检测是雷达目标检测领域中重要的研究课题之一。目前已经有一些新体制雷达系统可以实现对海面的高空间分辨-高多普勒分辨-多极化精细化观测，但与之相对应的基于精细化信号处理的海面目标检测面临着海面低速、漂浮小目标回波微弱，空时变海杂波异常复杂，海杂波特性认知难度大，小目标回波模型难以建立，海面小目标和海杂波之间样本的类别非平衡的问题，传统方法已经无法满足海面低速、漂浮小目标检测的实际应用需求，必须依靠新型高分辨多极化雷达系统进行海杂波特性的在线精细化感知和精细化利用。本项目着重解决了复杂海杂波背景下海杂波特性动态自主感知以及基于机器学习的目标多维特征联合检测的问题，形成了一系列的研究成果。本项目提出了岸基空时变海杂波参数分布估计的多帧递归参数估计框架，实现了快扫描模式大场景对海探测环境下海杂波的精细化动态估计。同时，在类别非平衡情况下提取了多种差异性海杂波与目标的差异性特征，设计了多目标类型覆盖的虚警可控的多种特征检测器，并在多组数据集上提升了检测性能。这些重要的研究成果已经成功应用于多个实验雷达或实际雷达系统中，对于空时变海杂波背景下的微弱小目标探测能力性能的提升提供了理论支持和算法支撑。