特征协助跟踪的多基地声呐网络结构化深度区分特征研究

To overcome the problems such as high clutter rate, and dim target with low detection rate sensors in multistatic sonar network, increasing attention has been paid to preprocessing technology for feature aided tracking, which suppresses the clutter by using the correlation between features derived from the return and the target, and provides more suitable representation. In this regard, this project proposes a structured deep discriminative feature learning paradigm for feature aided tracking, which takes together both horizontal combination and vertical mining of multiple sensors and features. First, we will establish a unified probability distribution representation for multistatic sonar features, and study their statistical similarity measure to fully measure the discrepancy between two distributions while allowing for some perturbation of each distribution. Then, based on the idea of structured sparsity, we will study the modeling and learning strategy for structured complementary features, reasonably fusing the multilevel complementary information of features of different sensors and types, and effectively suppressing the clutter. Moreover, based on the idea of unsupervised deep learning, we will also study the modeling and learning strategy for high order discriminative features, further capturing the abstract high order discriminative power in the network features, while improving the adaptability of learnt features to the deployed marine environment. Finally, the proposed paradigm will be applied to several typical trackers, so as to evaluate its effectiveness in target tracking. These studies will effectively suppress the clutter, while significantly improve the target measure quality and tracking performance, which is of great importance to safeguarding national marine sovereignty and upgrading the navy’s combat power.

为克服多基地声呐网络杂波率高、目标微弱、传感器检测概率低等问题，业界日益重视特征协助跟踪的预处理技术，以利用回波特征与目标的相关性，抑制杂波，提供更适宜的表征。对此，本项目提出特征协助跟踪的结构化深度区分特征学习范式，综合考虑多传感器、多特征间的横向融合与纵向挖掘。首先建立多基地声纳特征统一的概率分布表征，研究既能度量分布间的差异又能包容分布间扰动的统计相似性模型与学习策略；接着研究基于结构化稀疏的结构化互补特征的建模与学习，合理融合多传感器、多特征间的层次化互补信息，有效抑制杂波；此外，还将研究基于无监督深度学习的高阶区分特征的建模与学习，进一步捕获网络特征中抽象的高阶区分性，改善学得的特征对所部署的海洋环境的自适应性；最后，将其接入各种典型跟踪器，仿真评估其目标跟踪效用。通过上述研究，可有效抑制杂波，显著提高目标测量质量与跟踪性能，对维护国家海洋主权、提升海军战斗力意义重大。

本项目严格按照研究计划开展系统研究工作，即建立结构化深度区分特征学习范式，基础特征的概率分布表征与统计相似性度量研究，基于结构化稀疏的结构化互补特征研究，基于无监督深度学习的高阶区分特征研究。并在任务所规定的四个研究内容均取得了相关创新性成果，具体内容如下：①为了能够刻画声纳性能作为几何布局函数的粗略特性，仿真实现了多基地声纳数据生成模型。在此基础上深入研究了多种多基地声呐网络多级融合跟踪架构，为获得鲁棒而高性能的监控提供了丰富的技术手段。②在多基地声呐网络基础特征、目标与干扰特性深入研究的基础上，提出一种高效的时间维高阶间隙度特征用以刻画高背景杂波下的微弱目标。此外，将其扩展到多假设跟踪中，显著改善了目标跟踪与环境杂波抑制的效能。③为使学习器能够更好的发掘不同性质的特征与来自不同传感器的特征间潜在的一致性与多样性，并从中受益，我们提出将层次化聚类过程引入到标准的多核学习框架中，并将1-D的矢量融合推广到2-D的矩阵范数约束下的结构化融合。并系统分析了算法的Bousquet和Elisseeff稳定性边界。④研究了基于卷积神经网络的微弱信号检测和恢复问题。提出了一种由深卷积神经网络及其可视化组成对应的新框架。结合特定设计的多任务损失，可实现端到端的训练同时实现对微弱信号的检测和恢复。利用这一深层结构，低信噪比的性能极限可达到平均-24dB，某些情况下为-26dB。在对项目的深入执行过程中，我们系统梳理了目标微弱、环境背景杂波率高、传感器测量误差高度非线性等水声探测的三大挑战性难题，据此作为项目负责人申获国家级项目 5 项，总经费1200余万；作为第一/通信作者发表SCI论文7篇；作为第一/通信作者发表中文核心期刊论文4篇，作为第一完成人受理发明专利8项；培养博士研究生2名，硕士研究生9名；本项目中主持研发的5项技术在海军研究院与上海船舶电子设备研究所等行业研究所得到应用，获省部级奖励3项。成果突出，国防应用前景大。