随着现代信息科技的迅速发展，多维度的电子战逐渐成为现代战争的主流。目前，利用陆基、空基和天基等多种装备子系统构成的一体化监测系统，以其覆盖范围广等优点，获得越来越广泛的应用。在现代电子战中，保持一体化系统对目标的定位跟踪能力是至关重要的。目前仅利用目标辐射源的无源被动定位技术主要包括：测向定位，到达频差定位，到达时间和到达时差定位等，其中到达时差（Time Difference of Arrival, TDOA）由于仅需测量时间差，系统结构简单，被广泛用于被动定位中。然而由于目标的高动态化以及地形遮挡等因素，单个装备子系统内的不同监测站点会存在对目标的非共视问题，进而导致该测量时刻的TDOA测量矩阵不满秩，无法对该时刻的目标进行定位。

本文以雷达辐射源目标为处理对象，针对非共视时刻无法定位的问题，研究如何基于深度学习利用脉冲描述字（Pulse Descriptive Word, PDW）串进行雷达同源匹配以提升非共视时刻定位参数的利用率并保证一体化系统对目标的定位能力。基于PDW同源匹配算法可实现以下两种功能：一是将不同装备子系统在相近时刻测量得到并已分选完成的PDW数据串进行同源匹配，一旦确定为同源，即可对不同装备子系统的TDOA参数加以融合利用进行定位；同时也可对同一装备子系统不同时刻的PDW数据串进行同源匹配，以此实现在多监测目标情况下对目标进行编批。由于各个装备子系统与目标的距离各不相同以及同一装备子系统内的同源编批需求，这就极有可能导致两组用来匹配的PDW数据串是异步的，即两组参数不在同一时差窗内。另一方面，由于雷达脉冲分选算法性能不足以及系统监测条件高动态变化等原因，导致PDW数据存在丢失或野值。由于传统匹配方法需要脉冲对齐，因此对上述情况下的PDW数据匹配准确率较低。如何通过技术手段打破上述非共视情况下监测数据的融合壁垒，该问题亟待解决。

本文的主要研究工作分为以下几个方面：

（1）针对传统方法难以保证异步数据精确匹配难题，本文提出了基于自注意力机制的PDW同源匹配算法，将传统数据分类方法与自注意力机制相结合。首先通过传统数据分类方法对载频、脉宽等重要匹配参数信息进行分类，主要分为以下三类：无捷变类型、周期性捷变类型与非周期性捷变类型。将数据类型不一致的直接判定为非同源类型，无捷变类型参数则直接通过参数信息的欧式距离进行对比判断。然后针对周期性捷变数据与非周期性捷变数据训练不同的自注意力机制网络，网络更具针对性，大大提升了网络匹配的准确性，并引入残差单元以提升网络匹配准确率。算法整体研究流程主要分为以下几个部分：数据分类、数据集制作、网络训练与数据输入算法进行匹配等。最后对算法进行性能分析，并与传统聚类算法和其他网络结构下的匹配效果进行对比分析。经仿真分析验证，在参数测量误差较小时，基于自注意力机制的PDW同源匹配算法的匹配准确率在95%及以上，并且随着参数测量误差的增大，该算法的匹配准确率始终维持在90%以上。

（2）针对所提基于自注意力机制的PDW同源匹配算法面向错漏数据匹配准确率过低问题，本文提出了面向非理想PDW数据的深度学习同源匹配算法。在实际情况中，由于雷达脉冲分选算法性能不足以及系统监测条件高动态变化等原因，用于匹配的PDW数据经常发生存在丢失或野值的“错漏”现象，导致基于自注意力机制的PDW同源匹配算法匹配准确率骤降。原因在于上述算法在处理错漏数据时，使用传统数据分类方法错误率较高，网络结构也较为简单，无法对复杂情况下的数据进行准确匹配。本算法基于自注意力机制，直接使用网络模型进行匹配，在所提基于自注意力机制的PDW同源匹配算法的网络模型基础上重构网络结构，应用数据掩码等各类数据预处理技巧，并引入Embedding以提高对错漏数据的匹配准确率。算法研究流程分为以下三部分：数据集构建、网络训练与数据输入网络进行匹配。最后对算法进行性能分析，在不同数据丢失率、野值数据比例以及错漏数据比例下对算法进行测试，并与基于自注意力机制的PDW同源匹配算法进行性能对比分析。在错漏数据比例不超过25%时，面向非理想PDW数据的深度学习同源匹配算法的匹配准确率能够维持在95%以上。

With the rapid development of modern information technology, multi-dimensional electronic warfare has gradually become the mainstream of modern warfare. At present, the integrated monitoring system composed of various equipment subsystems such as land-based and space-based has been widely used due to its advantages such as wide coverage. In modern electronic warfare, it is crucial to maintain the ability of the integrated system to locate and track the target. At present, the passive positioning technology that only uses the target radiation source mainly includes: direction finding positioning, frequency difference of arrival positioning, time of arrival and time difference of arrival positioning, etc., where time difference of arrival (Time Difference of Arrival, TDOA) only needs to measure the time difference, the system has a simple structure and is widely used in passive positioning. However, due to factors such as high dynamics of the target and terrain occlusion, different monitoring stations in a single equipment subsystem will have the problem of non-common view of the target, which will lead to the TDOA measurement matrix at this measurement moment is not full of rank, and the target at this moment cannot be located.

This paper takes the radar radiation source target as the processing object, and aims at the problem that the target position cannot be obtained at the time of non-common view, and studies how to use Pulse Descriptive Word (PDW) strings for radar homology matching based on deep learning to improve utilization of positioning parameters at the time of non-common view and ensuring the positioning ability of the integrated system to the target. Based on the homology matching algorithm, the following two functions can be realized: one is to carry out homologous matching on the pulse descriptor parameters measured and sorted by different equipment subsystems at similar times. Once determined to be homologous, the TDOA parameters of different equipment subsystems can be fused and utilized; at the same time, the PDW parameters of the same equipment subsystem at different times can also be homologously matched, so as to realize the batching of targets in the case of multiple monitoring targets. Due to the different distances between each equipment subsystem and the target and the same-source batching requirements in the same equipment subsystem, it is very likely that the two sets of PDW data strings used for matching are asynchronous, that is, the two sets of parameters are not in the same time difference window. On the other hand, due to the insufficient performance of the radar pulse sorting algorithm and the high dynamic changes of the system monitoring conditions, there will be loss or outliers in the PDW data. Since the traditional matching method requires pulse alignment, the matching accuracy of PDW data in the above cases is low. How to use technical means to break the above-mentioned barriers to the fusion of monitoring data in the case of non-common view, this problem needs to be solved urgently.

The main research work of this paper is divided into the following aspects:

(1) Aiming at the problem that traditional methods are difficult to ensure accurate matching of asynchronous data, this paper proposes a PDW homology matching algorithm based on self-attention mechanism, which combines traditional data classification methods with self-attention mechanism. Firstly, the important matching parameter information such as carrier frequency and pulse width are classified by the traditional data classification method, which are mainly divided into the following three categories: non-agile type, periodic agile type, and non-periodic agile type. Inconsistent data types are directly judged as non-homologous types, and non-agile data is directly compared and judged by the Euclidean distance of parameter information. Then, different self-attention mechanism networks are trained for periodic agile data and non-periodic agile data. The network is more targeted, which greatly improves the accuracy of network matching, and introduces residual units to improve network matching accuracy. The overall research process of the algorithm is mainly divided into the following parts: data classification, data set production, network training and data input algorithm matching, etc. Finally, the performance analysis of the algorithm is carried out, and the matching performance is compared with the traditional clustering algorithm and other network structures. It is verified by simulation analysis that when the parameter measurement error is small, the matching accuracy of the PDW homologous matching algorithm based on the self-attention mechanism is 95% or above, and with the increase of the parameter measurement error, the matching accuracy of the algorithm is always maintained above 90%.

(2) Aiming at the low matching accuracy of the proposed algorithm for error-missing data, this paper proposes a deep learning homology matching algorithm for non-ideal PDW data. In actual situations, due to the lack of performance of the radar pulse sorting algorithm and the high dynamic changes of the system monitoring conditions, the PDW data used for matching often occurs missing or there are outliers. The accuracy of homology matching algorithm based on self-attention mechanism drops sharply. The reason is that the traditional data classification method used by the above algorithm has a high error rate and a relatively simple network structure when dealing with wrong and missing data, and cannot accurately match data in complex situations. This algorithm is based on the self-attention mechanism, directly uses the network model for matching, reconstructs the network structure based on the network model of the proposed PDW homology matching algorithm based on the self-attention mechanism, and applies various data preprocessing techniques such as data mask, and introduce Embedding to improve the matching accuracy of wrong and missing data. The algorithm research process is divided into the following three parts: data set construction, network training, and data input network matching. Finally, the performance analysis of the algorithm is carried out, and the algorithm is tested under different data loss rates and outlier data ratios, and the performance is compared with the PDW homology matching algorithm based on the self-attention mechanism. When the proportion of wrong and missing data does not exceed 25%, the matching accuracy of the deep learning homology matching algorithm for non-ideal PDW data can be maintained above 95%.