标签共享子空间多源迁移学习方法及在雷达辐射源识别中的研究

项目名称： 标签共享子空间多源迁移学习方法及在雷达辐射源识别中的研究

项目编号： No.61472305

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 其他

项目作者： 方敏

作者单位： 西安电子科技大学

项目金额： 80万元

中文摘要： 以多观测站新型雷达辐射源分类识别为背景，探索一种新的基于多源标签共享子空间迁移学习方法。分析多源域预测标签之间的相关性，研究目标域样本多标签化处理方法，扩展目标域样本为多标签样本。研究一种新的基于多标签共享子空间的多源迁移学习算法，避免独立评测每个源域雷达辐射源可迁移性带来的偏差。对无监督迁移成分析域匹配方法进行归纳式扩展，研究一种归纳式迁移成分分析的域匹配方法。在此基础上建立源节点预测器模型，提升迁移成分析的域匹配方法中少量带标签目标域样本的指导作用，为源域中雷达辐射源数据的映射和筛选提供理论基础。带有标签噪声的源域样本对可迁移性影响较大，结合自适应聚类算法研究具有区域筛选功能的迁移成分分析域匹配方法。选取局域范围内标签一致的样本，抑制受噪声影响的源域样本的迁移，使得归纳式迁移成分分析具有抗噪能力。结合多源共享子空间迁移方法，设计一种适用于雷达辐射源新体制、型号识别的迁移学习方法。

中文关键词： 标签共享子空间；迁移学习；多源；迁移成分分析；雷达辐射源

英文摘要： This project concerns on problems of recognition of novel radar emitter under multiple observation stations. It proposes a new transfer method based on the shared subspace of multiple sources and studies the classification of radar emitters by this method. This method studies the correlations between predicted labels of multiple source domains, explores the way to transform examples in the target domain into ones in the multi-label setting. The proposed method is a new multi-label shared subspace for multiple source domain transfer learning, to avoid the bias that individually assessing the transferability of radar emitter in each source domain causes. This project inductively expands the unsupervised transfer component analysis domain adaptation and studies an inductive transfer component analysis domain adaptation method, based on which source node predictive models are built. This can improve the guidance of a few labeled target examples in the transferable component domain adaptation, and provide theoretical foundation for the map and selection of data from radar emitter in the source domain. The source examples with label noises have much impact on transferability. Thus combining with adaptive clustering algorithms, a transferable component analytic method with local selection is proposed. First we choose those examples with the same label in the local neighborhood. This hinders the transfer of source examples influenced by noise, and enables inductive transfer component analysis to be anti noise. This project introduces multiple source domain shared subspace knowledge transfer method, and designs a transfer learning method suitable for the recognition of novel radar emitters with multiple observation station.

英文关键词： label shared subspace;transfer learning;multi-source;transfer component analysis;radar emitter