基于精细可靠度评估的优化证据推理及应用

Evidential reasoning is expert at the representation and combination of uncertain information, and it has become a research hotspot. In the multi-source information fusion, the different sources of evidence may have different reliabilities, but the reliability of the belief of each focal element in evidence cannot be accurately estimated using the existing reliability evaluation methods. In the traditional evidence fusion methods, the optimization of the fusion model based on the fusion performance evaluation criteria is not considered, and this makes the accuracy of the fusion result cannot be guaranteed. In this program, the optimized evidential reasoning technique based on refined reliability evaluation will be studied. The evidence inner confusion reliability matrix will be estimated using training knowledge, and then it will be studied how to carefully modify the evidence under the power-set according to the reliability matrix. By doing this, it is expected to greatly improve the evidence accuracy. For the dynamic information, the statistic feature of the historic evidence will be extracted, and the compatibility measure of evidence will be defined. Then the online fast evaluation of the relative reliability of evidence will be studied. In the fusion procedure, the unified combination model for complex information sources will be constructed, and the comprehensive evaluation criteria of fusion performance will be proposed. Then, the optimal evidential reasoning technique will be studied based on the joint optimization of the multiple criteria, which could be able to efficiently increase the robustness and reliability of the fusion system. These proposed methods will be analyzed and improved through the experiments, and the new methods will be applied in the multi-source information fusion target identification. This program is potential to open a new research topic for the reliability evaluation of information source and the evidential reasoning.

证据推理因其在不确定信息表示与融合上的独特优势而成为了研究的热点。信息融合时，不同信源证据可能有不同的可靠度，目前的可靠度评估方法不能精细地评估证据中各焦元信息的可靠性，且现有证据推理方法没有考虑利用融合性能评价指标对融合模型自动优化，这使不同情况下融合结果的准确性难以得到有效保证。本项目拟研究基于精细可靠度评估的优化证据推理。充分利用训练知识估计证据内可靠度混淆矩阵，基于此研究幂集下证据精细修正方法，从而提高证据准确性。针对动态信息，提取历史证据统计特征，定义证据相容度量测，研究证据相对可靠度在线快速评估方法。构建复杂信息统一融合模型，定义融合性能综合评价指标，研究多指标联合优化证据推理，以增强融合系统的鲁棒性和可靠性。通过实验分析不断完善提出的系列方法，并将新方法应用到多源信息融合目标识别。本项目将为信源可靠度评估和证据推理开辟一种新的研究思路。

多源信息融合能够增强目标识别的准确性与可靠性，证据推理因其在不确定信息表示与融合上的独特优势而成为了一种重要的高层信息融合方法。在多源信息融合识别中，不同信源证据可靠度一般不同，传统可靠度评估方法难以精细评估证据中各焦元信息可靠性，无法通过可靠度加权处理改善信源质量，且现有证据推理方法在处理异构高冲突等复杂信息时难以对融合模型自动优化，这难以保证不同情况下多源证据融合的可靠性。因此，本项目研究了基于精细可靠度评估的优化证据推理。在信源证据可靠性评估中，构建了证据内可靠度矩阵，提出了基于邻域样本误差的证据精细内可靠度混淆矩阵估计算法，设计了证据置信谨慎重分配规则，能够有效改善信源质量。在动态高冲突证据融合中，定义了证据柔性相容度量测，揭示了信源证据相对可靠度，可合理区分信源冲突性和互补性，进而通过对证据加权处理能够降低欺骗干扰等对融合结果的影响。当多源异类信息目标辨识框架不一致时，设计了多粒度证据等价转换规则，构建了异构复杂信息统一融合模型，提出了多源信息联合优化证据推理，定义了融合性能评价指标，通过对参数优化学习提高目标识别率。在多传感器融合空中目标识别中，针对部分类别目标训练样本缺失的情况，提出了不完备信息融合目标检测识别方法，发展了基于生成判别混合模型的空谱特征融合雷达目标识别技术，构建了时空多源动态信息优化融合识别模型，能够对野目标准确检测，显著提升融合识别的准确性和鲁棒性。研究成果有效解决了机载多传感器融合识别系统中多种探测模式下特征信息大差异、高冲突的问题，能够对空中和海面目标准确识别，对相关工程型号项目提供了有力技术支撑。. 项目研究成果在 IEEE TCYB/TFS/TIP/TGRS/TNNLS、中国科学：信息科学等国内外著名期刊和国际信息融合大会上共发表论文16篇，其中SCI期刊论文11篇，国际会议论文5篇，授权国家发明专利3项，受理国家发明专利5项，培养博士生5人、硕士生7人，举办信息融合学术研讨会4次，项目负责人入选国家青年拔尖人才，获陕西省科学技术一等奖、中国航空学会青年科技奖。