卷积目标分类网络的可信评估建模方法  

Credible Evaluation Modeling Approach for Convolutional Object Classification Networks

作　　者：曹恩龙 刘赫炎 孙智孝[1] 罗庆[1] 刘环宇 CAO En-long;LIU He-yan;SUN Zhi-xiao;LUO Qing;LIU Huan-yu(AVIC Shenyang Aircraft Design&Research Institute,Shenyang 110000,China;Institute of Automated Testing and Control,School of Electronics and Information Engineering,Harbin Institute of Technology,Harbin 150001,China;Information Countermeasure Technology Research Institute,Faculty of Computing,Harbin Institute of Technology,Harbin 150001,China)

机构地区：[1]航空工业沈阳飞机设计研究所,辽宁沈阳110000 [2]哈尔滨工业大学电子与信息工程学院自动化测试与控制研究所,黑龙江哈尔滨150001 [3]哈尔滨工业大学计算学部信息对抗技术研究所,黑龙江哈尔滨150001

出　　处：《测控技术》2022年第10期66-72,共7页Measurement & Control Technology

基　　金：国家自然科学基金项目(61671170)。

摘　　要：为了提升目标分类系统的可靠性和智能空战场景的适应性,需要对目标分类系统的可信性进行定量评价,现存的目标分类系统缺少完善的可信性评价方法,所以提出了以数据为驱动,围绕目标分类系统“数据+算法+平台”三要素的可信性评估方法与准则。提出了基于黑盒和白盒的可信性评估模型,构建了目标分类系统泛化能力评估模型和目标分类系统鲁棒能力评估模型,多方面评估系统的工作能力,从根本上保证了目标分类系统的可靠性。通过生成测试用例并使用评估模型对分类网络结果进行可信性评估,形成了目标分类系统的可信性评估方法,可用于支撑、完善目标分类系统的可信性评价技术体系,提升目标分类系统的可靠性分析能力,同时为分类算法的可用性提供建议。搭建了目标分类系统可信性评估平台,该软件平台集成了数据生成模块、系统泛化性指标计算模块和系统鲁棒性指标计算模块,从泛化性、鲁棒性不同角度对目标分类系统进行了评估。In order to improve the reliability of the target classification system and the adaptability of the intelligent air combat scene,it is necessary to quantitatively evaluate the credibility of the target classification system.The existing target classification system lacks a complete credibility evaluation method.Therefore,a data-driven credibility evaluation method and criteria based on the three elements of “data + algorithm + platform” of the target classification system is proposed.The credibility evaluation model based on black box and white box is proposed,and the generalization ability evaluation model and robust ability evaluation model of target classification system are constructed.The working ability of the system is evaluated in many aspects,which fundamentally ensures the reliability of the target classification system.By generating test cases and using the evaluation model to evaluate the credibility of the classification network results,the credibility evaluation method of the target classification system is formed,which can be used to support and improve the credibility evaluation technology system of the target classification system,improve the reliability analysis ability of the target classification system,and provide suggestions for the availability of the classification algorithm.The credibility evaluation platform of the target classification system is built.The software platform integrates the data generation module,the system generalization index calculation module and the system robustness index calculation module,and evaluates the target classification system from different perspectives of generalization and robustness.

关 键 词：目标分类系统 可信性 泛化能力 鲁棒能力 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构] V241[自动化与计算机技术—计算机科学与技术]