工程科学学报，第39卷.第10期：1584-1590,2017年10月 Chinese Journal of Engineering,Vol.39,No.10:1584-1590,October 2017 D0L:10.13374/j.issn2095-9389.2017.10.018;htp:/journals..usth.edu.cn 基于深度卷积神经网络的地磁导航方向适配性分析 肖晶)，齐晓慧)，段修生)区，王俭臣) 1)陆军工程大学，石家庄0500032)中国人民解放军驻西北工业大学军事代表室，西安710065 区通信作者，E-mail:sjzdxsh@163.com 摘要针对地磁导航方向适配性分析时人工提取的特征主观性较强且难以表达深层的结构性特征的问题，提出一种基于 深度卷积神经网络(convolutional neural network,CNN)的地磁导航方向适配性分析方法.首先，利用Gabor滤波器的方向选择 特性建立了6个典型方向的适配特征图：然后，设计了卷积神经网络对深层次的方向适配特征进行提取，并通过混和粒子群 算法(hybrid particle swarm optimization,HPSO)对卷积神经网络的训练参数进行优选：最后，通过仿真实验对所提方法进行了 验证.结果表明，该方法可有效避免复杂的计算以及人工特征提取的盲目性，实现了地磁导航方向适配性分析的自动化，且 所提方法的准确率高于传统的BP网络和支持向量机，对地磁导航和航迹规划具有指导意义. 关键词地磁导航；适配性分析；方向适配性；卷积神经网络；Gbor滤波器 分类号TP391.1 Direction-matching-suitability analysis for geomagnetic navigation based on convolu- tional neural networks XIAO Jing),QI Xiao-hui),DUAN Xiu-sheng,WANG Jian-chen2) 1)Army Engineering University,Shijiazhuang 050003,China 2)Govemment Representative Office in Northwestem Polytechnical University,Xi'an 710065,China Corresponding author,E-mail:sjzdxsh@163.com ABSTRACT Aimed at the problems of artificial direction matching features being too subjective to analyze magnetic matching suita- bility and deep architectural features that can't be extracted,a new matching suitability analysis method based on a convolutional neural network(CNN)is proposed.First,direction-matching-suitability feature maps in six typical directions are established using the Gabor filter's direction selection characteristics.Second,a CNN is designed to extract the deep direction features.The training parameters of the CNN are optimized with a hybrid particle swarm optimization(HPSO)algorithm.Finally,simulation experiments are conducted to verify the proposed method.Results show that the method can effectively avoid complicated calculations and blindness when artificially extracting direction features,and the direction-matching-suitability analysis for magnetic navigation can be achieved automatically.The method's analysis accuracy is higher than in the traditional BP neural network(BPNN)and support vector machine(SVM),and has practical implications for geomagnetic navigation and route planning. KEY WORDS geomagnetic navigation;matching suitability analysis;direction matching suitability;convolutional neural networks; Gabor filter 地磁导航是一种基于地球地理信息的导航方式，要补充.地磁导航的精度不仅与导航算法有关，还与 具有无源、无辐射、隐蔽性强和误差不随时间积累的特 地磁图的特征密切相关：在磁场特征明显、信息量丰富 点，可进行全天候、全地域导航，是惯性导航方式的重 的区域进行导航能够有效提高导航的精度和实时性. 收稿日期：2016-12-05 基金项目：武器装备军内科研重点资助项目(2014551)工程科学学报,第 39 卷,第 10 期:1584鄄鄄1590,2017 年 10 月 Chinese Journal of Engineering, Vol. 39, No. 10: 1584鄄鄄1590, October 2017 DOI: 10. 13374 / j. issn2095鄄鄄9389. 2017. 10. 018; http: / / journals. ustb. edu. cn 基于深度卷积神经网络的地磁导航方向适配性分析 肖 晶1) , 齐晓慧1) , 段修生1) 苣 , 王俭臣2) 1) 陆军工程大学, 石家庄 050003 2) 中国人民解放军驻西北工业大学军事代表室, 西安 710065 苣通信作者, E鄄mail: sjzdxsh@ 163. com 摘 要 针对地磁导航方向适配性分析时人工提取的特征主观性较强且难以表达深层的结构性特征的问题,提出一种基于 深度卷积神经网络(convolutional neural network, CNN)的地磁导航方向适配性分析方法. 首先,利用 Gabor 滤波器的方向选择 特性建立了 6 个典型方向的适配特征图;然后,设计了卷积神经网络对深层次的方向适配特征进行提取,并通过混和粒子群 算法(hybrid particle swarm optimization, HPSO)对卷积神经网络的训练参数进行优选;最后,通过仿真实验对所提方法进行了 验证. 结果表明,该方法可有效避免复杂的计算以及人工特征提取的盲目性,实现了地磁导航方向适配性分析的自动化,且 所提方法的准确率高于传统的 BP 网络和支持向量机,对地磁导航和航迹规划具有指导意义. 关键词 地磁导航; 适配性分析; 方向适配性; 卷积神经网络; Gabor 滤波器 分类号 TP391郾 1 Direction鄄matching鄄suitability analysis for geomagnetic navigation based on convolu鄄 tional neural networks XIAO Jing 1) , QI Xiao鄄hui 1) , DUAN Xiu鄄sheng 1) 苣 , WANG Jian鄄chen 2) 1) Army Engineering University, Shijiazhuang 050003, China 2) Government Representative Office in Northwestern Polytechnical University, Xi爷an 710065, China 苣Corresponding author, E鄄mail: sjzdxsh@ 163. com ABSTRACT Aimed at the problems of artificial direction matching features being too subjective to analyze magnetic matching suita鄄 bility and deep architectural features that can蒺t be extracted, a new matching suitability analysis method based on a convolutional neural network (CNN) is proposed. First, direction鄄matching鄄suitability feature maps in six typical directions are established using the Gabor filter蒺s direction selection characteristics. Second, a CNN is designed to extract the deep direction features. The training parameters of the CNN are optimized with a hybrid particle swarm optimization (HPSO) algorithm. Finally, simulation experiments are conducted to verify the proposed method. Results show that the method can effectively avoid complicated calculations and blindness when artificially extracting direction features, and the direction鄄matching鄄suitability analysis for magnetic navigation can be achieved automatically. The method蒺s analysis accuracy is higher than in the traditional BP neural network (BPNN) and support vector machine (SVM), and has practical implications for geomagnetic navigation and route planning. KEY WORDS geomagnetic navigation; matching suitability analysis;direction matching suitability; convolutional neural networks; Gabor filter 收稿日期: 2016鄄鄄12鄄鄄05 基金项目: 武器装备军内科研重点资助项目(2014551) 地磁导航是一种基于地球地理信息的导航方式, 具有无源、无辐射、隐蔽性强和误差不随时间积累的特 点,可进行全天候、全地域导航,是惯性导航方式的重 要补充. 地磁导航的精度不仅与导航算法有关,还与 地磁图的特征密切相关:在磁场特征明显、信息量丰富 的区域进行导航能够有效提高导航的精度和实时性