摘要
声呐多传感器观测资料的数据融合问题是声呐设计中的一个重要课题.对于多基阵的声呐信号处理系统,面临的问题往往不仅是单个基阵多传感器的数据融合问题,还有多个基阵的数据融合问题.本文的研究基于数据融合的基本规则,即任何形式的数据融合,从统计平均观点来说,总体观测误差绝不大于任何单个传感器或单个基阵的观测误差.观测资料的增加从统计平均意义来说,只会带来好处,即使有垃圾资料的加入.这一结论对于人工智能领域中的深度学习来说,具有类似结果.文中提出一种用于数据融合的深度学习算法.对于独立的或相关的观测资料进行最佳的线性融合,剔除野值,进行决策级的分块数据融合,以获得统计平均意义上最小误差的结果.系统计算机模拟表明,对于受随机干扰的观测数据,采用野值过滤和数据融合的深度学习方法,能够使决策级误差显著降低.
The data fusion of multiple sensor observations is an important topic in sonar design.For multiple array sonar signal processing systems,data fusion problem is often not only the problem of multiple observations of a single array but also the data fusion for multiple arrays.In this paper,the basic principle of data fusion is studied,that is the ensemble observation error of data fusion should be,in the statistical average,no larger error than the error of any individual observation data,which is considered as one of the entities in data fusion, regardless of any trash data addition.In other words,in the data fusion process,the increase of observation data quantities should always result in some advantages in the sense of statistical average meaning.This conclusion is the same as deep learning in artificial intelligence.A deep learning algorithm is proposed in this paper for multiple observation data fusion.The optimum linear weighted combination for the independent or dependent multiple observation data is considered as a method of carrying out decision level data fusion.The wild values are picked up before further data processing,and the input data are segmented into several blocks.Data fusion is performed in each'block such that we can get minimum observation error results.The results of the system simulation conducted show that in the case where the observation data are interrupted by the interferences,i.e. the wild value,the deep learning algorithm in data fusion,derived in this paper,can considerably reduce the observation error in decision level data fusion.
作者
李启虎
卫翀华
薛山花
Qihu LI;Chonghua WEI;Shanhua XUE(Institute of Acoustics,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《中国科学：信息科学》
CSCD
北大核心
2018年第12期1614-1621,共8页
Scientia Sinica(Informationis)
关键词
声呐信号处理
多传感器观测
数据融合
深度学习
sonar signal processing
multiple sensor observations
data fusion
deep learning
