卫星时序数据挖掘节点级并行与优化方法  

Node level parallel and optimization method of satellite time serial data mining

作　　者：鲍军鹏[1] 杨科 周静 BAO Junpeng;YANG Ke;ZHOU Jing(School of Electronic and Information Engineering,Xi'an Jiaotong University,Xi'an 710049,China;Military Region of Ningxia,Yinehuan 750021,China)

机构地区：[1]西安交通大学电子与信息工程学院,西安710049 [2]宁夏军区,银川750021

出　　处：《北京航空航天大学学报》2018年第12期2470-2478,共9页Journal of Beijing University of Aeronautics and Astronautics

基　　金：航天器在轨故障诊断与维修重点实验室课题~~

摘　　要：智能卫星技术对卫星时间序列数据挖掘提出了越来越多的需求。通常卫星数据计算量都非常大,若串行执行则需要较长时间。以卫星异变过程多类型特征分析过程为典型代表,针对窗口划分与向量相似度计算、特征提取、傅里叶变换、聚类等常见数据挖掘操作,探讨了在多核CPU和GPU的典型异构计算节点中对时序数据挖掘过程进行并行优化的多种策略,包括向量化方法、多进程方法、GPU计算等方法。对这几种优化策略的适用情况进行了实验分析对比。结果表明,针对不同任务情况综合使用多种优化策略具有显著提升效果。Intelligent satellite technology requires more and more data mining operations for satellite time series data. Usually,satellite data amount is very big that needs a lot of computation,so it will take a very long time to complete the computation in serial program. The satellite anomaly process multi-features analysis procedure is such a typical representation,which performs many common data mining operations,including windows segmentation,computation of vector similarity,feature extraction,Fourier transformation,and clustering. The paper discusses several speed-up and parallel optimization strategies for a time series data mining procedure on a typical heterogeneous computing node with multi-cores CPUs and GPUs,including vector optimization,multi-process parallelization,and GPU computation. We test and compare these optimization strategies in different usage conditions. The experiment results show that the combined use of them can achieve obvious efficiency improvement for different task.

关 键 词：航天大数据 数据挖掘 智能卫星 并行化 GPU 

分 类 号：V19[航空宇航科学与技术—人机与环境工程] TP311.11[自动化与计算机技术—计算机软件与理论]