基于滑动多窗口的时间序列流趋势变化检测  

Sliding Multi-Windows Based Trend Change Detection on Time Series Stream

作　　者：李晓光[1] 宋宝燕[1] 张昕[1] 

机构地区：[1]辽宁大学信息学院,辽宁沈阳110036

出　　处：《电子学报》2010年第2期321-326,共6页Acta Electronica Sinica

基　　金：国家自然科学基金(No.60703068);国家自然科学基金(No.60873068)

摘　　要：趋势变化检测在时间序列流中有着非常广泛的应用.针对可变长的趋势变化检测问题,提出一种基于滑动多窗口的趋势变化检测方法,通过动态生成大尺度窗口,来适应可变长的趋势变化检测.针对内存约束下长趋势变化检测问题,提出一种基于增量PLA的窗口数据近似表示方法,给出了其欧式距离下的误差分析,进而提出一种误差修正方法来降低漏检率.大量实验表明,本文提出的检测方法具有高准确率且时间效率很高.Trend change detection has been applied widely to applications of time series stream. For the issue of detecting the change of variable length, a detection approach based on sliding multi-windows is proposed to scales up the sliding window to detect variable change. For the issue of long-term change detection with a memory constraint, a synopsis of incremental PLA is proposed to approximate to the original data,and then the error under the L2 distance is analysis theoretically,by which an amendatory L2 is given to reduce the ratio of true negative. The approach in this paper achieves quite high detection accuracy and efficiency within the extensive experiments.

关 键 词：时间序列流 趋势变化监测 分段线性逼近 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论]