基于厚尾分布的异常信号检测方法  

Abnormal Signal Detection Method Based on Heavy-tailed Distribution

作　　者：王俐莉[1,2] 刘力维[3] 熊艳晔[1] 

机构地区：[1]海军指挥学院科研部,江苏南京210016 [2]解放军理工大学指挥自动化学院,江苏南京210007 [3]南京理工大学理学院,江苏南京210094

出　　处：《南京理工大学学报》2011年第1期16-20,共5页Journal of Nanjing University of Science and Technology

基　　金：国家自然科学基金(60874118)

摘　　要：为了检测具有厚尾分布信号的异常值,提出将基于EM算法的统计诊断方法应用于正态逆高斯信号的异常信号检测。为了避免计算贝塞尔函数的困难,考虑将正态逆高斯信号看作缺失变量,通过EM算法求解参数的极大似然估计,简化了计算过程;通过Q函数代替对数似然函数,提出了在正态逆高斯分布的信号中基于Q函数的影响度量方法,分别给出了数据删除模型与局部影响分析的诊断统计量,理论和实例计算结果表明:该文所提出的检测方法对于正态逆高斯信号的检测效果明显。In order to detect the abnormal signals with a heavy tailed distribution,the statistical diagnosis method based on expectation maximization（EM）algorithm is applied to abnormal signal detection for normal inverse Gaussian signal.To avoid the difficulties of calculating Bessel functions,the normal inverse Gaussian signal is treated as missing data.The maximum likelihood estimation of parameters is obtained using the EM algorithm,simplifying the calculation.The Q function takes place of the log-likelihood function and a Q-based measurement for the diagnosis method of normal inverse Gaussian distribution is proposed.The results of the calculation example and computer simulation demonstrate the superiority of the methods proposed here.

关 键 词：期望最大化算法 极大似然估计 COOK距离 Q函数 异常信号 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]