A new type of change-detection scheme based on the window-limited weighted likelihood ratios
A new type of change-detection scheme based on the window-limited weighted likelihood ratios
复制标题
一种基于窗口有限加权似然比的新型变化检测方案
DOI：
10.1016/j.eswa.2017.10.051
复制
发表时间：
2018
影响因子：
8.5
通讯作者：
Lianjie Shu
中科院分区：
文献类型：
--
作者：
Fangyi He;Tiantian Mao;Taizhong Hu;Lianjie Shu
过程监控已被广泛认为是系统监控的一个重要和关键的工具，用于检测异常行为和质量改进。在制造过程或工业系统中，失控变化的几个来源，例如工具磨损、逐渐的设备劣化、振动、不一致的材料等，往往导致过程参数的动态变化，或导致系统的特殊残留信号。检测这种微弱或衰减的信号是一项具有挑战性的任务。虽然窗口受限的广义似然比（wl-GLR）方案被广泛用于变点检测并且具有一些优点，但是当其用于检测弱或衰减信号时可能表现不佳。本文提出了一种新的变化检测方案，窗口限制加权似然比（wl-WLR）方案，以改善wl-GLR方案。为此，首先定义了一种新的统计距离度量，称为GLR散度，然后分析了它的性质。wl-WLR方案被设计为监视移动窗口中GLR发散的加权平均，并且wl-GLR方案可以被视为wl-WLR方案的特殊情况。两种类型的重量泛函的wl-WLR格式的介绍和研究。基于校准样本提供了选择最佳权重参数的数值算法。广泛的仿真研究表明，所提出的wl-WLR方案检测微弱或衰减信号，其性能是强大的，即使权重参数估计不准确。本文有在线补充材料。
Process monitoring has been widely recognized as an important and critical tool in system monitoring for detection of abnormal behavior and quality improvement. In manufacturing processes or industrial systems, several sources of out-of-control variation, such as tool wear, gradual equipment deterioration, vibration, inconsistent material etc., often result in dynamic changes in the process parameter, or result in special residual signals of the systems. Detecting such weak or decaying signals is a challenging task. Although the window-limited generalized likelihood ratio (wl-GLR) scheme is widely used in changepoint detection and has some advantages, it may perform poorly when it is used to detect weak or decaying signals. This paper proposes a new change-detection scheme, the window-limited weighted likelihood ratio (wl-WLR) scheme, to improve the wl-GLR scheme. To do this, a new statistical distance measure called the GLR divergence is first defined and then its properties are analyzed. The wl-WLR scheme is designed to monitor the weighted average of the GLR divergences in a moving window, and the wl-GLR scheme can be viewed as a special case of the wl-WLR scheme. Two types of weight functionals are introduced and investigated for the wl-WLR scheme. Numerical algorithms to select the optimal weight parameters are provided based on a calibration sample. Extensive simulation study favors the proposed wl-WLR scheme for detecting weak or decaying signals, and its performance is robust even if the weight parameters are not accurately estimated. This paper has online supplementary materials.