The Research On Change-point Detection Method To Identify The Local Changes In Linear Profile

In the process of industrial production, quality of a process is not only characterized by aunivariate or multivariate quality characteristics, but sometimes is characterized by arelationship between a response variable and one or more explanatory variables, such aslinear, nonlinear and so on. For the monitoring of linear profile, most current researches arebased on the assumption that global changes were taken in the linear profile, this means thatthere are shifts in the global parameters such as intercept and slope. For this change pointdetection problem, firstly the intercept or slope parameters of linear profile model can beeasily estimated. Then the monitoring of linear profile can be converted into the monitoringof multivariate vector of linear model.However, the possibility that profile have global changes becomes slimmer as theimprovement of working accuracy in manufacturing. So some effective detection methodsshould be found to monitor local changes in profile. Because the existence of local changes inprofile can affect the estimation of global model parameters, the accuracy of monitoringestimated parameters will decline.In Phase I stage of SPC (Statistical Process Control), we are not only interested in thedetecting the occurrence of changes, but also interested in detecting the position and amountof changes. Similarly, for the linear model, we not only monitor the changes of modelparameters, but also detect the accurate position of changes.So for the situation local changes taken in profile, we propose a Phase I change-pointdetection method based on T2Statistic. By reasonably segmenting the profile throughclustering method, we regard the characteristic values of all these segments as multivariate,and then combine the T2statistic to design change point detection method. And then bysimulation method we analyze the performance under different situations, such as differentscale of local changes, different positions of changes and different segmenting amount. Thesimulation results show that this method is an effective tool to identify the local changes oflinear profile in Phase I analysis.