Identifying Change-point In Polynomial Profiles Based On Data-segmentation

In industrial production, quality engineers recognize that statistical process control(SPC) is a significant tool to monitor the condition of the quality characteristic.Retrospective analysis (Phase I) and monitoring analysis (Phase II) are the two corecomponents in constructing control chart. Phase I is aimed at extracting baseline dataand forming statistical model to reflect manufacturing’s condition through investigatehistorical data set. On the other hand, Phase II means to compare the observation datain the future with the basic statistical model obtained from Phase I. If these newobservation cannot approximate to the model, it will diagnose that assignable causeslead to out-of-control process. These causes should be timely studied and then removedso that the quality could recover stability. There are six different types of process,including ordinary, multi-stage, profile quality characteristic, auto-correlated, highyield and specific process. Profile monitoring is the use of control charts for cases inwhich the quality of a process or product can be characterized by a functionalrelationship between a response variable and one or more explanatory variables. Unlikethe linear profile’s simple structure, the non-linear profile has relatively less attainmentbecause of highly complexity. Regression model is the initial method to analyze thephase I of non-linear profiles, but it lacks of sensitivity for local characteristic changes.In order to deal with localized shift during uncertain input value range withinprofile, the paper proposed a new ideology which consists of two major parts: a data-segmentation component to concisely detect local change’s location by overlaying gridpoints and a change-point detection component via the maximum likelihood estimate.Simulated data set of a polynomial profile are used to illustrate the effectiveness of thepropose strategy with comparison of typical T2multivariable statistics. From the respectof err I, err II and non-signal probability, it demonstrates new method’s goodperformance. Then this paper has discussed the capacity of identification for differenttypes of coefficient change under different proportion of local shift.In the end, this paper discussed the advantage and disadvantage of the newframework, and made conclusions. Some promising research plan for the topic of localshift has been put forward.