Detecting Outliers In Complex Profiles With Nonlinear Reduction Methods

After statistical process control(SPC)was put forward by Shewhart in 1924,quality management came to the time of statistical quality management.Then,statistical methods have permeated into every part of quality improvement,including designing,processing,managing and services.As SPC being mature,new branches came out.One of these is profile monitoring,which considers the function relationship within several quality characters of products.SPC can be divided into two phases: in Phase I a model is built based on history data set,which has reduced the noise and other interference;in Phase II,new products are monitored one by one to identify whether they're outliers or other special points.In the field of profile monitoring,the present technologies has been studied for years,while in the situation that data set is with pretty high dimension or that the function can hardly be fitted,these technologies show their drawbacks.To reduce the difficulties and guarantee the accuracy of outlier detection,a new method is put forward in this dissertation,which tries to reduce the dimensions first by nonlinear dimension reduction methods before modelling.Outlier-detection is an important part in profile monitoring.Outliers can interfere the modelling process,which may influence the on-line process,so it's really necessary to deal with the outliers.With dimensions getting larger,the detection becomes harder,the ?~2 control chart becomes less sensitive.The combination we proposed can directly deals with this problems.This dissertation tries to use the combination of nonlinear dimension-reduction methods with ?~2 control chart to identify outliers in profiles in Phase I.After discussing the detail about the combination,some simulations and case studied will be showed to confirm the usage and the ability of the method we proposed.