晶圓製造廠以階段迴歸法預測流動時間之交期模式

Abstract

交期如能準確的預測，則可以準時交貨及縮減在製品存貨使企業更具競爭 力。傳統交期模式中有關流動時間的預估大多與工作特性或現場狀況有關 ，然而卻很少考慮到彼此間的交互作用或非線性關係。因此本文採用迴歸 模式選取自變數與應變數最佳的變數型態及彼此間的相互關係等等，以求 得最佳的迴歸交期模式。由現場所獲得的實際資料確定自變數間相關性及 非線性關係以迴歸模式 來預測流動時間乃本文所欲探討的重點，且有別 於傳統交期指派來預測流動時間乃本文所欲探討的重點，且有別於傳統交 期指派模式以總體觀察值來預測 流動時間，測，在可容許的誤差模式以 總體觀察值來預測流動時間，本文採階段性的迴歸模式來預 範圍內合併 簡化預測公式，以求得更簡便更準確的流動時間預測值，乃本文的第二項 重點。再迴流型態工廠有別於一般的零工型工廠，其製程資料有可能是同 一批批量，但是不同層次製程在相同機器上加工生產，故階段迴歸預測乃 以途程中各階段為單位各自構建迴歸模式,，而不以各產品批量在各機器 所經過的流動時間構建迴歸模式來預測流動時間,以因應再迴流特性。並 以傳統交期法與階段迴歸法預估出的流動時間估計值，以各項績效指標來 評定優劣以論證階段迴歸的效度與信度。 Prediction of cycletime for a wafer fabrication has always an important issue. The flowtime is based on the processing time of operations and inter-operation time. The wafer interoperation time consists of queue time,move time and waiting time.With the regression prediction,the stage regression flowtimeestimation concept constructs an individual regressionmodel for each operationstage to estimate its individual operation stagesflowtime. The routing estima-ted cycletime is the sum of the individual estimated operation flowtime. The objective of this research is to predict precisely routing cycletime.Althoughthe stage flow time estimation conceptwill reduce forecast error, the estimation model for each operation stage has to be constructed.The will be too compl-ex especially in wafer fab which overhundred operation stage are common.A subregression merge algorithm approach is then used to group those operation stageregression equations that have the same combination of independent variables but different coefficients together.With the proposed merge stage regression model, the balance between flowtimeestimation error and model complexity can bbe achieved.The normal performance criterion index to evaluate due-date assignment models are can be summarized in terms of mean absolute deviation, mean square error, and % of late jobs. in this sample ,the complexity reduce 62%while forecast error increases a little bit than stage regression approach model without grouping. The resultshas demonstrated that staged regression estimation model is superior to sevenother traditional due-date estimation models in waferfabrication.