學習曲線在TFT-LCD產業之應用研究

Abstract

薄膜電晶體液晶顯示器(TFT-LCD)是屬於高精密技術與高材料成本的產業，所以，掌控製程不良率（或良率）的變動是相當重要的工作。但是目前TFT-LCD製造廠在生產數量規劃、產品製造成本制訂與品質管制等方面，依然採用一般固定不良率的方式，由於各製程之不良率變動會影響到整個生產數量規劃、產品製造成本制訂與品質管制等作業的合理性，因此，須將不良率的變化納入考量，才能反應出實際的生產狀況與生產成本。 經過初步分析TFT-LCD的製程資料發現，每批產品的不良率在生產初期都比較高，隨著生產數量的增加有逐漸降低之趨勢，類似傳統的學習現象。因此，本研究欲將能顯現產品實際之生產學習過程的「學習曲線」加入生產數量規劃、產品製造成本制訂與品質管制作業中，以修正目前不恰當的規劃與管制方式，同時比較加入不良率學習效應前後之差異。 本研究構建多變數與單變數兩種學習曲線，由最終的實證比較結果顯示，單變數學習模式在剛開始生產的變異較多變數學習模式為大，往後生產期間內之單變數學習模式的應用成效較多變數學習模式普遍為好，探究其原因可能在於單變數學習模式考量了變數間的共線性。分析比較生產數量規劃、產品製造成本制訂與品質管制作業加入不良率學習模式前後之差異，結果發現這些作業在有考量學習效應的情況下，較未考量學習效應之狀況減少許多不必要的投入量，且有助於訂定符合實際生產情況的產品製造成本，以及運用修正後之管制圖來對真實的生產狀況進行品質管制。 上述之應用成效顯示本研究所發展出來的學習模式，可提供TFT-LCD製造廠商未來在生產類似產品時，一項值得參考的預測工具。

TFT-LCD (Thin Film Transistor-Liquid Crystal Display) industry is high technology and high material cost. Relatively, controlling the change of defect rate (or yield rate) in TFT-LCD manufacturing process is very important. But today’s TFT-LCD manufacturer still use general fixed defect rate in planning output, formulating products’ manufacturing cost, quality control, and so forth. Any change of manufacturing process’s defect rate will affect the rationality of operations in planning output, formulating products’ manufacturing cost and quality control. So, we must take the defect rate’s change into account for reflecting the real productional condition and cost. Through the initiative analysis on data of TFT-LCD manufacturing process, finding out the defect rate of every batch of products is higher in the early production. And the defect rate will decline gradually with the growth of output. This is just like the traditional learning phenomenon. Therefore, for the purpose of correcting the irrelevant operations in production planning and control, this study would like attaching the “learning curve” that can show the real productional learning process on the operations of planning output, formulating products’ manufacturing cost and quality control. Furthermore, comparing the difference of taking or not defect rate’s learning effect into these operations. This study establishes two learning curves of multivariate and univariate. From the final comparing results, the variance of univariate learning curve is larger than multivariate in the early production. But in the after production duration the application of univariate learning curve is better than multivariate. The possible reason is that univariate learning curve considers the colinear between variables. After analyzing and comparing the difference of taking defect rate’s learning effect into output planning, formulating products’ manufacturing cost and quality control, we find that these operations considering learning effect can reduce more unnecessary input than disregarding learning effect. Additionally, these operations with learning effect are helpful for making products’ manufacturing cost that can match actual production condition. And using corrected control chart to control the quality of real production status. From the above application of learning curve proposed in this study, the results indicate that can provide TFT-LCD manufacturer a useful forecasting tool to produce similar products in the future.