訂單生產策略下氧化銦錫導電玻璃產業之良率預測研究

Abstract

由於蘋果（Apple）iPhone在技術端的突破，使得觸控面板成了科技產業近來備受關注的新題材。觸控式面板上游材料所需要的「透明導電薄膜」（ITO Film）幾乎都掌握在日商手中，這部分的技術與專利，台灣業者無法打入市場，只有在「氧化銦錫(ITO)導電玻璃」的製造有極大的機會與經驗。本研究將以觸控面板產業的上游零件氧化銦錫導電玻璃產業做為研究對象，透過建立有效率的良率預測模式，來改善生產規劃的方法，更準確地回應生產和市場需求的變化。生產規劃的重要活動之一便是良率預測，經由有效準確的預測，方可進行如主生產排程(Master Production Planning；MPS)與物料需求規劃(Material Requirement Planning；MRP)等細部之計畫與控制。

時間序列的預測模式基本上可分為兩種型態：定性方法（qualitative methods）與定量方法（quantitative methods）。本研究使用時間序列的定量方法做為本研究的研究方法，選擇常用預測模式中的移動平均法、指數平滑法及趨勢分析法針對個案公司的歷史良率資料進行資料計算，並得到三種結果，輔以誤差評估公式之平均絕對標準差(Mean Absolute Deviation；MAD)及平均絕對百分誤差(Mean Absolute Percent Error；MAPE)來評估使用三種預測模式及目前所使用的固定良率法，來比較何者有較優之預測能力。

分析個案公司的歷史良率資料並使用上述的預測方法進行良率預測，得到的績效結果最好至最差依序為：趨勢分析法＞指數平滑法＞移動平均法＞固定良率法。使用趨勢分析法進行良率預測時，以MAPE做為績效評估所得到的結果均<10%，依據Lewis對預測的能力之分級，屬於高度準確的預測，而且所得的結果具線性趨勢，可提供趨近實際良率的預測值。因此，由個案公司的資料驗證所得的結論，將可供氧化銦錫導電玻璃產業做為良率預測的計算參考，經由更準確的良率預測以計算出符合業務需求量的基材投入量，進而減少成品庫存量，節省庫存成本。惟本研究無法由歷史資料得到季節變化趨勢，因此建議後續研究應在線性趨勢中加入季節因子分析作為修正；或進一步利用迴歸分析與其他可能影響良率的相關因子比較；或是使用定性法找出直接影響良率的相關因子做相關之研究。

With the breaking-through of touch screen technology by Apple Company, touch screen becomes the focus of TFT industry and is now widely rolling out in the market of smart phone. However, key technology and patents of critical raw material of tough screens(ITO film) is still mastered by Japanese Companies. The only opportunity for Taiwan’s touch screen upstream manufacturing companies is in ITO glass. Through strategically production planning and controlling, stock can be reduced dramatically. One of the imperative production controlling is yield forecast. With accurate yield forecast, master production planning and material requirement planning can be conducted efficiently. Forecasting model of time serial method can be divided into two types: qualitative method and quantitative method. In this research, quantitative method is chosen as researching method and Moving Averages, Exponential Smoothing and Techniques for Trend are used as forecasting models. Mean Absolute Deviation and Mean Absolute Percent Error are used as evaluation methods to measure the performance of yield forecasting models. Using historical yield data of company A as the input of 4 forecasting models, Techniques for Trend is proofed as the best model for predicting manufacturing yield, meanwhile, fixed yield method is the worse. The evaluation result MAPE of Techniques for Trend model is all less than 10%. According to Lewis‘s forecasting accuracy category, it is in high accuracy level. Through measuring the deviation of yield forecast in company A, Techniques for trend is proofed to be better than the other forecasting models. The results obtained in this research may not represent all the companies in the industry but could be an important input of ITO glass industry.