以RFID為基的晶圓廠即時排程系統之發展

Abstract

在半導體晶圓廠的生產規劃中，如何達到生產運作的全自動化，提高廠區設備的生產效率，是一極為重要的研究方向。從接到訂單時，依客戶重要性或訂單相關性，排單進入廠區生產；在訂單進入廠區之後，除了產品本身的加工時間，還要根據機台的使用率有不等的等候時間，最後在到期日前要交貨；所以排程是晶圓廠中的重要部分，可以有效控制時間，縮短生產週期及減少延遲個數。 本論文的主要目的在藉由RFID (Radio Frequency Identification)技術為基礎，提出一個排程和重新排程系統。此系統有三個主要的模組，分別為即時決策排程器(real-time decision-making scheduler)、考慮機台損壞和瓶頸機台的重新排程法則及彈性製造系統模擬平台(Simulation Platform of Flexible Manufacturing System, SPFMS)。 在決策排程器的建構上，我們利用RFID收集系統狀態，並以GA based FFEI (Genetic Algorithm based Fuzzy-entropy-based Feature Evaluation Index)選出在三個績效指標：產出量(Throughput)、平均流程時間(Mean Flow Time)、延遲交貨個數(Tardy Number)中顯著的屬性(salient attribute)，再送進支向機(Support Vector Machine)做訓練，如此屬性可以找到與之相對應的派工法則，應用在適當的時間點做排程的依據。重新排程法則的提出則是為了因應晶圓廠內有不可預期失效值的發生，如機台損壞、瓶頸機台，而影響系統績效衡量；所以，我們考慮機台的平均無故障時間(Mean Time between Failure)和平均故障修復時間(Mean Time to Repair)造成的屬性變化，再將這些屬性送進決策排程器做重新排程，使得績效更佳。在模擬平台的實現上，我們利用模擬工具eM-Plant建構彈性製造系統模擬平台，並以合於國際標準的300mm晶圓廠和TRC fab為範例實做，來驗證我們所提出的排程法則跟重新排程法則。 最後，所實現的模擬平台與排程控制系統經驗證後，在績效指標的表現上，SEMATECH的結果單就平均流程時間(Mean Flow Time)比較，約減少3%(203.32－＞197.307小時)；而延遲個數則平均減少9個(63－＞52個)，約降低17%；平均產出量(Throughput)則增加了約0.6%(176852－＞177951 wafer)。TRC fab的結果則平均流程時間(Mean Flow Time)約減少4%(451.23－＞427.86小時)；而延遲個數則平均減少3個(15－＞12個)，約降低25%。 藉由GA based FFEI可以找出表現個別績效指標較好的顯著屬性，也可以選擇綜合三個績效指標的屬性，透過這樣的選擇再進行模擬後，此模擬結果證實，我們所提出的架構及方法達到了全自動化之生產運作的目標，不但線上收集資料迅速，於排程效能上也優於單一派工法則，更能有效解決機台損壞和瓶頸機台所影響的績效表現。

In the fab planning, how to achieve a fully automated manufacturing to replace human operator in performing runtime scheduling and dispatching decisions is a very important issue. This thesis constructs a real-time scheduling and rescheduling system based on a new technology, RFID (Radio Frequency Identification), for fully automated fabs. The scheduling and rescheduling system consists of three modules: a real-time decision-making scheduler, a rescheduling mechanism for machine breakdowns and bottleneck, and a simulation platform for flexible manufacturing system (FMS). In order to develop a real-time decision-making scheduler, we use RFID to collect system attributes and choose the salient attributes by GA based FFEI (Genetic Algorithm based Fuzzy-entropy-based Feature Evaluation Index). Then, the salient attributes corresponds to dispatching rules by using support vector machine (SVM). The proposed rescheduling framework is used when unknown machine breakdowns and bottleneck happen. We define the bottleneck tools with utilization of machines or queuing length, and then implement the rescheduling strategies. Next, we develop the simulation platform by using object-oriented software, eM-Plant. The implemented scheduling and rescheduling system have been demonstrated by two example manufacturing models, a 300mm factory fab of SEMATECH and TRC fab. In SEMATECH model, the mean flow time is reduced 3%(203.32－＞197.307 hours) and tardy numbers are decreased 17% (63－＞52 lots). We also have more throughput (176852－＞177951 wafer). In TRC fab model, the mean flow time is reduced 4% (451.23－＞427.86 hours) and tardy numbers are decreasing 25% (15－＞12 lots). The experimental results of simulation are better than previous researches. The dispatching rule sets proposed by SVM scheduler have higher performance than single dispatching rule. This scheduling and rescheduling framework can consider time of machine breakdowns and bottleneck and improve the system immediately.