多站批次機臺之生產線派工決策

Abstract

本研究以現實環境中一座晶圓代工廠做為實驗案例，探討此晶圓廠製程之派工決策。有別於一般傳統製造業，晶圓製造廠內常有序列工作站與批次工作站兩種加工機臺，而一次加工多項物件的批次工作站會因為加工批量大小的設定與造成下游機臺產生需求不固定(lumpy demand)的情況，導致批次工作站的派工決策將嚴重影響了晶圓廠的生產績效。梁日徽 (2014)針對晶圓廠內的批次工作站提出以最大等候時間(T_max^i)做為新的派工決策，批次工作站i在滿足下列其中一個情況開始進行加工：第一，暫存區(buffer)中含有一個工件之等候時間達到T_max^i；第二，暫存區之工件數量達到滿批量。該研究以模擬法與基因演算法進行最佳解求解，然而在求解過程中需要耗費大量時間進行運算，無法有效率的進行求解。為解決梁日徽 (2014)計算時間冗長之缺點，本研究以類神經網路取代模擬法做為晶圓廠的快速績效預測工具，並同樣以基因演算法做為近似最佳解搜尋工具。實驗結果表明，使用本研究之NN-GA可以快速求解，同時，求出之近似最佳解與梁日徽 (2014)研究中所求出之近似最佳解非常接近。

This research investigates the job dispatching decision in a semiconductor fab, which includes a number of batch workstations. A batch workstation can process more than one job simultaneously; the dispatching decision (when to start a batch machine) has a substantial effect on batch size and affects manufacturing performance. Liang (2014) proposes to use the maximum waiting time (T_max^i) criterion to dispatch batch workstations, that is, batch workstation i starts processing when one of its WIP (work-in-process) waits up to T_max^i or a full batch of WIP jobs are available. He used discrete-event simulation and genetic algorithm to find a near optimal solution; yet his approach requires extensive computational efforts. To reduce computational efforts, this research uses sampled simulation results to establish a neural network (NN) for quickly evaluating the performance of dispatching decisions, and develops a genetic algorithm (GA) to find a near-optimum solution. Experiment results indicate that the NN-GA approach greatly reduces the computational efforts while the obtained solution is very close to that of Liang (2014).