基於邏輯導向的生產排程知識庫系統

Abstract

生產排程在半導體前段工廠中扮演相當重要的角色。好的排程機制可以依據真實現況，合理地調配整體資源，為企業帶來更佳的效益。由於多樣化的生產製程，生產環境的變化，以及不同的生產需求，在大部份的工廠中，由生產排程專家依據其知識及經驗法則來進行生產排程，藉以達成較優異的生產績效。但是，往往必須有足夠的產業經驗，並且歷經數年的實際磨練後，才能培養出一位稱職的排程專家; 除此之外，每個工廠的生產環境不同，所需的排程知識亦不盡相同，如何將排程知識保留下來，並且能輕易地傳承，也是這個領域非常重要的問題。 欲解決上述問題，我們提出一套完整的生產排程知識庫系統。在這篇論文中，為了讓生產排程的知識能妥善地保存、輕易地複製沿用以及簡單地表達，進而定義了以邏輯為導向，且易於機器讀取的圖形化XML-based知識表達法(以下簡稱為XLogic)，知識擷取方法(以下稱為KA)跳脫過去訪談式的KA，提出圖形化的知識擷取工具(以下簡稱Logic Map Builder)，直接由專家以圖形拖拉的方式維護排程知識元件(以下稱為Logic Component)，產生排程邏輯圖(以下稱為Logic Map)，不僅便於操作，並且更容易了解整個排程的過程。架構上可分為三階段：Meta Knowledge Acquisition(以下稱為MKA)階段，專家可依據知識的重要度及順序性去建構Meta Knowledge(以下稱為MK)；KA階段，由專家透過圖形化的Logic Builder產生排程Logic Map進而轉換至NORM-based的規則庫中；排程推論階段，專家經由程式介面，輸入Lot相關排程資料後，再利用DRAMA推論引擎，推論產生排程結果。 針對排程精確度與準確度於系統與專家的比較上，我們發現大部分的情況皆能彼此符合。透過專家的回饋，證實本系統確實有保存延續排程知識之能力，並可作為新進排程人員之教學系統。

Lot scheduling plays a very important role in semiconductor wafer Fab. According to the up-to-date market situation, a good Lot scheduling mechanism can help wafer Fab to meet constraints and allocate reasonable resources to gain higher profit. The complexity of wafer Fab Lot scheduling comes from various process flows, dynamic manufacturing environment and variant market requirement. In order to increase average performance, in the most wafer Fabs, the Lot scheduling expert (Production Controller) utilizes her/his knowledge and experience to do the Lot scheduling. However, it's difficult to train a comprehensive Lot scheduling human expert who must have enough industry experience and real Fab practice through a couple of years. In addition, different Lot scheduling know-how is required in different wafer Fab. How to share and re-use the existing knowledge is also the main issue in this domain. For those reasons, we propose a suite of Lot scheduling knowledge-based system. In this thesis, we define a logic-oriented XML-based knowledge representation for the usage of knowledge repository, knowledge re-use and graphic expression. Besides, we design a graphic knowledge acquisition tool (Logic Map Builder) to elicit the Lot scheduling logic construction (Logic Map). Therefore, human expert can directly construct the Logic Map by the drag-drop function. There are three phases in this system. Firstly, Meta-Knowledge Acquisition Phase is designed to acquire the importance and sequence of knowledge components, says meta-knowledge. Secondly, Knowledge Acquisition Phase is designed to acquire Lot scheduling Logic Map. Finally, Lot Scheduling Phase is designed for knowledge inference by DRAMA to receive relevant facts and generate Lot scheduling result. The comparison of precision and accuracy between the results of our system and human expert is processed. The results of most cases are totally matched. From the feedback of human expert, we assure that the system not only has the capability to reposit and re-use Lot scheduling knowledge, but also can be used as a Lot scheduling tutoring system.