- 學位論文
應用類神經網路理論於化學氣相沉積設備之故障偵測與分類
Fault Detection and Classification of PECVD Equipment Using Neural Network
摘要
現今，再TFT產業的發展中，薄膜製程為其極為重要的角色，其中變態化學氣相沉積製程為非金屬薄膜製程中舉足輕重的製程。化學氣相沉積的運用已有經數年的發展，而在TFT面板產業已逐漸變成主要的方式呈現，業界篩選出最有優勢的製作方式的製程。面對現今大尺寸面板的時代來臨，針對設備與TFT-LCD製程面都需要更為講究效益，因此降低機台故障率與提升製程良率為目前TFT業界的終極目標。本篇論文研究以類神經網路的理論為基礎以歐基里德原理為輔，偵測PECVD的設備上產生電漿的射頻供應系統的故障和PECVD設備的故障，並將故障種類做分類。因為目前機台雖會針對射頻供應系統的故障做感測，但發生故障的原因仍需現場的技術人員仔細的尋找，才能發掘真因。為此，本篇研究出如何在機台故障時偵測真正的故障位置，以縮短復機時間，提昇公司機台的嫁動率。 本研究利用PECVD的機台所得到的數據運用類神經網路訓練模擬機台正常模式並結合真實異常的資料，交互比對並訂定實際故障波型。利用SOFM原理與歐基里德原理仔細釐清確實的故障因子，來判別模糊區域的故障現象。可以確實彌補目前機台故障偵測的能力不足與強化預警的效果，不僅可以減少故障時判斷的時間增加嫁動率，也可以協助判斷製程不良導致的產品不良。
並列摘要
Currently, the thin film processes are the major process in the TFT industry. Among all the thin film processes, the chemical vapor deposition process, or CVD, is especially important for processing nonmetallic thin films. Although, over the years, different CVDs are developed and adopted in the manufacturing processes, a common and more advantageous CVD has emerged for the TFT panel display production process. As the large-size panel display becomes the mainstream product, manufacturers need to be particularly benefit and cost-conscious about the TFT-LCD manufacturing processes and equipments. Therefore, lowering the equipment faculty ratio and improving the product non-defect ratio become the urgent goals for all TFT-LCD manufacturers. This study adopts the Artificial Neural Network theory, coupled with the Euclid principle, to detect possible faults for the PECVD equipments and their plasma RF supply subsystems. These commonly reported faults are then categorized. Although the PECVD equipments have a built-in fault detection mechanism for detecting the RF supply subsystem, the fault isolation process for root cause is still largely carried out by skilled on-site technicians. To improve the fault isolation process, this study adopts the regional analysis approach to correctly point out the equipment faulty location, reduce the equipment down time, and improve equipment utilization rate. The proposed regional analysis approach first uses an Neural Network model to gather normal operational data and abnormal data of a PECVD equipment, and then compares and contrasts both data sets to generate practical faulty waveforms. This is achieved by using SOFM and Euclid principles to clarify the faulty factors and isolate fuzzy regions, which is a distinguished advantage of the regional analysis approach. This approach can complement the equipment delectability and strengthen the warning function for a PECVD equipment. With this approach, the equipment down time can be significantly shortened and equipment utilization rate improved, and most importantly, the product defect rate, caused by inaccurate judgments for problem identification in the production process, can be drastically lowered.