- 學位論文
TFT-LCD缺陷檢驗之搜尋與分類模式
Defect Search and Classification in TFT-LCD Inspection
摘要
現今TFT-LCD的解析度與尺寸的需求逐年在增加，而製程的缺陷檢驗系統必須以最快速的方式去搜尋出面板缺陷的發生與否。雖然TFT-LCD產業已運用大量的自動化技術於製程中，目前前、中、後段製程的缺陷檢測仍然仰賴大量人工目視進行，其原因在於機器尚無法完全取代人員對於缺陷搜尋與缺陷判別的適應性與靈活性。因此，對於面板檢驗工作而言，品質決策仍需以人眼視覺為依歸。本研究利用凝視追蹤系統與卡片分類方法蒐集TFT-LCD製程模組端檢驗員之心智模式資料，藉此瞭解生手在訓練過程的變化與專家與生手對於缺陷檢驗方式與策略的差異。實驗結果顯示，訓練階段有助於改善搜尋形式、縮短搜尋時間與判定時間及增加判別面板缺陷之正確率，尤其以檢驗動作指導之第二階段訓練對於生手心智模式的養成最具影響。因此，程序式知識的獲得乃是績效提升的關鍵。根據注意力特徵整合理論，本研究建議訓練第二階段之檢驗動作指導與其第三階段之缺陷介紹互調，以提升缺陷檢驗訓練之績效。另外，實驗結果發現，專家大部分照著規範行事，但在一些缺陷檢驗中衍生出自己的搜尋策略。因此，有一些顯而易見的缺陷反而花費較多時間檢驗。此現象可以Rasmussen之SRK模型解釋此一人為失誤。本研究認為定期的績效考核以判斷檢驗員的檢驗方式，有助於確認與維持缺陷檢驗的績效。
並列摘要
As the demands of higher resolution and larger size of TFT-LCD, it becomes more critical for the TFT-LCD industry to find an efficient and effective way of defect inspection. Although the TFT-LCD industry has put lots of automatic techniques into practice, the defect inspection of Array, Cell and Module still heavily relies on human visual system. It is because of the sensitivity and adaptability of human eyes that cannot be replaced by machines. Therefore, it is the inspectors who make a final decision about the quality of product. The purpose of this study is to apply gaze tracking and cart sorting methods for collecting the data of inspectors’ mental model and their inspecting behaviors. The data was analyzed to understand the training effect on the novice inspectors and the differences between experts and novices on their defect inspection. The results showed that the training improved the search pattern, defect decision time, and accuracy of the inspection. Especially, the second phase of the training about the procedures of the inspection had the most influence on these variables. It concluded that the acquisition of the procedural knowledge was a key factor for enhancing the performance of defect inspection. Based on the feature-integration theory of attention, it was suggested that the second phase of the training should be switched with the third phase that common defects was introduced. We also found out that generally the experts followed the standard operation process. However, they also created their own search strategies and patterns. The results showed that, for some defects, the experts took more inspection time than the novices did based on their own search patterns. This phenomenon may be explained by Rasmussen’s SRK model on human errors. It was suggested that a periodical review process will be necessary to sustain the performance of defect inspection.