- 學位論文
模糊神經網路於半導體BGA影像辨識系統的製作
Identification of BGA Defects Using Fuzzy Neural Networks
若您是本文的作者，可授權文章由華藝線上圖書館中協助推廣。
摘要
近來，在政府之亞太科技島的政策下，台灣半導體工業為執行的重點之一，在高度的國際競爭壓力下，半導體工廠自動化成為決定勝負的關鍵。BGA(Ball Grid Array)是半導體工業中的重要製程之一，但在製程中它的檢測工作，卻需藉由人工目檢來完成，使得生產線無法達到全面自動化，而造成生產瓶頸。因此本文針對BGA所可能發生的五種缺點情形，製作離線(off-line)BGA影像辨識系統。此辨識系統是由BGA影像前處理、BGA特徵值的選定及模糊自我成圖(Fuzzy Self-Organizing Map, Fuzzy SOM)神經網路的應用檢測三部份所組成。在BGA特徵值的選取方面，乃根據X軸影像投影法所得到的灰階直方圖，擷取樣本平均數、樣本變異數、樣本偏態及樣本峰態等四個特徵值，再經由正規化的處理，使之成為Fuzzy SOM網路的輸入學習值。在Fuzzy SOM本身的推導與實作等三方面，是找出五種BGA缺陷的聚類情況，以檢測出半導體具何種缺陷。至於線上(on-line)的即時(real-time)檢測，則仍有待日後在實際的環境下繼續實施。
並列摘要
Recently, the R&D of semiconductors has been regarded as one of the focused industries under the government policy instructions in Taiwan. The semiconductor industry, however, is with the strongest competitiveness in the world; consequently, its automation manufacturing process becomes the major factor to win the competitiveness. BGA is one of such the important processes. In this thesis, the novel fuzzy self-organizing map (SOM) neural network system is developed to automate the identification process of the defects occurred on BGAs. The fuzzy neural system contains three components: pre-processing for BGA data extraction, BGA data representation and fuzzy SOM learning and identification. In the course of the pre-processing, the bar charts of gray values are thresholded to binary values by a statistical image data analysis, then projected onto an X-Y coordinate where the Y-axis stands for the sum of the binary values and the X-axis means the X-coordinate value on the image. From the bar chart of the binaries of sum, the mean, variance, skewness and kurtosis values are statistically computed to represented the data features input to the fuzzy SOM system. This system then learns the training set of the vectors that each has the four data features of a column of the grid balls on a piece of BGA contained six grid ball columns, a prescribed standard BGA piece. A set of real BGAs and four classes of BGA defects are experimented and compared with a normal BGA. The results of the experiments show the proposed fuzzy SOM has a high accuracy rate of the defect identification.
參考文獻
【32】 林信成， 彭啟峰， 1994，Oh! Fuzzy， 第三波文化事業股份有限公司。
【1】 Blank, L., 1980, Statistical Procedures for Engineering, Management, and Science, McGraw-Hill.
【2】 Chandrasekaran, V., and L. Z. Qiang, 1998, "Topology constraint free fuzzy gated neural networks for pattern recognition", IEEE Transactions on Neural Networks, vol.9, no.3, pp.483-503.
【3】 Chen, M. and D. A. Linkens, 1997, "Neuro-expert approach to characteristics recognition for process monitoring and performance evaluation", International Journal of Intelligent Systems, vol.12, pp.359-377.
【4】 Chou, P. B., A. R. Rao, M. C. Sturzenbecker, V. H. Brecher, 1997, "Automatic defect classification for semiconductor manufacturing", Machine Vision and Applications,vol.9.