兩種半導體封裝檢驗類神經網路模型之績效比較

A performance analysis of two artificial neural network models for semiconductor package inspection.

蕭東岳(Tung-Yueh Hsiao)

指導教授 ： 項衛中

中原大學/電機資訊學院/工業與系統工程研究所/碩士(2021年)

https://doi.org/10.6840/cycu202100494

摘要

在半導體封裝產業中，隨著封裝元件等製造技術不斷的縮小發展，使封裝晶片在檢驗時品質管控與效率上受到的考驗也隨之增加。透過自動化檢測與神經網路類別分類方法，能夠降低在檢驗封裝晶片時大量影像資料的檢驗時間，輔助傳統人工視覺判別缺陷的方法，以降低視覺判別上的個別差異性與檢驗錯誤。本研究比較與探討在半導體封裝檢驗製程中，將晶片影像資料輸入影像處理後，運用倒傳遞神經網路及卷積神經網路進行晶片缺陷判斷，依據判斷結果分析兩種類神經網路的績效。
本研究流程包含三個階段，第一階段主要是影像資料辨識與影像前處理，使用待檢驗晶片的影像資料，透過尋找影像上相似的特徵並加以裁切為單顆晶片影像；第二階段為單顆晶片影像的資料處理，依照正常與缺陷資訊來標記影像；第三階段為模型訓練與驗證，依四項指標:準確率、損失函數以及混淆矩陣中的偽陽性與偽陰性來比較兩種類神經網路模型績效。研究結果發現在資料有限的情況下，使用卷積神經網路訓練晶片影像資料能夠獲得比倒傳遞神經網路更好的四項指標；且若影像資料複雜度與精細度較高，則兩種類神經網路模型皆能夠建立出較為簡單的神經網路架構，以及較少的試驗次數即可得到好的預測結果。

關鍵字

半導體封裝晶片 ； 影像辨識 ； 倒傳遞神經網路 ； 卷積神經網路

並列摘要

In the semiconductor packaging industry, with the continuous development of packaging components and other manufacturing technologies, the requirements of quality and efficiency for the packaged chip inspection have also increased. Automatic inspection and neural network classification methods were applied to reduce the inspection time of image data for the packaged chips. They can help the manual vision inspection method to identify defects, and reduce the individual difference and human errors. In this research, the back propagation neural network and convolution neural network were used to determine the chip defects with the chip image data in the semiconductor packaging inspection process, and the performance of two types of neural networks was analyzed based on the testing results.
A three-stage process was developed. The first stage is image data recognition and image pre-processing. The image data of the whole substrate were used to find similar features on the chips and were cut into several single-chip images. The second stage is to label the chip images, according to pre-defined information, as normal or defective chips. The third stage is training and validating the back propagation neural network and convolution neural network models to classify the chips. The classification results of two kinds of neural network models were compared. They are four performance indicators: valid accuracy, valid loss, false positive and false negative rate. This research found that in the case of limited data, the convolution neural network model can obtain better results in four indicators than the back propagation neural network model. For higher complexity and fineness of the image data, simper neural network architectures and better performance of indicators with less trials can be obtained for both types of neural network models.

並列關鍵字

Semiconductor Package Chip ； Image Recognition ； Back-Propagation Neural Network ； Convolutional Neural Network

參考文獻

[1] 工研院產業學習網。全文網址：

https://college.itri.org.tw/event/525-smart-machinery2020.html

[2] SEMI 2018：台灣半導體產業2021年產值估達3兆元。全文網址：

http://iknow.stpi.narl.org.tw/post/Read.aspx?PostID=14787

[3] Chao-Ton Su, Taho Yang and Chir-Mour Ke, "A neural-network approach for semiconductor wafer post-sawing inspection," IEEE Transactions on Semiconductor Manufacturing, May 2002, vol. 15, no. 2, pp. 260-266.