以大數據分析建置刀具異常磨耗偵測系統

Applying Big Data Analysis to Construct an Abnormal Wear Detection System for Dicing Tools

呂英軒(Ying-Hsuan Lu)

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(2020年)

https://doi.org/10.6840/cycu202000308

摘要

在半導體封裝製程中，晶圓切割是直接影響產品良率的重要步驟，透過預測方法提前找到將發生磨耗異常的製程就可以有效避免物料成本與停機所造成的損失。然而現今的半導體封裝廠在封裝製程中有不同晶圓切割機，每臺晶圓切割機也會因不同客戶需求而執行不同切割程式，產生了由不同機台與切割程式所造成的製程差異。本研究探討能否由少量的初始加工數據就能推估製程是否將發生異常，以及使用了三種預測模式來驗證不同機台與不同切割程式間是否具有預測模式的通用性。
本研究為了達到快速偵錯與預測模式評斷的目的，首先透過具有時間序列特性的刀具磨耗資料萃取其自迴歸特徵值，再以支持向量機、決策樹、隨機森林、類神經網路建置刀具異常磨耗偵測系統，並以平均準確度當作預測模式的評斷標準。研究結果指出在資料有限且同機台同切割程式的情況下，使用資料量少的預測模式可以普遍達到78%至98%的平均準確度，若要達到更高的平均準確度則需使用不同的預測模式；資料量多的預測模式在不同機台或不同切割程式的情況下通用，其平均準確度可以達92%以上。

關鍵字

自迴歸分析；支持向量機；決策樹；隨機森林；類神經網路

並列摘要

In the semiconductor packaging process, wafer dicing is an important process that directly affects product yield. Using predictive methods to find a process, that may cause abnormal wear in advance, could effectively avoid material scrap and machine idling cost. However, a semiconductor packaging plant might have different wafer cutting machines, and each wafer cutting machine executed various programs for different customer needs. Different machines and different cutting programs would cause different process results. This study explores the possibility of predicting whether the process will be abnormal by a small amount of initial processing data. Three prediction modes are proposed to find if a single prediction mode is effective for different cutting machines and different cutting programs.
A procedure was applied for the purpose of rapid error detection and prediction mode evaluation. First, using the autoregressive method to fit the tool wear data with time series parameters. The tool abnormal wear detection system was designed with Support Vector Machine, Decision Tree, Random Forest, and Artificial Neural Network methods. Average prediction accuracy rate of these methods was used as the performance index of the prediction mode.
The results of the study indicated that the prediction model with small amount of data can achieve 78% to 98% average accuracy rate for the same cutting program and same cutting machine cases. For higher average prediction accuracy rate, different prediction modes need to be used. Once the machine or cutting program is not the same one, prediction models using more data can be applied for those cases, and those models can achieve at least 92% average prediction accuracy rate.

並列關鍵字

Autoregression ；Support Vector Machine；Decision Tree；Random Forest ；Artificial Neural Network

參考文獻

Box, G. E. and Jenking, G. M. (1976), Time Series Analysis: Forecasting and Control, Revised, San Francisco: Holden Day.

Birant D, Comparison of decision tree algorithms for predicting potential air pollutant emissions with data mining models , Journal of Environmental Informatics ,17(1):46-53, 2011

De Myttenaere, A.; Golden, B.; Le Grand, B.; Rossi, F. Mean absolute percentage error for regression models. Neurocomputing 2016,192, 38–48.

Groover, M. P. (2007). Fundamentals of Modern Manufacturing： Materials Processes, and Systems： John Wiley Sons.

Huang, P. T. (2001). A neural networks-based in-process adaptive surface roughness control (NN-IASRC) system in end-milling operations. (Doctoral dissertation), Iowa State University, US.