Identifying the Time of Step Change and the Source of Mean Shift in Multivariate Process Using Convolutional Neural Networks

碩士 === 國立成功大學 === 工業與資訊管理學系 === 107 === Detecting and identifying the mean shift in the manufacturing process has always been an important issue. An effective detection method can help engineers figure out the root causes of the shift so as to improve or restore the underlying manufacture process. I...

Full description

Bibliographic Details
Main Authors: Kai-ChunChien, 簡愷均
Other Authors: Tai-Yue Wang
Format: Others
Language:zh-TW
Published: 2019
Online Access:http://ndltd.ncl.edu.tw/handle/drxa5g
Description
Summary:碩士 === 國立成功大學 === 工業與資訊管理學系 === 107 === Detecting and identifying the mean shift in the manufacturing process has always been an important issue. An effective detection method can help engineers figure out the root causes of the shift so as to improve or restore the underlying manufacture process. If one has to monitor multiple quality characteristics simultaneously, the multivariate control chart can be used as an effective monitoring tool. The MEWMA control chart is one type of multivariate control chart, which considers previous data, and performs better than Hotelling’s T2 control chart when dealing with small shifts. Therefore, we take MEWMA statistics as our training feature. In addition, control charts can just detect the shift but can not tell more shift information. In this research, we have constructed a monitoring process that includes two models, both of which utilize neural networks, which are known for having excellent learning abilities. The monitoring process can not only detect the mean shift, but also provide us with the quantification of the shift. In this study, we have decided to use convolution neural networks, which are able to extract the effective features in two-dimensional data. So we combine the raw data with MEWMA statistics as our input vector. There is a total of two phases in the monitoring process. Firstly, we work to detect the mean shift. If it is determined that a shift exists, we subsequently perform additional tests in acquire the shift quantification of each quality characteristic. We use ARL and accuracy to evaluate the first and the second model, then we perform a sensitivity analysis by appointing different MEWMA weights, window sizes, and correlation coefficients. Our results show that our way of combining raw data and MEWMA statistics as our training features, and by using CNN in extracting the features, is more effective compared to using SPC as well as the previous machine learning methods. The advantage also stands in the identification of extra information regarding the shift quantification.