Tensorial Data Modeling and Analysis for Manufacturing Process Quality Control Using Sensor Data

Tensorial data modeling and analysis;Industrial and Operations Engineering;Engineering;Industrial & Operations Engineering

Zerehsaz, YaserPlumlee, Matthew ;

University of Michigan

关键词: Tensorial data modeling and analysis;    Industrial and Operations Engineering;    Engineering;    Industrial & Operations Engineering;   

Others  :  https://deepblue.lib.umich.edu/bitstream/handle/2027.42/138580/yzereh_1.pdf?sequence=1&isAllowed=y

瑞士|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

One of the main engines driving the new era of industrial data analytics is advanced sensing technologies and distributed computing. Although a vast amount of data can be garnered with the aid of these sensing and computing technologies, several challenges aggravating the data analysis arise at the same time. One important question to pose is how to develop the new data analysis and modeling methodologies that can be applied to the new set of complex datasets containing functional and tensorial data as instances moving away from the conventional data structures.The objective of this dissertation is to study and address three major challenges typically occurring when working with complex multi-stream and functional sensing data. First, while the classical data analysis tools are designed to work with single-sensor data represented by matrices or vectors, sometimes the multisensory data are more efficiently represented by high-order arrays to preserve the original structure of the data. A dataset is said to have a multi-stream structure, meaning that it contains more than two informative dimensions.In Chapter 2 of this dissertation, a high-order-based monitoring method is suggested for monitoring a tensorial dataset of tool wear measurements. In addition to achieving high monitoring performance, the developed tensor-based chart is capable of providing correlation pattern analysis. This is useful in discerning assignable causes of unusual patterns of tool wear to enhance the process diagnosis ability. The second challenge addressed in this dissertation is to deal with outlier observations that are inevitable when collecting sensor data. Chapter 3 develops a new robust decomposition method that can handle the outlier observations. The critical point in this study is that the proposed method is the first decomposition technique considering both correlated noise components and outlier observations. The correlated noise can be seen in time series, longitudinal and image data. The chapter starts with developing a robust low-order matrix decomposition method, and it proceeds to extending the concepts and mathematical formulas to a high-order tensor setting. The method is applied to a dataset with the purpose of surface defect monitoring using real billet images taken from a hot rolling process. The third challenge studied in this dissertation is how to find a mathematical relationship between a quality response variable and some process variables that are not necessarily scalar or functional. Indeed, some of the predictors are in the form of tensors in addition to the regular functional or scalar predictors. The challenge is how to estimate the parameters of such general regression models. Chapter 4 focuses on developing a flexible yet parsimonious model for predicting scalar response variable utilizing some tensorial and functional predictors. The developed model is called functional linear regression with tensorial predictor (FLRTP). The advantage of this methodology compared to classical functional data analysis and linear regression methods is that it can handle both functional and tensorial predictors without performing vectorization on the tensorial predictors. This is helpful since the multi-stream structure of the predictor is preserved and the number of parameters to be estimated is kept at a reasonable amount. The performance of all methods is evaluated using simulation and real-world studies.