Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


38020.pdf (3.05 MB)

ETD Abstract Container

Abstract Header

A Data Augmentation Methodology for Class-imbalanced Image Processing in Prognostic and Health Management

Yang, Shaojie
http://rave.ohiolink.edu/etdc/view?acc_num=ucin161375046654683

Abstract Details

2020, MS, University of Cincinnati, Engineering and Applied Science: Mechanical Engineering.
Machine vision is commonly used in the field of prognostics and health management (PHM) for industrial applications, including defect reduction, robot assistant, quality inspection, safe work environment, reading characters, and packing inspection. It provides degradation information on parts that are too small to be seen by the human eye, and can also help classify the root cause of the fault without damaging parts or installing additional sensors. Many deep learning-based models have been studied for machine vision-based applications, but few studies focus on imbalanced data issues, and most studies are based on data that has a good amount of diversity. Modern deep-learning-based models require class-balanced data to avoid overfitting. In many industrial applications, it is difficult to collect class-balanced data to train deep learning models. For that reason, choosing the appropriate method to significantly increase the diversity of data available for industrial applications is critical for training modern deep learning-based models. Therefore, this research focuses on developing a data augmentation methodology for deep learning-based fault diagnosis modeling, to improve the quantity and diversity of class-imbalanced data without actually collecting new data. In this thesis, a cross-class data augmentation approach using convolutional autoencoder latent space interpolation is proposed for industrial image processing applications to overcome the presents of class-imbalanced dataset. During training stage, a latent space augmentation model is constructed into the traditional geometric transformation image augmentation approach. New samples are synthesized by interpolating extracted features from convolutional autoencoders. The developed approach has been validated using degradation assessment of cutting wheel and wafer map failure pattern recognition. And proposed method is benchmarked with geometric transformations and generative adversarial networks (GAN) using convolutional neural network models.
Jay Lee, Ph.D. (Committee Chair)
Jay Kim, Ph.D. (Committee Member)
Manish Kumar, Ph.D. (Committee Member)
76 p.
Mechanical Engineering
Data Augmentation; Image Processing; Machine Vision; Deep Learning; Prognostics and Health Management

Recommended Citations

Citations

  • Yang, S. (2020). A Data Augmentation Methodology for Class-imbalanced Image Processing in Prognostic and Health Management [Master's thesis, University of Cincinnati]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=ucin161375046654683

    APA Style (7th edition)

  • Yang, Shaojie. A Data Augmentation Methodology for Class-imbalanced Image Processing in Prognostic and Health Management. 2020. University of Cincinnati, Master's thesis. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=ucin161375046654683.

    MLA Style (8th edition)

  • Yang, Shaojie. "A Data Augmentation Methodology for Class-imbalanced Image Processing in Prognostic and Health Management." Master's thesis, University of Cincinnati, 2020. http://rave.ohiolink.edu/etdc/view?acc_num=ucin161375046654683

    Chicago Manual of Style (17th edition)

ucin161375046654683
401
© 2020, some rights reserved.Creative Commons License A Data Augmentation Methodology for Class-imbalanced Image Processing in Prognostic and Health Management by Shaojie Yang is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License. Based on a work at etd.ohiolink.edu.
This open access ETD is published by University of Cincinnati and OhioLINK.