Theses and Dissertations

Advisor

Marufuzzaman, Mohammad

Committee Member

Ma, Junfeng

Committee Member

Wang, Haifeng

Committee Member

Da Silva, Bruno

Committee Member

Tanger, Shaun M.

Date of Degree

5-16-2025

Original embargo terms

Visible MSU Only 1 year

Document Type

Dissertation - Campus Access Only

Major

Industrial and Systems Engineering

Degree Name

Doctor of Philosophy (Ph.D.)

College

James Worth Bagley College of Engineering

Department

Department of Industrial and Systems Engineering

Abstract

In this study we propose a deep learning method to optimize the classification of wood chip moisture content levels using the Vision Transformer and then ultimately increase the classification performance by creating synthetic images using the diffusion transformer model. In the first chapter of our study, we complete a detailed explanation of how the moisture content levels of 10 different wood chips were gathered ranging from 2 to 50$\%$. This chapter serves as a foundation for subsequent sections, illustrating the challenges associated with the current data collection process, which is both time-consuming and inefficient. Accurately determining moisture content for wood chips is significant in different industries such as pellet mills, bio refineries, and paper mills, and for that reason, in the second chapter of our study, we concentrate on developing an advanced computer vision model named MoistViT. To achieve the MoistViT, we complete an extensive evaluation of fourteen different Vision Transformer models and supplement them with the Bayesian Optimization Hyperband method to achieve superior performance when classifying moisture content levels. In our third chapter, we concentrate on creating high-definition synthetic wood chip images while using the computational limitations faced by many researchers when training advanced models. The importance of creating the generated wood chip images comes from the fact that current data-driven methods used to measure the moisture content of wood chips need a large amount of data to train the models to eventually achieve state-of-the-art results and the current method used to create the wood chip images is destructive and time-consuming. Therefore, we propose a diffusion transformer that is used to generate images in a fraction of the time needed by current methods. The diffusion model obtains strong results, achieving high-quality synthetic images that are then used to augment existing smaller datasets to train computer vision classification models for real-time industrial applications. Finally, the last two chapters present a detailed analysis of the results achieved by both proposed models, along with a comprehensive explanation of their development. With these explanations, we aim to improve model interpretability, mitigating the "black box" nature often associated with deep learning methods.

Sponsorship (Optional)

This work is supported by the Sustainable Bioeconomy through Biobased Products and Engineering for Agricultural Production and Processing programs, project award no. 2020- 67019-30772 and 2022-67022-37861, from the U.S. Department of Agriculture’s National Institute of Food and Agriculture.

Share

COinS