Date of Award

Fall 12-15-2023

Level of Access Assigned by Author

Open-Access Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Chemical Engineering

Advisor

Douglas Bousfield

Second Committee Member

Andre Khalil

Third Committee Member

Yifeng Zhu

Additional Committee Members

Peter Stechlinski

Martin Holmvall

Abstract

Pulp mills and papermakers require careful control of input raw materials. The paper pulp composition, consisting of blends of different wood fiber types, affects multiple final product properties in interacting ways and impacts process operating conditions. Manual estimation of composition by classification and counting using microscopy is time consuming, repetitive, error-prone, and fibers are not always identifiable. Using a dataset of 359,840 fibers from 12,690 images of either hardwood or softwood fibers from 423 microscopy slides with data partitioned into 60% training, 20% validation, and 20% testing splits by slide, and a sequence of principal components analysis, Gaussian mixture, image analysis, and convolutional neural network models this work demonstrates a system capable of processing 4.92 megapixel microscopy images with 3 color channels at a rate of 30 seconds per image using a 4gb Nvidia Jetson Nano computer with a fiber-segment level test accuracy of 91%. The variation in accuracy between slides is statistically significant and follows a beta-binomial distribution, which controls the required number of slides for confident estimation of actual process mixture composition; the described implementation requires 10 slides for a 90% interval of ±3.25% of the estimated composition. Additionally, anomalous cotton fibers, not present in training data, are correctly identified with a rate of 33% false negatives and 5% false positives. The entire process is visualized, enhancing interpretability, and understanding of fundamental fiber structures. The complete system enables papermakers and pulp mills to improve control of the input concentrations of component fibers and appropriately adjust corresponding operating conditions to achieve desired properties. Studying the classification results, we the identify the influence of confounding factors in our data; changing confounding factors from one slide to the next influences not only the species of fiber, but also the observation conditions, such as illumination, imaging, and slide preparation. Then, by simulating a dataset of microscopy slides, in which the influence of such confounders is not present, we demonstrate that it is not the simplicity of the objects of interest that limits the use of high capacity models for learning, but hypothesize the presence of an easily learnable feature that varies from slide to slide and is detectable among many objects from the same slide. Mitigating this feature could greatly improve learning of otherwise relevant but subtle fiber features.

Files over 10MB may be slow to open. For best results, right-click and select "save as..."

Share