Machine-learning-enabled optimization and online monitoring for efficient and high-quality smart drying

Drying is an important process in the food industry that plays a critical role in both food production and preservation. Industrial scale drying processes and systems involve multiple interacting process parameters, conflicting production objectives, and highly uncertain sample characteristics, which make process control extremely challenging. Current industrial practice lacks the necessary decision-making tools to simultaneously achieve high energy efficiency and food quality. To address these challenges, this dissertation develops a suite of machine-learning-based process control tools to enable smart drying with improved process efficiency and product quality. The contributions of this dissertation are summarized as follows. It is important to devise a drying strategy to optimize drying efficiency, energy consumption, and product quality, especially under intricate input-output relationships with process uncertainties. Chapter 2 develops an uncertainty-aware, machine-learning-based response surface methodology for apple drying. New drying experiments are designed to resemble industrial practice with variable slice thickness. Variable-response relationships are modeled using machine learning models; Monte Carlo simulations are applied to quantify process uncertainties; and a constrained optimization approach identifies feasible design spaces and optimal parameter combinations. The proposed method achieves a 17.9% energy savings and a 19.0% reduction in drying time. Physical phenomena in drying can be measured by heterogeneous data modalities, with each carrying unique and complementary information. Effectively leveraging multi-modal data is essential for improving the performance of predictive modeling but remains challenging. Chapter 3 develops a multi-modal data fusion framework for accurately predicting final moisture content in apple drying. Tabular data and high-dimensional images are integrated through an encoder-decoder network to capture both process conditions and sample variability. Experimental results demonstrate predictive accuracy improvements of 19.3%, 24.2%, and 15.2% compared to tabular-only, image-only, and standard data fusion models, respectively. It is also shown that the proposed method is robust to varying modality ratios and can effectively capture process variabilities. Accurate real-time forecasting of the drying readiness (the optimal drying endpoint) is crucial for minimizing energy consumption and ensuring product quality. Chapter 4 presents a multi-modal fusion framework for online cookie drying readiness prediction. The model integrates in-situ video data and tabular process parameters using modality-specific encoders and a transformer-based decoder. The proposed model achieves a 15-second average prediction error, outperforming the state-of-the-art method by 65.7%, while balancing accuracy, model size, and efficiency. The framework is extensible to various other modality fusion tasks for effective online monitoring. Dynamic changes in food attributes during drying directly reflect product quality, and accurately predicting the trajectories of these attributes provides valuable insights into determining optimal drying conditions. Chapter 5 develops a data-driven approach for zero-shot prediction of surface color trajectories during food drying. The method learns component function parameters to represent color evolution under unseen conditions, with DCT preprocessing and enhanced by multi-modal data fusion and similarity- informed training selection. The method is validated on two case studies: cookie and apple drying, significantly outperforming baseline models by 93.2% and 87.30%, respectively.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Mechanical Engineering

Grantor

University of Illinois Urbana-Champaign

Year dc:date

2025

Author dc:creator

• Li, Shichen

Contributors dc:contributor

• Shao, Chenhui
• Ferreira, Placid Matthew
• Salapaka, Srinivasa M
• Wang, Pingfeng

• machine learning
• data-driven modeling
• smart drying
• response surface methodology
• multi-modal data fusion
• online monitoring
• zero-shot learning

Statement dc:rights

• Copyright 2025 Shichen Li All Rights Reserved

Language dc:language

en

Handle dc:identifier

https://hdl.handle.net/2142/132453

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/132453