In modern society, complex fluids such as polymer melts, colloidal suspensions, slurries, liquid crystals, and emulsions are widely used in continuous processes for high value industries. These fluids, which are typically non-Newtonian, exhibit comple...
In modern society, complex fluids such as polymer melts, colloidal suspensions, slurries, liquid crystals, and emulsions are widely used in continuous processes for high value industries. These fluids, which are typically non-Newtonian, exhibit complex rheological behaviors including shear thinning, shear thickening, yield stress, viscoelasticity, normal stresses, and thixotropy. As a result, it is significantly more challenging to determine optimal processing conditions for these complex fluids compared to simple Newtonian fluids like water or oil.
Furthermore, in real industrial processes, unintended variations in raw material grade or composition can occur, and the internal structure and resulting rheological properties of the fluids may change during processing. These dynamic changes complicate process control, potentially leading to reduced product quality, increased defect rates, and ultimately resulting in economic inefficiency and environmental issues.
This dissertation proposes a real-time characterization approach for complex fluids during processing by applying machine learning (ML) to sensor signals. The central hypothesis is that fluids with distinct rheological properties generate unique sensor signals, and that these signals can be learned and classified using ML techniques.
To validate the proposed approach, two experimental systems were investigated. First, lithium-ion battery anode slurries with varying internal structures and compositions were circulated through a pipe system. Pressure and flow rate signals were collected and classified using an echo state network, a type of recurrent neural network. The effects of different preprocessing methods, diagnostic time, and sensor types on model performance were also examined.
Second, pressure signals were acquired during the extrusion of polymer blends with systematically varied compositions. These signals were analyzed using MultiRocket-Hydra, a convolutional neural network, to perform regression and classification tasks. The suitability of both tasks for real-world industrial scenarios was confirmed, and the impact of diagnostic time on model performance was evaluated.
This study is the first to demonstrate that complex fluids can be effectively diagnosed using a purely data-driven approach based on sensor signals. It is also expected to serve as a valuable early-stage diagnostic framework for controlling and optimizing complex fluid processes across a wide range of industrial applications.