Degree
Doctor of Philosophy (PhD)
Department
Chemical Engineering
Document Type
Dissertation
Abstract
One of the limitations of conventional monitoring tools in crystallization is the inability to deal with high solids concentration. Image-based monitoring, in which RGB images of the solution are captured and analyzed by an object detection software, has been a promising alternative. The software used in these tools primarily depends on hand-coded heuristics to distinguish between the signal and the noise. With the recent success of supervised deep learning, a newer paradigm has emerged in which the heuristics can be learnt from labeled images. This approach is founded on the idea that it is easier to develop labels for data than it is to come up with complex heuristics. In this work, a new image-based sensor using deep learning-based Convolutional Neural Networks is proposed and tested.
Population balance models are used to keep track of the crystals during growth and these models can be optimized to generate optimal variable trajectories. Since models are not perfect, repeated optimization is required at runtime to keep the process under control, which can be hard depending on the complexity of the model. This work also introduces the use of Deep Reinforcement Learning to develop optimal control profiles offline. The DRL algorithm can be trained offline in the presence of process disturbances and deployed online without any additional training, thereby alleviating some of the problems associated with conventional control frameworks.
Date
5-3-2022
Recommended Citation
Manee, Vidhyadhar, "A Data-Based Framework for Monitoring and Controlling Particulate Systems" (2022). LSU Doctoral Dissertations. 5846.
https://repository.lsu.edu/gradschool_dissertations/5846
Committee Chair
Romagnoli, Jose
DOI
10.31390/gradschool_dissertations.5846