Author ORCID Identifier

https://orcid.org/0009-0005-8503-7519

Semester

Summer

Date of Graduation

2024

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Chemical and Biomedical Engineering

Committee Chair

Dr. Debangsu Bhattacharyya

Committee Co-Chair

Dr. Srinivas Palanki

Committee Member

Dr. Srinivas Palanki

Committee Member

Dr. Xin Li

Committee Member

Dr. Nagasree Garapati

Committee Member

Dr. Yuhe Tian

Abstract

Abstract

Development of Probabilistic Dynamic Model Building and Bayesian Machine Learning Approaches

Samuel Adeyemo

The recent years have seen a tremendous increase in the use of artificial intelligence (AI) and machine learning (ML) for the development of data-driven mathematical models needed for performing real-time optimization, model-based control, performance optimization, dynamic data reconciliation, and process performance monitoring. However, the development of data-driven models is faced with some challenges including lack of model interpretability, sensitivity of algorithm to noise in training data, limited extrapolation capabilities and violation of conservation laws. Drawing motivation from these existing gaps, this work aims to develop robust algorithms for constructing interpretable predictive data-driven models from noisy process data. Novel strategies are developed to enhance model extrapolation capabilities by guaranteeing the satisfaction of conservation laws for both steady state and dynamic models.

By carrying out Bayesian inferencing in the Expectation-Maximization framework, a data-driven approach that simultaneously estimates the noise in training data and their possible correlation while estimating the posterior probability distribution of model parameters conditioned on the available data is developed. An algorithm for model selection is proposed using branch and bound seeking a parsimonious model structure by promoting sparsity and penalizing redundancy in model parameters. The resulting Bayesian Identification of Dynamic Sparse Algebraic Model (BIDSAM) algorithm incorporates model parameter estimability measures in the model selection criterion minimizing the computational cost of the algorithm.

To guarantee the conservation of mass by model predictions, two novel algorithms are developed for both static and dynamic models. The first algorithm involves a sequential approach for solving for model parameters and carrying out data reconciliation while the second approach exploits the model structure in the proposed BIDSAM algorithm to enforce equality constraints on the model parameters such that model predictions are guaranteed to satisfy mass and energy balances.

Finally, this work draws cue from the computational neuroscience literature where predictive coding (PC) is used in the training of networks modeling the Bayesian brain. In this work, sparse hierarchical models are constructed for systems with (spatio)temporal distributions and local computations are engaged for parameter estimation. This results in hierarchical BIDSAM (H-BIDSAM) algorithm which automatically detects unknown input delays and is robust to presence of noise with cross- and autocorrelations, in the training data.

Embargo Reason

Publication Pending

Available for download on Thursday, July 31, 2025

Share

COinS