|dc.description.abstract||Smart manufacturing (SM) is a new paradigm of manufacturing that leverages rapid advance in information technologies, industrial Internet-of-Things to improve the flexibility and adaptability of manufacturing, which is expected to fuel the next industrial revolution. The increasingly available data is one of the major driving forces of SM, where data-driven modeling plays a central role in harnessing the power of the advanced infrastructure. However, developing data-driven models for SM applications poses unique challenges for SM practitioners. This dissertation discussed three data-driven modeling projects.The first project develops a fault prediction model for a steel manufacturing process using casting and steelmaking data. Due to the limitations in instrumentation and the complexity of the underlying mechanisms, the data generating process is highly uncertain and time-varying. To address this challenge, this project incorporated mechanistic knowledge including process flowsheet, chemistry, and transport phenomena into the data-driven modeling process. Specifically, the upstream steelmaking data is used to develop the predictive model, while the casting data is used to develop an auto-labeler for the steelmaking data. The resulting model can predict faults 10 minutes in advance.
The second project develops a methodology to identify critical process parameters using observational process data. Because the majority of data collected from manufacturing processes are observational, they are intrinsically insufficient to provide causal information, which is necessary for active tasks such as process control and optimization. To address this challenge, this project explored using causal inference to bridge the gap between data-driven methodologies and causal objectives. Specifically, this project discussed a seemingly paradoxical result where the fault prediction model developed in the steelmaking project is used to guide process improvement. Then this project proposed a metric to identify critical process parameters using causal inference, which is illustrated using a simulation case study.
The third project explores using nonlinear data-driven models for model-based control. Due to the nonconvexity of many nonlinear data-driven models, using them for optimal control and optimization can lead to suboptimal solutions and increased computation expense. To address this challenge, this project explores using input convex neural networks (ICNNs) for model-based control, which guarantees the control problems to be convex optimization problems. Specifically, this dissertation presented different formulations of ICNNs including convolutional ICNNs, Bayesian last layer ICNNs, and partial ICNNs. Additionally, this project discussed the integration of ICNNs in a model predictive control framework. Through two simulation case studies, the unique benefits of the proposed method are shown.||