(145b) Challenges in Applying Data Analysis to Industrial Processing: Inverse Problem Viewpoint | AIChE

(145b) Challenges in Applying Data Analysis to Industrial Processing: Inverse Problem Viewpoint


Sturnfield, J. - Presenter, The Dow Chemical Company
Trahan, D. W., The Dow Chemical Company
Shuang, B., Dow
Production data analysis is often a regular part of the evaluation and troubleshooting for operations. This is especially true when looking at issues of quality, conversion rates, production rates, or rates of degradation in the process. Ideally, it would be valuable to understand what is actually happening inside the production units. In particular, production problems would be easier to address if it could be determined where and when in the process significant changes are occurring. This can be considered an inverse problem of determining the process configuration that gives the measurements seen.

The underlying physics can provide insights to data analysis and is probably necessary to solve the “Inverse Problem”. For example, the physics of a fixed bed reactor will produce relationships about the flows, pressure drop, temperature, heat generated (absorb), and stream composition changes. Applying these physics in an industrial setting will require numerous assumptions about the process. There may be competing assumptions that need to be considered. For the reactor example, these assumptions will be capture in parameters that designate the flow resistance and reaction kinetics. The competing assumptions will generate different model structures and possibly different types of parameters. The analysis of the production data will drive refinements of these parameters.

This presentation examines the challenges and tools for analyzing production data and for determining the appropriate physic driven models that can capture the behavior of the production units. This will include approaches for doing time series analysis, uncertainty quantification, principle component analysis, machine learning, objective function determination and data structure analysis. In particular, we will look at how these tools can provide different insights into the underlying production behavior.