Persons and affiliations

Date

Description

The monitoring capability of a chemical process is highly dependent on the chosen
sensor configuration. Hence, the decision between different sensor systems under
given constraints is crucial for a chemical plant’s operation. A common goal is to
achieve the best trade-off between spent sensor cost and obtained process
information. Often, the value of information is derived from a rigorous equation system
describing the process, for instance via the state prediction error covariance matrix or
Fischer information matrix [1]. While these approaches are well established, they have
shortcomings, in particular when it comes to the task of sensor selection in brownfield
chemical plants. Namely, linear system dynamics, Gaussian sensor uncertainties that
are independent, and a completely rigorous description of the process are common
assumptions that are hardly met by real-world, large-scale brownfield chemical plants.
In our contribution, we propose an approach to assess the information value of new
sensors that mitigates aforementioned shortcomings. The method applies ideas from
uncertainty – and sensitivity analysis [2] to determine the optimal sensor configuration
under constraints of prediction precision and equipment cost. Variables that are not
measured are modeled via distributions, expressing prior knowledge about the
parameters; poorly understood process parts are expressed via data-driven models.
The presented approach is demonstrated on a CSTR-reactor and compared to the
results of [1], in terms of modeling efforts and determined optimal sensor configuration.

[1] K. R. Muske und C. Georgakis, „Optimal measurement system design for chemical
processes“, Aiche Journal, Bd. 49, Nr. 6, S. 1488–1494, June 2003
[2] F. Pianosi, K. Beven, J Freer, J. W. Hall, J Rougier, D. B. Stephenson, T. Wagener,
„Sensitivity analysis of environmental models: A systematic review with practical
workflow“, Environmental Modelling & Software, Bd. 79, S. 214–232, May 2016