(305e) Reconstructing the Topology of Complex Process Systems

Chemical process systems consist of a multitude of interacting components. These interactions produce complex dynamics and dependencies that can be physical (e.g., pump pressure, heat transfer) or chemical (e.g., chemical reactions, thermodynamic constraints). Furthermore, process integration, intensification, and new advances in industrial equipment consistently increase the scale and complexity of chemical process systems. Although first-principle models can be used to describe these interactions, such models can be challenging to derive, validate, and use for control and optimization purposes as the scale and complexity of the process increases. In addition, aiming towards increased automation and ultimately autonomy in industrial operations, motivates pursuing methods for automatically identifying and accounting for such complex interactions.

In this study we address the problem of reconstructing the pattern of causal interactions among variables of chemical process systems based on process data. Adopting a graph theory perspective, where a process system is represented as a graph with nodes (or vertices) corresponding to process variables and edges corresponding to dynamic interactions, this problem is also known as network topology identification. A potential instance of application is the automated identification of the topology of a chemical process, as a first step towards process identification and control.

We focus on causal discovery methods that are purely data-driven and non-parametric. Thus, these methods can be broadly applied to many process systems and do not require prior knowledge of their dynamics [1,2]. Furthermore, unlike many standard system identification procedures, these causal discovery methods do not require independent manipulation of inputs [3,4,5]. This enables us to perform topological identification based on operations at a steady state, making data collection simple and efficient for real-world processes. Similar, methods have found use in characterizing faults and improving reliability for chemical process systems [6, 7].

We explore the application of a suite of causal inference methods derived from the fields of information theory, Bayesian networks, and signal processing [2,4,5,8,9]. We consider a chemical process network comprising multiple reactors, a separation unit, and a recycle between the reaction and separation section. The time series data used is generated by simulating a stochastic differential equation model of the process via Euler-Maruyama numerical integration. We benchmark the performance of these methods on this data set and discuss the benefits and limitations of each approach.

References

[1]. Glymour, Clark, Kun Zhang, and Peter Spirtes. "Review of causal discovery methods based on graphical models." Frontiers in genetics 10 (2019): 524.

[2]. Schreiber, Thomas. "Measuring information transfer." Physical review letters 85.2 (2000): 461.

[3]. Ljung, Lennart. System identification. Birkhäuser Boston, 1998.

[4]. Spirtes, Peter, et al. Causation, prediction, and search. MIT press, 2000.

[5]. Pearl, Judea. Causality. Cambridge university press, 2009.

[6]. Bauer, Margret, et al. "Finding the direction of disturbance propagation in a chemical process using transfer entropy." IEEE transactions on control systems technology 15.1 (2006): 12-21.

[7]. Yang, Shu, and B. Wayne Bequette. "Observational process data analytics using causal inference." AIChE Journal (2022): e17986.

[8]. Materassi, Donatello, and Murti V. Salapaka. "Reconstruction of directed acyclic networks of dynamical systems." 2013 American Control Conference. IEEE, 2013.

[9]. Materassi, Donatello, and Murti V. Salapaka. "On the problem of reconstructing an unknown topology via locality properties of the wiener filter." IEEE transactions on automatic control 57.7 (2012): 1765-1777.