390392 Determination of Direct Causal Relationships Among Variables in Process Systems

Monday, November 17, 2014: 2:43 PM
403 (Hilton Atlanta)
Taha Mohseni Ahooyi1, Jeffrey E. Arbogast2, Warren D. Seider3, Ulku G. Oktem4 and Masoud Soroush1, (1)Department of Chemical and Biological Engineering, Drexel University, Philadelphia, PA, (2)Process Control & Logistics, Air Liquide, Newark, DE, (3)Department of Chemical and Biomolecular Engineering, University of Pennsylvania, Philadelphia, PA, (4)Risk Management and Decision Center, Wharton School,University of Pennsylvania, Philadelphia, PA

There is an increasing demand for systematic methods to determine the causal relationships among variables in complex systems. These causal relationships allow for developing complex models for use in the design, optimization and control of the systems, and detecting the hazard propagation pathways needed in risk management. Causal relation determination can be made using expert knowledge, tentative mathematical models, and/or historical data [1]. In the past decade, the research community has focused particularly on the last approach. Besides various heuristic and hybrid methods [2], the existing data-based causal relationship determination methods are mainly of two types: score and search methods, and conditional independence tests [3].  Despite many advantages of these existing methods, finding directed causal structures with higher accuracy and lower computational cost is still an open problem.

In this work, a novel data-based method of determining direct causal relationships among variables of a process is presented. Of particular interest is to overcome the challenges imposed by different types of cyclic causalities present in modern industrial process systems.  The cyclic causalities  are mainly due to control and process bidirectional relationships. A new measure of conditional independence that finds the skeleton of undirected structure and a new technique that directs causal arrows and has the capability to capture highly non-linear interactions are introduced. Such a directed causal structure is also useful to identify hidden variables within a system. The application and performance of the proposed method in detecting and diagnosing  process system abnormal operation and propagation of hazard will be shown by simulating several process examples.


[1] Neapolitan, R. E. Learning Bayesian Networks. Series in Artificial Intelligence. New York: Prentice Hall, 2004.

[2] Brown, L. E.; Tsamardinos, I.; Aliferis, C. F. A comparison of novel and state-of-the-art polynomial Bayesian network learning algorithms. In Proceedings of the Twentieth National Conference On Artificial Intelligence, Veloso, M. M.; Kambhampati, S. (eds). 2, AAAI Press, 739–745 (2005).

[3] Shaughnessy, P.; and Livingston, G. Evaluating the causal explanatory value of Bayesian network structure learning algorithms. Research paper, Department of Computer Science, University of Massachusetts Lowell (2005).

Extended Abstract: File Uploaded
See more of this Session: Advances in Data Analysis: Theory and Applications
See more of this Group/Topical: Computing and Systems Technology Division