Methode zur Integration zeitvariabler Kantengewichte in kausale Graphen für präzisere Modellanpassung

Diagnosing deviations between a simulation model and the real behaviour of a process plant is often done with causal directed graphs, but these usually use static edge weights and so cover changing operating states, ageing or wear only to a limited extent. The paper presents a method for adding time-varying, data-based edge weights to an existing causal graph, which is itself derived from an AutomationML description of the plant.

Simulation data are segmented into discrete operating states by actuator positions; for each state a time-varying dynamic Bayesian network is built, and edge weights computed with the Kullback-Leibler divergence are fed back into the graph per state. On a Modelica model of a laboratory process plant with 81 potential fault causes and about 12,000 simulations, the mean rank of the correct fault cause improved by 9.0% and the scaled absolute error fell by 28.8% compared with a static heuristic.

Keywords: Kausale Graphen, Time-Varying Bayesian Networks, AML, Simulation, Werkerassistenz, Industrie 4.0

English summary written by AutomationML e.V.; the paper and its German abstract are open access via the DOI above. Listed as part of AutomationML at Academic Conferences 2026.

Felix Gehlhoff, Malte Ramonat, Moritz Wagenknecht, Tom Westermann: Methode zur Integration zeitvariabler Kantengewichte in kausale Graphen für präzisere Modellanpassung. In: Lisa Underberg (Hrsg.): EKA 2026, Entwurf komplexer Automatisierungssysteme, 19. Fachtagung, Magdeburg, 2026. doi: 10.25673/123331