Dr. Sören Bergmann

Diskrete Simulation ist eine etablierte Methodik zur Untersuchung des dynamischen Verhaltens von komplexen Fertigungs- und Logistiksystemen. Konventionelle Simulationsstudien fokussieren auf einzelne Modellaspekte und spezifische Analysefragen. Der Umfang der ausgeführten Szenarien ist häufig gering. Das Konzept des Data-Farming verwendet das Simulationsmodell als Datengenerator für eine breite Skale von Experimenten und ermöglicht unter Nutzung von Data-Mining-Methoden eine wesentlich breitere Untersuchung des simulierten Systems sowie eine höhere Komplexität in den abgeleiteten Erkenntnissen. Anforderungen an Simulationssysteme und -modelle zur Durchführung von Data-Farming werden erläutert. Eine Erweiterung des Ansatzes ist die simulationsbasierte Robustheitsanalyse auf der Basis von Verlustfunktionen nach Taguchi. Beide Vorgehensweisen werden an einer Fallstudie aus dem Fahrzeugbau demonstriert.

Discrete event simulation is an established methodology for investigating the dynamic behavior of complex manufacturing and logistics systems. Traditionally, simulation experts conduct experiments for predetermined system specifications focusing on single model aspects and specific analysis questions. In addition to that, the concept of data farming and knowledge discovery is an ongoing research issue that consists of broad scale experimentation and data mining assisted analysis of massive simulation output data. As an extension to this approach, we propose a concept for investigating the robustness of complex manufacturing and logistic systems which are often very sensitive to variation and noise. Based on Taguchis loss function, we developed a concept including data farming and visual analytics methodologies to investigate sources of variation in a model and the factor values that make a configuration robust. The concept is demonstrated on an exemplary case study model.

Discrete-event simulation is a well-accepted method for planning, evaluating, and monitoring processes in production and logistics. To reduce time and effort spent on creating simulation models, automatic simulation model generation is an important area in modeling methodology research. When automatically generating a simulation model from existing data sources, the correct reproduction of dynamic behavior of the modeled system is a common challenge. One example is the representation of dispatching and scheduling strategies of production jobs. When generating a model automatically, the underlying rules for these strategies are typically unknown but yet have to be adequately emulated. In this paper, we summarize our work investigating the suitability of various data mining and supervised machine learning methods for emulating job scheduling decisions based on data obtained from production data acquisition. We report on the performance of the algorithms and give recommendations for their application, including suggestions for their integration in simulation systems.