Development of a methodology for hybrid metamodeling of hierarchical manufacturing systems within a simulation framework

Scope and method of study. This research effort was directed at conceptualizing and validating a methodology to create "models of models" (i.e., metamodels) for use in situations where aggregate behavior of some system sub-models can be substituted for detailed behavior. The metamodel is used in concert with a detailed representation of other system sub-models. Through use of the developed methodology, a workcenter within a multi-workcenter manufacturing model can be represented in metamodel form while preserving the time-in-system distribution of parts moving through the workcenter. Metamodels are developed based on statistics collected from observation of their detailed counterparts during controlled experimentation. The statistics are used to estimate a set of empirical time-in-system cumulative distribution functions for the workcenter metamodel. The set comprises curves for several values of workcenter stage utilization. Metamodels are validated by statistically comparing experimentally generated metamodel distributions of time-in-system with detailed model distributions of time-in-system.;Findings and conclusions. Metamodels were created for six different workcenter structure/operating characteristic scenarios. These metamodels were compared to their detailed counterparts under two different decision making cases at two different utilization levels. Two statistical tests and a pragmatic test were formulated to analyze the consistency of results. The first statistical test was based on the hypothesis that if the detailed model resulted in a significant difference in mean time-in-system across the decision alternatives the metamodel should also. The results of this test were entirely consistent. The second statistical test was based upon the hypothesis that the interaction effect between decision alternative and model type should not be significant. The results of this test were mixed but predominantly undesirable. The third test was designed to measure the pragmatic impact of the interaction in absolute terms. A statistic was formulated to measure the percentage difference between the achieved results and results calculated in the absence of an interaction effect. The results were very satisfying with a worst case error of 5.35%.