| description abstract | Although multiple-model structural identification (MM ST-ID) approaches appear to offer clear, conceptual benefits over single-model approaches, they have not yet been employed within a transparent scenario that will allow quantitative comparison, critique, and refinement. To fill this gap, the research reported in this paper aimed to (1) implement and compare current MM ST-ID approaches on a physical laboratory model to establish their accuracy and identify their merits and shortcomings, and (2) identify the ability to refine MM ST-ID methods by weighing observations based on their correlation with the desired predictions. The scenario implemented used modal parameters as the observations, and static displacements and strains as the desired predictions. The various MM ST-ID methods were evaluated based on how their prediction distributions agreed with the actual responses of the physical model. The results indicated that while all methods were successful in bounding the actual responses, the Bayesian updating approach proved to be the most efficient in terms of required number of simulations, and was able to produce prediction distributions with the smallest bounds (while still incorporating all measured responses). In addition, the mean of the MM ST-ID prediction distributions did not coincide with the model that had the largest weight (i.e., the highest likelihood), which indicates that single model approaches not only are unable to provide estimates of variability, but may produce biased predictions. Finally, through a second set of scenarios, the research reported in this paper showed how prediction distributions may be improved by weighing observations based on their correlation with the desired predictions. | |
