System Identification through Model Composition and Stochastic Search

Abstract

System identification methodologies are useful for identifying characteristics of structural systems using measurement data. However, incorrect systems might be identified when many combinations of system characteristics result in the same predicted responses at measured locations. The reliability of identification is affected by a number of factors that most previous work has overlooked. This paper presents a system identification methodology that explicitly treats factors that affect the success of identification. Rather than simply determining parametric values, this methodology also involves identification of model characteristics including boundary conditions. Due to inevitable modeling errors, models that provide absolute minimum differences between predictions and measurements are rarely correct models. In such situations, the challenge is to define a population of candidate models that result in such differences being below threshold values that are determined by the magnitude of modeling errors. The methodology is illustrated using a case study in civil engineering. This work contributes to providing engineers with general strategies to meet interpretation challenges associated with sensor data.