Bayesian synergistic filter for online model class learning and model updating

In this work, a novel Bayesian synergistic filter (BSF) for online model class learning and model updating is presented. Most existing parametric model updating approaches are conducted by employing a prescribed model class. The model updating performance relies on the configuration of this prescribed model class. In practice, it is a challenging task to construct a model class that is sufficient but not unnecessarily overcomplicated. Moreover, the deficiencies between the prescribed model class and the actual structure remain throughout the analysis. To address these issues, the proposed BSF reconfigures the model classes and performs the model updating in a recursive manner. At each time step, the plausibilities of the model classes are computed and they serve as cues to trigger the self-calibration learning of the model classes. Consequently, the deficiencies of model classes can be corrected as the filter propagates. The calibratable model classes provide the crucial baseline for reliable model updating. The proposed BSF is applicable for tracing the structural response and estimating the model parameters of both linear and nonlinear time-varying dynamical systems using incomplete noisy response measurements.