Variational Inference for Nonlinear Structural ‎Identification

Research interest in predictive modeling within the structural engineering community has recently been focused on Bayesian inference methods, with particular emphasis on analytical and sampling approaches. In this study, we explore variational inference, a relatively unknown class of Bayesian inference approaches which has potential to realize the computational speed, accuracy, and scalability necessary for structural health monitoring applications. We apply this method to the predictive modeling of a simulated Bouc-Wen system subject to base vibration and compare the performance of this inference approach to that of the unscented Kalman filter. From this investigation, we find that though variational inference is more computationally intensive than the unscented Kalman filter, it exhibits superior performance and flexibility.

Keywords:

system identification , predictive modeling , Bayesian inference , unscented Kalman filter , nonlinear systems