Dual estimation of partially observed nonlinear structural systems: a particle filter approach

Dual estimation consists of tracking the whole state of partially observed systems, and simultaneously estimating unknown model parameters. In case of nonlinearly evolving systems, standard filtering procedures may provide unreliable model calibrations, either because of estimates affected by bias or due to diverging filter response. In this paper, we propose a particle filter (PF) wherein particles, i.e. system realizations evolving in a stochastic frame, are first sampled from the current probability density function of the system and then moved towards the region of high probability by an extended Kalman filter. We show that the proposed filter works much better than a standard PF, in terms of accuracy of the estimates and of computing time.