Particle filter scheme with mutation for the estimation of time-invariant parameters in structural health monitoring applications

In this study, a novel method is presented for non-linear, non-Gaussian online state and parameter identification, developed for use in structural health monitoring (SHM) problems. The algorithm consists of a particle filter (PF) that combines the use of the standard PF with mutation operators. The algorithm aims at alleviating the sample impoverishment problem, which is a well-known limitation of the standard PF, yielding it inefficient for demanding non-linear identification problems. To overcome this hurdle, we introduce here an alternative approach, influenced by the principles of evolutionary computation. After the standard PF steps are performed to a point where the sample diversity drops below some threshold, the unfit particles are replaced by either the fittest particles or the current weighted estimate of the state. Next, the time-invariant components of the particles are mutated under some mutation probability, and the new sample is then propagated to the next time step. This process is well suited for joint state and parameter estimation problems, as is usually the case in SHM techniques. As a result, the loss of diversity associated with the standard PF is overcome, and the new PF with mutation is shown to outperform the standard PF and the unscented Kalman filter for the case of high process noise. The method is validated through an established benchmark problem found in the literature, lying outside of the structural identification concept, and a previously referenced 3DOF structural system with hysteresis elaborating the SHM aspect. Copyright (c) 2012 John Wiley & Sons, Ltd.