Identification of vibrating structures subject to non stationary excitation : A non stationary stochastic realization problem

The vector signal delivered by accelerometers measuring the vibrations of a structure subject to a non stationary excitation may be modelled as a Gauss-Markov signal with constant state-transition and observation matrices, but non stationary excitation noise. It is shown that some stochastic realization algorithms remain robust even used on a single (non stationary) sample of the signal ; we developp some algorithms which are recursive in the order of the model and we investigate the effect of model reduction.