Robust identification of time-varying system dynamics with non-white inputs and output noise

We developed a new technique to identify time-varying system dynamics from an ensemble of input-output realizations. With this new approach, a matrix equation is generated for every sampling time using input autocorrelation and input-output crosscorrelation functions estimated across the ensemble, and a pseudoinverse is used to solve for the impulse response function (IRF). The technique was tested on a simulated time-varying system using various combinations of input signal bandwidth, output signal-to-noise ratio (SNR), and number of realizations. Simulation results showed that our pseudoinverse approach outperforms a previous least-squares method when the input is strongly coloured and the SNR is low. The new technique is also more efficient computationally.