Order Estimation of Computational Models for Dynamic Systems with Application to Biomedical Signals

Parametric models, in particular Autoregressive Moving Average (ARMA) models and their affiliates, are widely used in computational models of biomedical signals to fit a model to a recorded time series. An important step in this system identification process is the estimation of the model order. This paper provides the results of a systematic study of a previously developed technique based on the eigenvalues of the data covariance matrix to estimate the order of univariate ARMA models. A modified model order selection criterion which gives more robust results is used and the effect of the pole-zero positions on the correctly identified model orders is highlighted. Furthermore, the approach is extended to allow for the model order estimation of univariate Autoregressive (AR) and Moving Average (MA) models.