Efficient algorithms for Volterra system identification (Q2732896)

The identification task of discrete-time finite-dimensional (polynomial) symmetric Volterra models is considered in the linear regression and least squares framework. It is shown that the normal equations for such a Volterra system driven by a zero mean Gaussian process possess a unique solution if and only if the spectral density of the input does not vanish at least at \(m\) distinct frequencies, where \(m\) is the memory length of the system. A systematic account of the multichannel embedding approach is provided, and a mapping of a single-input single-output Volterra model to an equivalent multi-input single-output multichannel linear model is proposed. Using this representation, efficient algorithms for the least squares estimation of the Volterra kernel values are derived for batch and adaptive processing. Also, various order-recursive schemes for the structure determination of the unknown system are discussed, and a new algorithm for hierarchically structured Volterra models is proposed. The presented techniques are illustrated by means of simulation examples.