Time-varying quadratic model selection using wavelet packets

Model selection and system identification for cases where the model is required to have both characteristics of time-variance and nonlinearity is considered. To enable identification from a single input/output observation record, the time-variation is approximated by a weighted sum of orthogonal sequences. Wavelet packets are chosen for these sequences and an adapted basis for each time-varying coefficient is selected via the best basis algorithm. Individual wavelet packets are then selected via a multiple hypothesis test which determines those packets that are significant to each approximation, and which may be discarded from the model.