A Construction Method of Wavelet Network Based on Local Time-frequency Information

摘要

In this paper, a novel construction method of wavelet network (WN) is proposed. At first, the basis library is divided into some sub-library according to time-frequency information of local input samples. Furthermore, basis selection is repeated in the sub-library instead of the library so that the candidates are drastically decreased. Finally, the wavelet network is composed by some sub-wavelet networks (sub-WNs) and some auxiliary neural units which are designed to adaptively locate the right sub-WNs to process the given input. The proposed method has not only the advantage of being computationally less expensive than the conventional method but also the noteworthy points that can adaptively adjust the scale of sub-library through controlling the scale of local input. Both the theoretic analysis and the simulation results for the function learning show the effectiveness of the proposed method.