A competitive wavelet network for signal clustering.

Abstract

This correspondence proposes a novel signal clustering method based on the unsupervised training of a wavelet network. The synaptic weights are parameterized by wavelet basis functions, which are adjusted by a competitive algorithm that makes use of the neighborhood concept proposed by Kohonen. The robustness of the wavelet network with respect to noise is illustrated in a simulated problem, in which dynamic systems are grouped on the basis of their step responses. An example involving clustering of electrocardiographic signals taken from the MIT-BIH database is also presented. In this case, the ability of the proposed network to perform clustering at successive resolution levels is illustrated. The possibility of interpreting the information encoded in the network at the end of training is also discussed.