A Novel Signal Modeling Method Using the Wavelet Transform.
Stavros A. Karkanis, D.A. Karras, B.G. Mertzios · 1998
This paper presents a novel methodology for improving the modeling of non-smooth signals, based on the wavelet transform and neural network function approximation techniques. It is suggested that the enhanced time frequency localization properties of the wavelet transform could provide the means for extracting highly informative features for signal modeling tasks, especially for very spiky signals. These features are used as inputs to neural network nonlinear regression models of the MLP (MultiLayer Perceptron) type in addition to the usually used current and past signal values organized in sliding windows. While the majority of contemporary research efforts to improve the generalization capability of function approximation and classification techniques, like neural networks, employ the reduction of the number of effective parameters, it is demonstrated here that significant improvements in the generalization capability could be achieved by introducing additive informative features. Although our methodology involves an increase of the input space dimensionality it is exhibited, through extensive experimentation, that it leads to significantly better signal modeling of non-smooth peaked signals.