Wavelet Transform Based Nonlinear Prediction of Signals

Estimating signals from time series is a common task in many domains of science and has been addressed for a long time by specialists. Predicting a signal from recorded time series remains however a very specific task, a great challenge. The wavelet transform provides multi-resolution analysis and allows accurate time-frequency localization of different signal properties. This paper presents a nonlinear prediction method implemented on artificial neural network based learning structure. From a discrete wavelet transform provided tree structure, specific coefficients are obtained and predicted with the already mentioned method, the reconstruction of signal is carried out using these new coefficients. The predicted signal is compared with the original one through parameters as the absolute mean error, using different analyzing functions and different learning structures. To evaluate the prediction, noise of different levels is added and the absolute mean error is recomputed and compared after every prediction.