Subband neural networks for noisy signal forecasting and missing data reconstruction

A subband multirate architecture is presented for data signal prediction and reconstruction. The architecture is based on a uniform filter bank that divides the input signal into many narrow band signals each of them predictable using a low order neural network. The main advantage of this approach consists in the extension of the forecast horizon using a low complexity network. In the case of missing data, this approach allows the reconstruction of a long data sequence from forward-backward predicted samples. Due to the multirate processing the subband networks input consists of a decimated version of the input signal. It follows a great reduction of the convergence time. Moreover, in presence of additive input noise, we can observe an improvement of the generalization performances. The experimental tests, conducted using noisy artificial nonlinear chaotic series, show a comparison between fullband and 4-channels subbands approaches involved in prediction and reconstruction tasks.