Adaptive filtering for bandlimited noise using an error backpropagation neural network

Standard noise filtering schemes are often inadequate for eliminating noise energy without seriously distorting the desired signal. As an extension of previous work (see IEEE Trans. Instr. and Meas., vol.40, pp.820-825, 1991) the authors demonstrate the use of a feedforward backpropagation error correction neural network for the filtering of absorption spectra exhibiting the effects of bandlimited Gaussian noise. The noise is bandlimited to the same spectral region as the information bearing signal. The neural network is trained by using synthesized data sets as in the earlier work. The additive noise is generated by sampling a Gaussian distribution of random numbers such that the synthesized noise sequence added to the synthetic spectrum has a variance of 1. The noise sequence of Gaussian random deviates is further bandlimited to the Nyquist frequency dictated by the implied sampling rate.<<ETX>>