Using Adaptive Artificial Neural Networks for Reconstructing Irregularly Sampled Laser Doppler Velocimetry Signals

Abstract

This paper is concerned with the problem of analyzing turbulent flow signals that are irregularly sampled by a laser Doppler velocimeter. The temporal irregularity of the sampling is the main problem addressed due to the difficulties it introduces in the use of traditional analysis techniques. The main contribution of this paper is to assess the adequateness of introducing a signal-modeling strategy by means of temporal delay-based artificial neural networks (ANNs) that are adaptively and continuously trained online on the irregularly sampled real data using an evolutionary-based strategy called the multilevel Darwinist brain. The networks are used to model the signal and thus are able to produce a regularly sampled signal straight from the irregular one. It is important to note that they are being trained to model the signal and not just interpolate it; they are able to generate spectral signatures that are very close to that of the original signal