On the Analysis of Turbulent Flow Signals by Artificial Neural Networks and Adaptive Techniques

Abstract

Artificial Neural Networks (ANNs) and evolution are applied to the analysis of turbulent signals. In a first instance, a new trainable delay based artificial neural network is used to analyze Hot Wire Anemometer (HW) signals obtained at different positions within the wake of a circular cylinder with Reynolds number values ranging from 2000 to 8000. Results show that these networks are capable of performing accurate short term predictions of the turbulent signal. In addition, the ANNs can be set in a long term prediction mode resulting in a sort of non linear filter able to extract the features having to do with the larger eddies and coherent structures. In a second stage these networks are used to reconstruct a regularly sampled signal straight from the irregularly sampled one provided by a Laser Doppler Anemometer (LDA). The irregular sampling dynamics of the LDA signals is governed by the arrival of the seeding particles, superimposing the already complex turbulent signal characteristics. To cope with this complexity, an evolutionary based strategy is used to perform an adaptive and continuous online training of the ANNs. This approach permits obtaining a regularly sampled signal not by interpolating the original one, as it is often done, but by modeling it.