Reconstruction of Flow Field With Missing Experimental Data of a Circular Cylinder Via Machine Learning Algorithm

Abstract

In this study, artificial neural networks (ANNs) have been implemented to recover missing data from the particle image velocimetry (PIV), providing quantitative measurements of velocity fields. Due to laser reflection or lower intensity of particles in the interrogation area, the reconstruction of erroneous velocity vectors is required. Therefore, the distribution of time-averaged and normalized flow characteristics around a circular cylinder has been demonstrated as streamwise and cross-stream velocities at Re = 8000. These velocity components have been given for different regions at x/D = 0.5, x/D = 1.25, x/D = 2, and y/D = 0. These stations have been chosen to estimate missing data for near-wake, mid-wake, far-wake, and symmetry regions. The missing data ratios (A*) for 0.5 <= x/D <= 2 are A* = 3.5%, 7%, and 10%. In addition, these values are A* = 4%, 8%, and 12% for y/D = 0, while A* = 7.5% for the shaded region. The increment of area positively affects the estimation results for near-wake and mid-wake regions. Moreover, the errors tend to decrease by moving away from the body. At y/D = 0, increasing the area negatively influences the prediction of the results. The mean velocity profiles of predicted and experimental data have also been compared. The missing data have been predicted with a maximum percentage error of 3.63% for horizontal stations. As a result, the ANN model has been recommended to reconstruct PIV data.