A novel variational Bayesian deep learning framework for reconstructing the particle image velocity field of pump-jet propulsor

Abstract

This study establishes a novel super-resolution (SR) framework for reconstructing the low resolution (LR) velocity field of pump-jet propulsor (PJP) obtained by particle image velocimetry (PIV) method, which combines down-sampled skip-connection/multi-scale (DSC/MS) models and variational Bayesian (VB) theory to form the VB-DSC/MS model. The PIV method is can obtain the spatial-temporal velocity field of PJP, while the PIV method consumes expensive costs and a lot of time. The VB-DSC/MS method can investigate the real information of PIV velocity field and spatial-temporal prior knowledge to establish nonlinear relationships, and which will obtain the SR flow field. The investigate uses velocity field data obtained by improved delayed detached eddy simulation (IDDES) and PIV methods as data-sets, which learned the accuracy and uncertainty distribution of reconstructing the axial wake flow field at different conditions using VB-DSC/MS method. The results show that the SR method has a higher accuracy to reconstruct the contour and wake evolution of LR velocity field obtained by PIV method. It can accurately reconstruct the fluid acceleration region, hub low-velocity region, and mixed flow region of PIV field, and which the improvement factor can reach to 256. Although there is a higher uncertainty distribution at a larger scaling factors, all reconstructed velocity field data are within a 95% confidence interval.