Particle image velocimetry analysis with simultaneous uncertainty quantification using Bayesian neural networks

TL;DR

This paper introduces a Bayesian convolutional neural network approach for particle image velocimetry that provides simultaneous flow field estimation and uncertainty quantification, outperforming traditional CNN methods.

Contribution

It is the first application of Bayesian neural networks to PIV, enhancing flow analysis with uncertainty estimates and demonstrating improved accuracy using cross-correlation maps.

Findings

BCNNs with cross-correlation inputs outperform those with interrogation windows

The best BCNN captures 100% of true displacements within 95% confidence intervals

BCNNs can be adapted to multi-pass PIV with moderate accuracy loss

Abstract

Particle image velocimetry (PIV) is an effective tool in experimental fluid mechanics to extract flow fields from images. Recently, convolutional neural networks (CNNs) have been used to perform PIV analysis with accuracy on par with classical methods. Here we extend the use of CNNs to analyze PIV data while providing simultaneous uncertainty quantification on the inferred flow field. The method we apply in this paper is a Bayesian convolutional neural network (BCNN) which learns distributions of the CNN weights through variational Bayes. We compare the performance of three different BCNN models. The first network estimates flow velocity from image interrogation regions only. Our second model learns to infer velocity from both the image interrogation regions and interrogation region cross-correlation maps. Finally, our best performing network derives velocities from interrogation region cross-correlation maps only. We find that BCNNs using interrogation region cross-correlation maps as inputs perform better than those using interrogation windows only as inputs and discuss reasons why this may be the case. Additionally, we test the best performing BCNN on a full test image pair, showing that 100% of true particle displacements can be captured within its 95% confidence interval. Finally, we show that BCNNs can be generalized to be used with multi-pass PIV algorithms with a moderate loss in accuracy, which may be overcome by future work on finetuning and training schemes. To our knowledge, this is the first effort to use Bayesian neural networks to perform particle image velocimetry.