Accurate weight coefficient estimation of multi-camera light field PIV through backpropagation neural network

Abstract

Volumetric velocity measurement through multi-camera light field particle image velocimetry (LF-PIV) requires an accurate estimation of the weight coefficient (WC) of three-dimensional (3D) tracer particle distribution reconstruction. To achieve that, this study proposes a calibration method based on a backpropagation neural network (BP-NN) for the WC estimation of the multi-camera LF-PIV. The BP-NN model establishes a mapping relationship between the spatial voxels and pixels of the multi-cameras. The proposed method is compared with the direct ray tracing (DRT) method and it shows that the proposed method provides an accurate estimation of the WC. It also does not depend on the prior knowledge of angle separations of the multi-cameras as is required for the DRT method. The proposed method is initially evaluated by conducting synthetic tests of ring vortex field reconstruction and further verified by conducting experiments on a low-swirl injector (LSI) flow. Results show that the root mean square error of the ring vortex displacement field can be reduced from 0.71 voxels to 0.35 voxels by the proposed method. The relative errors of LSI flow axial and radial velocity components are smaller than 10%. Therefore, it demonstrates that the 3D flow velocity can be measured accurately by the multi-camera LF-PIV by incorporating the proposed BP-NN calibration method.