Publication Date

2023

Document Type

Dissertation

Committee Members

Zifeng Yang, Ph.D. (Advisor); George Huang, Ph.D. (Committee Member); Philippe Sucosky, Ph.D. (Committee Member); Hamed Attariani, Ph.D. (Committee Member); Bryan Ludwig, M.D. (Committee Member)

Degree Name

Doctor of Philosophy (PhD)

Abstract

In fluid flow experiments, there have been numerous techniques developed over the years to measure velocity. Most popular techniques are non-intrusive such as particle image velocimetry (PIV), but these techniques are not suitable for all applications. For instance, PIV cannot be used in examining in-vivo measurements since the laser is not able to penetrate through the patient, which is why medical applications typically use X-rays. However, the images obtained from X-rays, in particular digital subtraction angiography, are projective images which compress 3D flow features onto a 2D image. Therefore, when intensity techniques, such as optical flow method (OFM), are applied to these images the accuracy of the velocity measurements suffer from 3D effects. To understand the error introduced in using projective images, a vertical square tube chamber was constructed to achieve various water flow rates with variable dye injection points to perform dye visualization velocimetry (DVV). The results from DVV were compared with PIV measurements to quantify the error associated with DVV. Results from DVV were comparable with PIV, but a machine learning correction method, more specifically multilayer perceptron (MLP), was needed to adjust the DVV results. To train the MLP model, CFD simulations were conducted to generate detailed velocity distributions in the tube and projected dye images which would be used for DVV analysis and thus used as input for training. These CFD simulations were compared with PIV measurements and dye visualization images to validate proper boundary conditions and meshing. For the laminar case, MLP reduces the error associated with DVV from 35% down to 6.9%. When MLP was used to correct instantaneous DVV measurements for the turbulence cases, the error decreased from 22% to 9.8% for measurements 20 mm downstream of the dye inlet. For a time-averaged turbulent case, MLP was able to decrease the v-velocity error down to 5% and reduce the error of DVV by 50% for shear stress near the dye inlet.

Page Count

202

Department or Program

Ph.D. in Engineering

Year Degree Awarded

2023

ORCID ID

0000-0001-6641-8651


Included in

Engineering Commons

Share

COinS