Abstract
Particle Imaging Velocimetry (PIV), which is the whole flow field measurement technique using visualized images, has been recognized to be essential and very useful for analyzing two-or three-dimensional complex flow fields. Although lots of measurement methods for PIV have been reported, the PIV is clasified to the pattern tracking and the particle tracking methods in principle. The pattern tracking method such as the correlation method and the gray level difference method gives some erroneous vectors because of mismatching a tracer image pattern to another one. While velocity vector profiles measured by the particle tracking method such as the four-step PTV (Particle Tracking Velocimetry), the spring-model method, and the binary image correlation method, partly lack information on velocity vectors because no tracer particles exist in some parts of the flow field. It is required, therefore, to estimate a correct velocity vector field from the measured vectors with erroneous ones or to estimate velocity vectors in the unmeasurable parts. One promising approach for the estimation is to make an appropriate model of the field by using the measured information.
This paper proposes a new method using artificial neural network for estimating the whole flow velocity vector fields from measured velocity vectors. The neural network is trained by using measured velocity vectors as teaching data so that the derivatives of a certain scalar function agree well with the measured data. The continuity equation of flow is consequently satisfied in the estimated vector fields and the scalar function gives the stream function. The merit of this method is that measured data on velocity vectors are automatically corrected and the estimated data satisfy the continuity equation of flow. The pattern of streamlines is additionally obtained.
The method is applied to the PIV standard images developed based on a calculated velocity field that is available through the Internet (http://sap.gen.u-tokyo.ac.jp) in order to evaluate the effectiveness and acuracy. It is proved that the method gives much correcter velocity vector distributions than the measured ones.
