Velocity Estimation in the Shear Layer of a M_j=0.6 Jet Using Deep Neural Networks

In this presentation, we recast the Linear Stochastic Estimation method to take advantage of the predictive power of Deep Neural Networks (DNNs), and offer a quantitative comparison to traditional Linear Stochastic Estimation (LSE). High frame-rate Particle Image Velocimetry (PIV) (10kHz) was used to measure the flow-field of a Mach 0.6 axisymmetric jet. Time-series of the stream-wise and cross-stream components of velocity were recorded at 11 locations, spanning 0.1 to 0.75 diameters in the radial direction, simulating the output of crosswires at those locations. A reduced order model of the turbulent velocity fluctuations in the shear layer is then produced using Proper Orthogonal Decomposition (POD). The resulting model is used to train both a DNN and an LSE model to predict the velocity within the shear layer, given a subset of the raw velocity measurements. We show that, on average, the DNN is able to predict the velocity fluctuations more accurately than LSE, because of the non-linear relationship between the conditional and unconditional events.