Deep Learning Based Particle Image Velocimetry Technology And Its Application

Particle Image Velocimetry(PIV)is an importance whole-field flow measurement technique without intrusive probes,and has been applied to many areas,including life science,material science,environmental science,enegy science,industrial and agricultural production.In particular,it is applied to a wide range of flow measurements,including vortex formation in artificial heart valves,the flow over an aircraft wing in a supersonic wind tunnel,etc.The measurement results can be utilized to validate the CFD simulations or explore the complex flow phenomena.However,the reliability and accuracy of PIV measurement are prone to deterioration due to multiple causing factors,including the non-uniform particles distribution,improper laser lighting configuration,poor imaging recording/synchronizing system and misconfigured processing algorithms.As a result,a practical measurement often contains more than 5% outliers,and cannot provide the small scale flow structure which is smaller than the interrogation window.Thus,the study of PIV measurement error model is of great importance to fully understand the system error,progressive error and gross error.Based on the error analysis,a cascaded deep convolutional neural network framework(PIV-DCNN)was proposed to progressively regress the velocity vector.Different from the uniform motion assumption of cross-correlation method,PIV-DCNN learn the mapping function from large amount of synthetic training data.As a result,the system error from uniform assumption has been significantly reduced,especially in the case of complex turbulent flow.Regarding the progressive error of spatial gradient tensor of flow velocity field,a GT nueral network(GTnet)was proposed to extract the gradient tensor directly from particle images instead of traditional differential operation on the biased velocity measurement.Thus,the bias of velocity has little influence on the GT estimation due to the absence of differential procedure.From the outlier statistical model of PIV measurement,a Beyesian maximum a posteriori(MAP)was used to construct a novel outlier detection and vector field reconstruction algorithm.Due to the nonuniform nature of natural flow,the global noise variance is replaced in our modified vector field correction method(MVFC)by a locally estimated noise variance to improve the detection accuracy.Finally,a software package was devoleped with proposed algorithms,and has been applied to our PIV measurement instrument.Totally,the main innovative efforts in this dissertation are summarized as follow:1.A cascaded deep convolutional neural network framework for PIV analysis(PIV-DCNN)is proposed to extract fine flow structures of complex turbulent flow,and it benefits from the complex learnable model instead of the traditional manual model assumption.Based on large amount of synthetic training data,the trained PIV-DCNN model is a highly nonlinear function which can deal with complex flow with satisfactory accuracy.In order to improve the performance of single network,several networks are cascaded in a coarse-to-fine manner,and the final prediction is the average of three parallel network results.Outlier replacement and symmetric window offset operation glue them together.As a result,the PIV-DCNN could significantly reduce the systematical bias error and random error due to the novel gross structure and strong deep model capability,and the spatial resolution of PIV analysis has been improved considerably.2.A convolutional neural network(Gradient Tensor Network,or GTnet)is proposed to gaplessly estimate the differential variables from two center roughly aligned particle image patches.The GTnet prediction is thus independent of the biased velocity and differential operator,resulting in the decrese of progressive error.The accuracy of the GTnet was confirmed by comparing the results with the central difference operation using the same synthetic particle image.And experimental results indicated the advantages of GTnet over vector patterns with small flow structure.3.An adaptive spatial variable threshold outlier detection algorithm for raw gridded particle image velocimetry data using a locally estimated noise variance is proposed.This method is an iterative procedure with a reference vector field reconstruction step and an outlier detection step.A modified outlier detector is constructed using the locally estimated noise variance,and thus,the spatial variable threshold motivation is achieved.It turns out that a spatial variable threshold is preferable to a single spatial constant threshold in complicated flows such as vortex flows or turbulent flows.And experimental results validated the advantages of MVFC over large number of synthetic and real flow patterns.4.A software package for PIV analysis is developed,which integrates the above researches.And the algorithms are tested on this platform using the experimental particle images from our PIV instrument.