A high-resolution velocity field predicted method in near wall boundary layer by CNN
-
摘要: 本文设计并验证了基于卷积神经网络的边界层近壁流动高分辨率平均速度场预测方法:首先采用示踪粒子图像对数据集训练卷积神经网络,通过调整神经网络参数可以预测示踪粒子在数据集上的平均跨帧位移;然后使用该卷积神经网络预测像素空间中各像素位置的单粒子位移,得到高分辨率的平均速度场信息。将该方法用于预测湍流脉动较小的边界层近壁区的平均流动,能够将空间分辨率提高到单像素精度。误差分析发现,该方法获得的测速精度略优于传统单像素系综平均互相关算法,且对粒子浓度和示踪粒子图像对数目的要求明显低于后者。Abstract: A high-resolution velocity field prediction method in the boundary layer based on CNN-PTV is designed and verified in this paper. This method includes two stages, training process and prediction process. In the training process, a CNN model is trained by the particle image pairs. By optimizing the parameters, the exact ensemble particle movements are predicted by the CNN model. In the prediction process, a synthetic image including just a single particle is imported to the CNN model to estimate the flow information at this particular pixel space. Thus, a high-resolution velocity field is predicted. Comparing with the single-pixel ensemble correlation method, the CNN-PTV method has a higher precision. And the results of CNN-PTV method is insensitive to the frame numbers and particle density.
-
Key words:
- particle images /
- CNN /
- boundary layer
-
表 1 所测试CNN的不同配置参数及损失
Table 1. The parameters set in various tested cases
工况
(Cases)可训练参数
(Trainable parameters)批大小
(Batch size)学习率衰减
(Learning rate decay)卷积核大小
(Conv kernel size)最大卷积深度
(Deepset channel)损失
(Loss)Case 1 2567274 8 1×10–2 3,3,3,2,2,2 512 215 Case 2 2567274 8 1×10–2 3,3,3,2,2,3 512 115 Case 3 2567274 8 1×10–2 3,3,3,2,2,3 512 100 Case 4 2567274 8 1×10–2 3,3,3,2,2,3 512 103 Case 5 4980190 8 5×10–3 3,3,3,2,2,5 512 70 Case 6 9016659 8 1×10–2 7,7,5,5,3,3 512 61 Case 7 9016659 8 5×10–3 7,7,5,5,3,4 512 63 Case 8 25927463 8 5×10–3 9,7,7,5,5,3,2 1024 42 Case 9 25927464 16 5×10–3 9,7,7,5,5,3,3 1024 45 Case 10 25927464 24 5×10–3 9,7,7,5,5,3,3 1024 43 表 2 人工合成虚拟粒子图像对的参数设置
Table 2. List of particle image parameters
参数 物理意义 数值 X×Y 图像尺寸 64 像素×128 像素 dp 平均粒子直径 4 像素 σ(dp) 粒子直径标准差 20% ρp 粒子浓度,32 像素×32 像素窗口内粒子数 12.8 N 粒子图像对总数 2000 Zl 激光片光厚度 1 mm ΔZp 粒子在z方向的位移 0.2 mm σnoise CCD高斯噪声 5% -
[1] WINTER K G. An outline of the techniques available for the measurement of skin friction in turbulent boundary layers[J]. Progress in Aerospace Sciences,1979,18:1-57. doi: 10.1016/0376-0421(77)90002-1 [2] 赵荣娟,吕治国,黄军,等. 基于压电敏感元件的摩阻天平设计[J]. 空气动力学学报,2018,36(4):555-560.ZHAO R J,LYU Z G,HUANG J,et al. Design of skin friction balance based on piezoelectric ceramics[J]. Acta Aerodynamica Sinica,2018,36(4):555-560. [3] NGUYEN C V,WELLS J C. Direct measurement of fluid velocity gradients at a wall by PIV image processing with stereo reconstruction[J]. Journal of Visualization,2006,9(2):199-208. doi: 10.1007/BF03181763 [4] WILLERT C E. High-speed particle image velocimetry for the efficient measurement of turbulence statistics[J]. Experiments in Fluids,2015,56(1):1-17. doi: 10.1007/s00348-014-1892-4 [5] 申俊琦,王建杰,潘翀. 平板湍流边界层瞬时摩擦阻力的光学测量和统计分析[J]. 气体物理,2020,5(5):13-23.SHEN J Q,WANG J J,PAN C. Optical measurement and statistical analysis of instantaneous wall-shear stress in a turbulent boundary layer[J]. Physics of Gases,2020,5(5):13-23. [6] SHEN J Q,PAN C,WANG J J. Accurate measurement of wall skin friction by single-pixel ensemble correlation[J]. Science China Physics, Mechanics & Astronomy,2014,57(7):1352-1362. doi: 10.1007/s11433-014-5462-9 [7] NIE M Y,PAN C,WANG J J,et al. A hybrid 3D particle matching algorithm based on ant colony optimization[J]. Experiments in Fluids,2021,62(4):1-17. doi: 10.1007/s00348-021-03160-4 [8] DOSOVITSKIY A, FISCHER P, ILG E, et al. FlowNet: learning optical flow with convolutional networks[C]//Proc of the 2015 IEEE International Conference on Computer Vision. 2015: 2758-2766. doi: 10.1109/ICCV.2015.316 [9] ILG E, MAYER N, SAIKIA T, et al. FlowNet 2.0: evolution of optical flow estimation with deep networks[C]//Proc of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. 2017: 1647-1655. doi: 10.1109/CVPR.2017.179 [10] CAI S Z,ZHOU S C,XU C,et al. Dense motion estimation of particle images via a convolutional neural network[J]. Experiments in Fluids,2019,60(4):1-16. doi: 10.1007/s00348-019-2717-2 [11] LAGEMANN C,LAGEMANN K,MUKHERJEE S,et al. Deep recurrent optical flow learning for particle image velocimetry data[J]. Nature Machine Intelligence,2021,3(7):641-651. doi: 10.1038/s42256-021-00369-0 [12] LECORDIER B, WESTERWEEL J. The EUROPIV synthetic image generator (S. I. G.)[C]//Proc of Proceedings of the Particle Image Velocimetry: Recent Improvements. 2004. [13] WANG L W,PAN C,LIU J H,et al. Ratio-cut background removal method and its application in near-wall PTV measurement of a turbulent boundary layer[J]. Measurement Science and Technology,2021,32(2):025302. doi: 10.1088/1361-6501/abb483