Chinese Journal of Liquid Crystals and Displays, Volume. 40, Issue 7, 1023(2025)

Particle image velocimetry method based on ConvLSTM and LiteFlowNet architecture

Xin'ai LIU1, Juan MENG1、*, Hai DU2, and Zhiyuan LI1
Author Affiliations
  • 1College of Information Engineering,Dalian Ocean University,Dalian 116023,China
  • 2State Key Laboratory of Coastal and Offshore Engineering,Dalian University of Technology,Dalian 116023,China
  • show less
    Figures & Tables(15)
    Structure of the LiteFlowNet-CL network
    Structure of the CBAM
    Diagram of CBAM feature extraction
    Structure of the ConvLSTM
    Processing and analysis flow of LiteFlowNet-CL
    Examples of particle image pairs at different time instants
    Estimated velocity fields and streamlines for different models
    Estimated velocity fields and vorticity maps for different models
    Comparison of root mean square error(RMSE)results for different models
    RMSE error curves of different methods during the training process
    Outputs of LiteFlowNet-CL at different resolution levels
    • Table 0. [in Chinese]

      View table
      View in Article

      Table 0. [in Chinese]

      算法1 CBAM

      输入:输入特征图x(形状为C*H*W)。

      输出:经过处理后的特征图x

      (a)avg_pool=AdaptiveAvgPool2d(x);max_pool=AdaptiveMaxPool2d(x)。 /* 使用自适应池化*/

      (b)avg_out=fc2(relu1(fc1(avg_pool)));max_out=fc2(relu1(fc1(max_pool)))。 /* 全连接层 */

      (c)out=sigmoid(avg_out+max_out)。 // 合并池化结果并应用激活函数

      (d)x=x*out。//应用通道注意力权重

      (e)avg_out=mean(x,dim=1,keepdim=True);max_out=max(x,dim=1,keepdim=True)。  /* 计算每个通道的平均值和最大值 */

      (f)x=cat([avg_out,max_out],dim=1)。 // 将平均值和最大值在通道维度上拼接

      (g)x=conv1(x)。 // 利用卷积层处理特征图

      (h)x=sigmoid(x)。 // 应用激活函数

      (i) x=x*x+x。 // 应用空间注意力权重

    • Table 1. Description of the PIV dataset

      View table
      View in Article

      Table 1. Description of the PIV dataset

      种类参数条件数量
      均匀流(Uniform)|dx|∈[0,5]2 000
      反阶梯流(Back-step)

      Re=800

      Re=1 000

      Re=1 200

      Re=1 500

      600

      600

      1 000

      1 000

      圆柱绕流(Cylinder)

      Re=40

      Re=150

      Re=200

      Re=300

      Re=400

      500

      500

      500

      500

      500

      各项同性湍流(DNS-turbulence)无参数可描述4 000
      海洋表面流场(SQG)不同天数4 500
      JHTDB-channel无参数可描述2 000
      JHTIDB-mhd1024无参数可描述1 600
      JHTIDB-isotropic1024无参数可描述2 000
    • Table 2. RMSE of different flow fields

      View table
      View in Article

      Table 2. RMSE of different flow fields

      方 法Back-stepCylinderJHTDBSQGDNS-turbulenceUniform
      训练集测试集训练集测试集训练集测试集训练集测试集训练集测试集训练集测试集
      WIDIM-0.476-0.45-0.520-0.715-0.824-0.312
      HS-0.234-0.297-0.341-0.462-0.613-0.256
      PIV-LiteFlowNet-en0.0900.0890.0710.0720.1120.1130.1660.1680.1830.1870.0830.085
      LiteFlowNet-CL0.0250.0250.0360.0390.0940.0980.0760.0770.1260.1320.0360.042
    • Table 3. Ablation experiment

      View table
      View in Article

      Table 3. Ablation experiment

      实验CBAMResBlockConvLSTMRMSE
      1×××0.112 2
      2××0.102 8
      3×0.107 3
      40.100 4
      5××0.109 8
    Tools

    Get Citation

    Copy Citation Text

    Xin'ai LIU, Juan MENG, Hai DU, Zhiyuan LI. Particle image velocimetry method based on ConvLSTM and LiteFlowNet architecture[J]. Chinese Journal of Liquid Crystals and Displays, 2025, 40(7): 1023

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category:

    Received: Mar. 6, 2025

    Accepted: --

    Published Online: Aug. 11, 2025

    The Author Email: Juan MENG (mengjuan@dlou.edu.cn)

    DOI:10.37188/CJLCD.2025-0052

    Topics