Scholarlink

Particle Image Velocimetry (PIV) is widely used to measure the spatial distribution of flow. In this study, downstream flow distribution data was obtained within a narrow rectangular channel with a central blockage through PIV measurements. The study combines deep learning algorithms with cross-correlation computations to achieve refined reconstruction of the blocked flow distribution in the spatial domain, enabling the acquisition of more detailed flow field distribution data. The reconstruction is implemented using the DMLResNet model, a deep convolutional neural network featuring multi-level nested residual connections and an attention mechanism. Extensive quantitative evaluations of the DMLResNet model’s reconstruction results indicate its superior performance in reconstructing complex flow data distributions. When the velocity vector density is increased by a factor of 16, the reconstruction error from the DMLResNet model consistently remains below half the interpolation computational error. At the same time, this study compares and analyzes the distribution of vortices, turbulence intensity, etc. based on the refined reconstruction of the small-scale flow distribution data, compared with the original data of PIV post-processing. This comparison is conducted to more deeply study the evolution of the flow characteristics downstream of the narrow rectangular channel under the blocking conditions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Deep Learning Modeling Approach for the Reconstruction of Flow Field Parameter Distributions Towards PIV Fine Measurements

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Biao Liang,
  • Sichao Tan,
  • Jiangkuan Li,
  • Zhen Liu,
  • Shouxu Qiao,
  • Ruifeng Tian

摘要

Particle Image Velocimetry (PIV) is widely used to measure the spatial distribution of flow. In this study, downstream flow distribution data was obtained within a narrow rectangular channel with a central blockage through PIV measurements. The study combines deep learning algorithms with cross-correlation computations to achieve refined reconstruction of the blocked flow distribution in the spatial domain, enabling the acquisition of more detailed flow field distribution data. The reconstruction is implemented using the DMLResNet model, a deep convolutional neural network featuring multi-level nested residual connections and an attention mechanism. Extensive quantitative evaluations of the DMLResNet model’s reconstruction results indicate its superior performance in reconstructing complex flow data distributions. When the velocity vector density is increased by a factor of 16, the reconstruction error from the DMLResNet model consistently remains below half the interpolation computational error. At the same time, this study compares and analyzes the distribution of vortices, turbulence intensity, etc. based on the refined reconstruction of the small-scale flow distribution data, compared with the original data of PIV post-processing. This comparison is conducted to more deeply study the evolution of the flow characteristics downstream of the narrow rectangular channel under the blocking conditions.