Search

CN-121982222-A - Single-view schlieren three-dimensional density field reconstruction method based on multi-mode pneumatic embedding

CN121982222ACN 121982222 ACN121982222 ACN 121982222ACN-121982222-A

Abstract

The invention relates to the technical field of flow field measurement and three-dimensional reconstruction, and discloses a single-view schlieren three-dimensional density field reconstruction method based on multi-mode pneumatic embedding. The method comprises the steps of obtaining single-view schlieren/background schlieren images and target object surface pressure distribution, establishing three-dimensional reconstruction domain and projection mapping, respectively extracting schlieren characteristics and pressure tokens through a schlieren encoder and a pressure encoder, executing cross attention retrieval on two types of tokens by a three-dimensional query point on the basis of projection alignment, adaptively fusing through learning gating, outputting a density field by a three-dimensional decoder through fused characteristics, and introducing micro-projection consistency constraint and structure maintenance constraint to optimize reconstruction results. The method uses the surface pressure as strong boundary priori, relieves the depth ambiguity of single view reconstruction, and improves the spatial positioning accuracy and the structural fidelity of the density field.

Inventors

  • LI JIANAN
  • LONG YIN

Assignees

  • 西南科技大学

Dates

Publication Date
20260505
Application Date
20260302

Claims (10)

  1. 1. A method for reconstructing a three-dimensional density field of a single-view schlieren chart based on multi-mode pneumatic embedding is characterized by comprising the following steps: Step 1, acquiring single-view schlieren observation images and surface pressure data, wherein the single-view schlieren observation is recorded as The surface pressure input is noted as a scalar field p (S) defined on the aircraft surface S; Step 2, converting an input image into a multi-scale two-dimensional characteristic representation through a schlieren mode encoder, and converting a surface discrete pressure point set into a pressure token with space semantics through a pressure mode encoder; Step 3, constructing a three-dimensional query unit corresponding to a voxel grid center point, a sparse sampling point set or layered sampling points along a camera ray; step 4, enabling the three-dimensional query to read schlieren features and pressure features simultaneously through the cross attention module to generate fused three-dimensional semantic features; And 5, outputting a density field rho by a three-dimensional decoder, integrating the network output through discrete light rays by virtue of micro-renderable consistency constraint to obtain a predicted deflection field, and calculating loss in alignment with input observation.
  2. 2. The method of claim 1, wherein the schlieren mode encoder employs a skip-coded encoded-decoded backbone structure or an explicit feature pyramid structure, and the output contains both high resolution detail features and low resolution global structural features.
  3. 3. The method of claim 1, wherein the pressure encoder unifies pressure encoding into a set of tokens with coordinates, sampling points for each surface Constructing an initial feature vector Wherein The stabilization of the pressure value is indicated, For the three-dimensional position coding, Including local geometric information.
  4. 4. The method of claim 1, wherein the cross-attention fusion employs a double cross-attention fusion gated fusion structure, one branch having the query cross-attention to the schlieren tokens and the other branch having the query cross-attention to the pressure tokens, ultimately converged with a learnable gating.
  5. 5. The method of claim 4, wherein the gating junction formula is: wherein Is the initial hidden state of the query point, In order for the gating coefficients to be learnable, And The outputs of the pressure branch and the schlieren branch, respectively.
  6. 6. The method of claim 1, wherein the micro-renderable consistency constraint converts the density field of the network output into a predicted deflection field by discrete ray integration and calculates the L1 loss with the input observations.
  7. 7. The method of claim 1, wherein the loss function comprises a body field fitting term Structure holding item And projection consistency term The final total loss is 。
  8. 8. The method of claim 7, wherein the loss factor Early training improvement with staged scheduling And (3) with Weight is used to learn reasonable three-dimensional structure, and the middle and later stages of training are gradually improved Weights are used to enhance observed consistency.
  9. 9. The method of claim 1, wherein the initial query vector of three-dimensional query points is obtained by projection alignment sampling by first projecting Bilinear sampling is carried out on the two-dimensional feature map to obtain schlieren features Then three-dimensional position coding Splicing, obtaining query vector through linear layer 。
  10. 10. A computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the method for reconstructing a single view schlieren three-dimensional density field based on multi-modal aerodynamic embedding as claimed in any one of claims 1-9.

Description

Single-view schlieren three-dimensional density field reconstruction method based on multi-mode pneumatic embedding Technical Field The invention relates to the crossing field of computational fluid mechanics and deep learning, in particular to a single-view schlieren three-dimensional density field reconstruction method based on multi-mode pneumatic embedding. Background Under transonic, supersonic or even hypersonic conditions, surrounding flow fields of a high-speed aircraft generally show complex phenomena such as strong compression, shock wave-expansion wave interaction, boundary layer development and separation, shear layer entrainment and the like. These phenomena macroscopically determine the aerodynamic forces and aerodynamic thermal loads of the aircraft, directly correlating in engineering the structural safety margins, control quality and overall performance. Compared with surface pressure distribution, the three-dimensional density field can more directly reveal shock wave position, intensity and space topological structure, and is a key intermediate quantity for explaining pneumatic load causes, positioning abnormal flow structures and supporting model correction and reliability assessment. Schlieren/background schlieren imaging provides a relatively low cost, strong intuitionistic flow field observation path. The method has the advantages that the refractive index gradient induced light deflection effect can be captured in a larger view field range without invading a flow field, so that the method is sensitive to density gradient information. However, engineering practice is limited by factors such as experimental layout space, cost of a synchronous multi-camera system, calibration complexity, shielding and the like, and only single-view schlieren images can be obtained. Single view observations are essentially projection measurements in which each pixel in an image corresponds to a path of light through a three-dimensional flow field, and the imaging response is related to the integral of the refractive index (or density) gradient across the path. The direct consequence of projection measurement is the indistinguishability of depth direction information, in that different three-dimensional density distributions may produce very similar or even identical two-dimensional grain effects, especially in non-axisymmetric flow fields, strong three-dimensional structures or multi-structure superposition. The multi-resolution and the discomfort enable the three-dimensional density field to be restored only by using the single-view schlieren image, enough constraint is lacking naturally, and problems such as space positioning drift, foreground and background aliasing, excessive smoothness of a structure or a pseudo structure are easy to occur. Disclosure of Invention The invention aims to provide a single-view schlieren three-dimensional density field reconstruction method based on multi-mode pneumatic embedding, which is characterized in that surface pneumatic pressure distribution is used as available strong physical constraint information and is embedded into the reasoning process of single-view schlieren three-dimensional reconstruction, so that the problem of multiple solutions of single-view reconstruction is remarkably relieved. In order to achieve the above purpose, the invention adopts the following technical scheme: A method for reconstructing a three-dimensional density field of a single-view schlieren chart based on multi-mode pneumatic embedding comprises the following steps: Step 1, acquiring single-view schlieren observation images and surface pressure data, wherein the single-view schlieren observation is recorded as The surface pressure input is noted as a scalar field defined on the aircraft surface S; Step 2, converting an input image into a multi-scale two-dimensional characteristic representation through a schlieren mode encoder, and converting a surface discrete pressure point set into a pressure token with space semantics through a pressure mode encoder; Step 3, constructing a three-dimensional query unit corresponding to a voxel grid center point, a sparse sampling point set or layered sampling points along a camera ray; step 4, enabling the three-dimensional query to read schlieren features and pressure features simultaneously through the cross attention module to generate fused three-dimensional semantic features; And 5, outputting a density field rho by a three-dimensional decoder, integrating the network output through discrete light rays by virtue of micro-renderable consistency constraint to obtain a predicted deflection field, and calculating loss in alignment with input observation. Preferably, the schlieren mode encoder adopts a code-decode type backbone structure with jump connection or an explicit characteristic pyramid structure, and the output contains both high resolution detail characteristics and low resolution global structural characteristics. Preferably, the p