CN-122021292-A - Discontinuous flow field prediction method and system based on multi-precision data fusion

CN122021292ACN 122021292 ACN122021292 ACN 122021292ACN-122021292-A

Abstract

The invention discloses a discontinuous flow field prediction method and a discontinuous flow field prediction system based on multi-precision data fusion, wherein the discontinuous flow field prediction method constructs a flow field prediction task of a two-dimensional Riemann problem as a regression problem on a fixed structure grid, and adopts a low-order numerical method to respectively solve a control equation on two sets of grids with different resolutions to obtain corresponding solutions; meanwhile, the invention constructs a discontinuous flow field prediction model comprising a main branch, a gradient branch and a physical information branch, captures high-order discontinuous characteristics such as shock waves and grid convergence information contained in the multi-precision input respectively through the model, and embeds control equation constraint by means of the physical information branch, thereby ensuring that the prediction result accords with the physical rule while effectively predicting the high-frequency details of the flow field.

Inventors

XIAO ZHOUFANG
ZHANG TINGRUI

Assignees

杭州电子科技大学

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (10)

1. A discontinuous flow field prediction method based on multi-precision data fusion is characterized by comprising the following steps: constructing samples in the data set by using two sets of grids with different precision at physical quantities of adjacent time steps; The method comprises the steps of constructing an intermittent flow field prediction model, wherein the intermittent flow field prediction model comprises a main branch, a gradient branch and a physical information branch which are parallel, wherein the inputs of the main branch and the gradient branch are multi-precision inputs, and the inputs of the physical information branch are coordinate graphs of a predicted flow field in different directions and corresponding time steps thereof; The trunk branch comprises a first compression module and a characteristic fusion module, wherein the first compression module is used for carrying out cross reconstruction and channel compression on multi-precision input and inputting a processing result to the characteristic fusion module, the characteristic fusion module adopts a U-Net structure and comprises a first encoder, a first decoder and a first bottleneck layer, and the jump connection of the first encoder and the first decoder fuses the output of a gradient branch and a physical information branch in sequence; The gradient branch comprises a gradient field extraction module, a second compression module and a multi-scale feature extraction module; the multi-scale feature extraction module adopts a U-Net structure and comprises a second encoder, a second decoder and a second bottleneck layer, wherein in the multi-scale feature extraction module, the output of the second encoder is fused into a feature fusion module, and the output of the second decoder is used as the final output of a gradient branch; The physical information branch is used for carrying out channel expansion processing on input through the multi-layer perceptron, inputting the input into the third encoder, and merging the output of the third encoder into the characteristic fusion module; Training the discontinuous flow field prediction model by using the data set, and acquiring final flow field prediction results at different moments by using the trained discontinuous flow field prediction model.
2. The discontinuous flow field prediction method based on multi-precision data fusion according to claim 1 is characterized in that in a feature fusion module, the first encoder comprises a plurality of encoding layers, the first decoder comprises a plurality of decoding layers, the encoding layers and the decoding layers which are positioned on the same layer are connected in a jumping mode, the gradient branches and the outputs of physical information branches are fused on the jumping connection of the encoding layers and the corresponding decoding layers in sequence to obtain a first intermediate feature and a second intermediate feature, the inputs of the other decoding layers except the first decoding layer are the result of splicing the corresponding second intermediate feature with the output of the last decoding layer, and the input of the first decoding layer is the result of splicing the corresponding second intermediate feature with the output of the splicing layer.
3. The discontinuous flow field prediction method based on multi-precision data fusion of claim 2, wherein in the multi-scale feature extraction module, the second encoder comprises a plurality of encoding layers, the encoding layers in the second encoder are in one-to-one correspondence with the encoding layers in the first encoder, and the first intermediate features are obtained by respectively splicing and compressing the output of each encoding layer in the second encoder with the output of the corresponding encoding layer in the first encoder.
4. The discontinuous flow field prediction method based on multi-precision data fusion as set forth in claim 3, wherein the third encoder comprises a plurality of coding layers connected in sequence, the coding layers in the third encoder are in one-to-one correspondence with the coding layers in the first encoder, and the second intermediate features are obtained by multiplying the output of each coding layer in the third encoder with the corresponding first intermediate features element by element.
5. The discontinuous flow field prediction method based on multi-precision data fusion as set forth in claim 1, wherein the first compression module and the second compression module have the same structure and each comprise two compression sub-modules connected in series, wherein in the compression sub-modules, after input features are subjected to grouping normalization, the input features are divided into first features through a gating mechanism And second feature Respectively the first characteristics And second feature Splitting into first sub-features Second sub-feature 、 And will first sub-feature Respectively with the second sub-feature 、 The method comprises the steps of obtaining a first cross combination feature and a second cross combination feature by adding, splitting a spliced result of the first cross combination feature and the second cross combination feature into an upper layer feature and a lower layer feature, compressing the upper layer feature and the lower layer feature respectively, splicing compressed results, multiplying channel by channel, and obtaining output of a compression submodule by channel attention calculated by an average pooling layer and an activation function.
6. The discontinuous flow field prediction method based on multi-precision data fusion as claimed in claim 5, wherein features in the first compression sub-module are split according to the time steps of each channel, and features in the second compression sub-module are split according to the grids of each channel.
7. The discontinuous flow field prediction method based on multi-precision data fusion of claim 5, wherein the method for compressing the upper layer features comprises the steps of respectively processing the upper layer features by using two independent convolution layers, and adding the processing results element by element to obtain the upper layer processing features; The lower layer characteristics are compressed in such a way that the lower layer characteristics are spliced with the lower layer characteristics after convolution treatment, so as to obtain the lower layer treatment characteristics.
8. The discontinuous flow field prediction method based on multi-precision data fusion according to claim 1, wherein the extraction of the gradient field is achieved through a gradient operator with fixed parameters, the gradient operator is composed of two convolution kernels, the multi-precision input is respectively convolved with the two convolution kernels, and the results are added to obtain a final gradient field.
9. The discontinuous flow field prediction method based on multi-precision data fusion of claim 1, wherein total loss is constructed to guide model parameter updating in a training process, and comprises data loss, physical information loss, trunk gradient loss and branch gradient loss.
10. A discontinuous flow field prediction system based on multi-precision data fusion is characterized by being used for executing the discontinuous flow field prediction method based on multi-precision data fusion, and the discontinuous flow field prediction system comprises a data acquisition module, a data preprocessing module and a discontinuous flow field prediction module, wherein the data acquisition module is used for acquiring physical quantities in grids with different precision, the data preprocessing module is used for preprocessing the data acquisition module to acquire input data adapting to the discontinuous flow field prediction module, and the discontinuous flow field prediction module is used for predicting flow fields with different time steps according to different inputs.

Description

Discontinuous flow field prediction method and system based on multi-precision data fusion Technical Field The invention belongs to the technical field of flow field prediction, and particularly relates to a discontinuous flow field prediction method and system based on multi-precision data fusion. Background In the field of computational fluid mechanics, the problem of shock waves, contact discontinuities and the like in compressible flows is of great significance. Shock waves are discontinuities where fluid undergoes abrupt transitions in physical parameters such as pressure, density, and temperature under supersonic or strong compression conditions, and are essentially intense energy conversion and dissipation processes. For example, in the aerospace field, the shape and position of shock waves directly determine the aerodynamic performance and thermal load of an aircraft, and in power engineering, the shock waves have a key influence on the efficiency and safety of impeller machinery. Therefore, the accurate prediction of shock wave generation, structure and interaction with complex structures such as turbulence are the core for solving many front-edge scientific and engineering problems. To calculate shock waves accurately, various high resolution numerical methods have been developed. In order to clearly describe a discontinuous structure and inhibit non-physical oscillation, the traditional high-resolution format generally depends on extremely high grid resolution and fine time-step control, which directly leads to geometric progression increase of the number of grids and calculated amount required in three-dimensional complex flow simulation, so that the simulation with high fidelity often needs to consume huge calculation resources and time, and severely restricts the practical application of the simulation in engineering design and large-scale parameter research. Meanwhile, the artificial intelligence technology represented by deep learning makes breakthrough progress in the fields of computer vision, natural language processing, content generation and the like, provides a brand-new methodology for solving the problems of complex pattern recognition and high-dimensional mapping, and brings a new idea for overcoming the bottleneck in the CFD field. In this context, flow field prediction methods based on deep learning exhibit the potential to break through this dilemma. Currently, research on flow field prediction methods based on deep learning presents a diversified technical path. Among them, the pure data driven end-to-end approach exhibits strong fitting capability over large-scale data sets. The method learns the complex nonlinear mapping relation between input and output in an end-to-end mode by performing supervised training on a large-scale data set, so that direct prediction of a flow field is realized, but physical consistency and generalization capability of the flow field are limited. Therefore, researchers have developed a large number of methods for introducing physical constraints at the same time, and the method can realize rapid prediction of a flow field by establishing complex nonlinear mapping between input and output in an end-to-end mode through supervised learning, but has insufficient physical consistency, so that extrapolation capacity is limited. To remedy these drawbacks, researchers have developed physical constraint embedding methods in synchronization to enhance the reliability of models while reducing the amount of training data required. In recent years, a hybrid method of fusing data driving and physical mechanisms has become an important trend, and aims to synergistically promote generalization and physical consistency of a model. Although intelligent simulation methods have found widespread use, they still face significant challenges in solving the problem of strong flow field discontinuities (such as shock waves). This is mainly due to the difficulty in first, that high quality flow field data containing high resolution discontinuity structures are often difficult to obtain, and second, that partial differential equation based physical constraints essentially assume that the solution is smooth and continuous, and that it is difficult to accurately describe the discontinuous regions of the solution itself, which if applied directly might instead smooth out the discontinuity structures in the flow field, resulting in prediction distortion. Therefore, a discontinuous flow field prediction method based on multi-precision data fusion is provided. Disclosure of Invention The invention aims to provide a discontinuous flow field prediction method and a discontinuous flow field prediction system based on multi-precision data fusion. In a first aspect, the present invention provides a discontinuous flow field prediction method based on multi-precision data fusion, the method comprising: constructing samples in the data set by using two sets of grids with different precision