A method for predicting blade meridian flow field based on explicit and implicit mechanism mining coding

Abstract

The present application discloses a method for predicting the meridian flow field of a blade based on explicit and implicit mechanism mining coding, which relates to the technical field of turbine blade design. The method comprises: obtaining the meridian plane grid and design operating parameters of the blade flow passage on an axial flow compressor described by a graph data structure; using a trained blade meridian flow field prediction model, according to the meridian plane grid and design operating parameters of the blade flow passage, predicting the physical field data of each node of the meridian plane grid of the blade flow passage; the present application mines the implicit differential correlation relationship within the physical data as a data feature in the data feature extraction process of the graph neural network model through an implicit mechanism mining mechanism, and explicitly encodes the turbine flow analysis mechanism into the model loss function, and trains the strategy for optimizing the blade meridian flow field prediction model through back propagation based on the model loss function, thereby improving the accuracy of the blade meridian flow field prediction and enhancing the model generalization and result confidence.

Description

Technical Field

[0001] The present application relates to the technical field of turbine blade design, and in particular to a method for predicting blade meridional flow fields based on explicit and implicit mechanism mining coding. Background Art

[0002] Blade design plays a crucial role in turbine machinery design, and analysis of its aerodynamic performance primarily relies on computational fluid dynamics (CFD) methods, which calculate the physical fields within the flow path. However, the large number of computational grids makes the calculation process time-consuming and labor-intensive, while the highly nonlinear characteristics of flow field data also lead to inefficient data analysis. In recent years, with advances in deep learning technology, academia has begun to explore the use of end-to-end data-driven approaches using neural networks to reduce design costs and improve efficiency. For example, in their 2024 paper “C(NN)FD–A Deep Learning Framework for Turbomachinery CFDAnalysis,” published in IEEE Transactions on Industrial Informatics, Bruni et al. used a 1.5-stage compressor rotor blade as the research object and constructed a C(NN)FD model to predict the effect of the rotor blade tip clearance on the flow field of multiple axial two-dimensional sections within the stage flow passage. In their 2024 paper “A fast prediction model of blade flutter in turbomachinery based on graph convolutional neural network,” published in Aerospace Science and Technology, Liu et al. used a graph data structure to describe the node and cell information of the CFD grid. Combining a graph neural network-based flow field prediction model and an aerodynamic damping prediction model, they predicted the flutter of a two-dimensional blade cascade under the influence of the flow field. Despite this, current data-driven aerodynamic performance prediction methods rely entirely on modeling flow field data, which limits their generalization capabilities.

[0003] In addition, the inherent "black box" nature of neural networks makes end-to-end data-driven methods difficult to trust in some critical tasks. Since Raissi et al. proposed the physically informed neural network (PINN) in their 2019 paper "Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations" in the Journal of Computational Physics, the method has become a current research hotspot by encoding the governing differential equations within the physical system into the model's loss function, thereby partially alleviating the problem of the opacity of the model's working mechanism. However, in the field of actual engineering simulation, taking blade aerodynamic calculations as an example, the promotion of such methods still faces multiple limitations, such as complex computational domains, large data sizes, and the inclusion of empirical formulas in the mechanism knowledge. Summary of the Invention

[0004] The purpose of this application is to provide a blade meridional flow field prediction method based on explicit and implicit mechanism mining coding, which can achieve fast and accurate prediction of the meridional flow field of the axial flow compressor blade stage.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] A blade meridian flow field prediction method based on explicit and implicit mechanism mining coding includes the following steps:

[0007] Obtain the meridian plane mesh and design operating parameters of the blade flow path on an axial flow compressor described using a graph data structure.

[0008] Using the trained blade meridional flow field prediction model, the physical field data of each node of the meridional plane grid of the blade flow channel is predicted according to the meridional plane grid and design operating parameters of the blade flow channel; the trained blade meridional flow field prediction model is a model obtained by training the blade meridional flow field prediction model using a model training strategy with explicit encoding of the flow mechanism; the model training strategy with explicit encoding of the flow mechanism is to explicitly encode the turbine flow analysis mechanism into the model loss function, and optimize the blade meridional flow field prediction model through back propagation based on the model loss function during the training process; the blade meridional flow field prediction model is a graph neural network model based on an implicit mechanism mining mechanism; the implicit mechanism mining mechanism is a mechanism for mining implicit differential correlations within physical data as data features during the data feature extraction process of the graph neural network model.

[0009] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0010] The present application provides a blade meridian flow field prediction method based on explicit and implicit mechanism mining encoding. In this method, the implicit differential correlation relationship inside the physical data is mined as a data feature during the data feature extraction process of the graph neural network model through the implicit mechanism mining mechanism, and the turbine flow analysis mechanism is explicitly encoded into the model loss function. The blade meridian flow field prediction model is trained by back propagation based on the model loss function, so that the trained blade meridian flow field prediction model can predict the physical field data of each node of the meridian plane grid of the blade flow channel according to the meridian plane grid and design operating condition parameters of the blade flow channel described by the graph data structure; the above-mentioned model used in this application mines the implicit differential correlation inside the physical data during the data feature extraction process to improve the accuracy of the blade meridian flow field prediction, and also constrains the training process by explicitly encoding the turbine flow analysis mechanism into the model loss function, thereby improving the model generalization and result confidence. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0012] Figure 1 A flowchart of a blade meridional flow field prediction method based on explicit and implicit mechanism mining coding is provided in one embodiment of the present application.

[0013] Figure 2 A flowchart of constructing and training a blade meridional flow field prediction model in a blade meridional flow field prediction method based on explicit and implicit mechanism mining coding provided in one embodiment of the present application.

[0014] Figure 3 A flowchart of step B1 in a blade meridional flow field prediction method based on explicit and implicit mechanism mining coding is provided in another embodiment of the present application.

[0015] Figure 4 A schematic structural diagram of a blade meridional flow field prediction model in a blade meridional flow field prediction method based on explicit and implicit mechanism mining coding provided in another embodiment of the present application.

[0016] Figure 5 A schematic structural diagram of a spatial-spectral regularized graph neural operator unit in a blade meridional flow field prediction method based on explicit and implicit mechanism mining coding provided in another embodiment of the present application. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0018] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0019] The embodiment of the present application provides a blade meridian flow field prediction method based on explicit and implicit mechanism mining coding. In an exemplary embodiment, Figure 1 As shown, the following steps are included:

[0020] A1. Obtain the meridian plane mesh and design operating parameters of the blade flow passage on an axial flow compressor described using a graph data structure.

[0021] A2. Using the trained blade meridian flow field prediction model, the physical field data of each node of the blade meridian mesh is predicted based on the blade meridian mesh and design operating parameters.

[0022] Specifically, the trained blade meridional flow field prediction model used in step A2 is a model obtained by training the blade meridional flow field prediction model using a model training strategy that explicitly encodes the flow mechanism; the model training strategy that explicitly encodes the flow mechanism is to explicitly encode the turbine flow analysis mechanism into the model loss function, and optimize the blade meridional flow field prediction model through back propagation based on the model loss function during the training process; the blade meridional flow field prediction model is a graph neural network model based on an implicit mechanism mining mechanism; the implicit mechanism mining mechanism is a mechanism that mines implicit differential correlations within physical data as data features during the data feature extraction process of the graph neural network model.

[0023] In this embodiment, the design operating parameters of the blade flow channel include rotor speed, inlet total pressure, inlet total temperature and outlet static pressure; the physical field data of each node of the meridian grid include pressure, density, temperature and velocity.

[0024] Before using the trained blade meridian flow field prediction model in step A2, it is obvious that the blade meridian flow field prediction method based on explicit and implicit mechanism mining coding in this embodiment also needs to build a blade meridian flow field prediction model and perform training, such as Figure 2 As shown, the process includes the following steps:

[0025] B1. Through computational fluid dynamics simulation, a flow field data sample set is constructed; any flow field data sample in the flow field data sample set includes the meridian grid of the blade flow channel described by the graph data structure, the design operating parameters and the physical field data of each node of the meridian grid. In this embodiment, Figure 3 As shown, step B1 specifically includes the following steps:

[0026] B11. Divide the geometric model of the blade flow channel on the axial flow compressor into a computational grid to obtain the meridian plane grid of the blade flow channel.

[0027] B12. The design operating parameters of the blade flow passage are sampled several times by using the Latin hypercube random sampling method to obtain several groups of design operating parameters.

[0028] B13. For any set of design operating parameters, perform fluid dynamics calculations based on the meridian plane grid and the design operating parameters to obtain the physical field data on each node grid of the meridian plane grid.

[0029] B14. The geometric model of the blade flow channel is calculated using the circumferential averaging method to obtain the physical field data of each node of the meridian plane grid.

[0030] Specifically, the geometric model of the axial compressor blades and flow passages is first divided into computational grids, and the four design operating parameters of rotor speed, inlet total pressure, inlet total temperature and outlet static pressure are sampled by the Latin hypercube random sampling method to obtain 1,000 sets of computational model samples under different design operating parameters; all computational model samples are calculated by the CFD method to obtain the pressure, density, temperature and velocity calculation results on each node grid in the flow passage; the geometric model of the blade flow passage is calculated by the circumferential averaging method to obtain the meridian data grid that is uniform in the axial and radial directions and used to characterize the meridian flow field and the physical field data on each node.

[0031] B15. Convert the meridian plane grid, design operating parameters and physical field data of each node of the meridian plane grid of the blade flow channel into a form described by a graph data structure and serve as a flow field data sample; the flow field data sample includes sample features and sample labels; the sample features are the meridian plane grid and design operating parameters described by the graph data structure; the sample labels are the meridian plane grid and physical field data of each node of the meridian plane grid described by the graph data structure.

[0032] The physical field data of the meridian grid and each node of the meridian grid are defined by the mathematical definition of the graph G ={ V , E} expression, where V is the node set of the meridian grid, ERepresents the connection relationship between nodes, and considers that there is a connection edge relationship between nodes belonging to the same meridian grid unit; the graph data can be represented by a feature matrix and the symmetric adjacency matrix Expressed as:

[0033] .

[0034] in, n is the number of nodes, d is the node feature dimension, Indicates the i The first grid node j Node features; Representation node and There are edge connections between them. Representation node and There is no edge connection between them; the spatial coordinates of each node and the four design parameters of the sample are used as input features; the output features are the density, pressure, temperature and velocity values at the node, that is, the physical field data. i The meridian plane grid coordinates and the four design variables of rotor speed, inlet total pressure, inlet total temperature and outlet static pressure corresponding to the sample are used as graph node features, and the adjacency relationship between the meridian plane grid nodes is used as graph edge connection information to obtain the input parameter graph. ; Each sample i The pressure, density, temperature and velocity calculation results of the meridian grid nodes are used as the graph node features, and the adjacency relationship between the meridian grid nodes is used as the graph edge connection information to obtain the physical field graph. ; where subscript i Represents each sample, 1 and 2 represent the design operating parameters and physical field parameters, respectively.

[0035] The samples extracted by calculation i Input parameter diagram of As a sample feature, the meridian plane physical field map As sample labels, together with a flow field data sample described using a graph data structure, a sample set of blade computational grid information, design operating parameters, and meridian plane physical field data was obtained. The sample set size is 1000 flow field data samples, which are divided into training set, validation set, and test set in an 8:1:1 ratio, ensuring both model training and validation and generalization testing.

[0036] B2. Construct a graph neural network model based on implicit mechanism mining to obtain a blade meridian flow field prediction model; Figure 4As shown in the figure, the blade meridian flow field prediction model includes the first fully connected layer, the graph information extraction network module and the regressor module.

[0037] Specifically, the first fully connected layer is used to transform the features of each node from the input dimension to the model calculation dimension through a fully connected transformation; preferably, the model calculation dimension is selected as 64.

[0038] The graph information extraction network module is based on an encoder-decoder structure, including an encoder path and a decoder path; the encoder path includes several sequentially connected encoder blocks, which are used to sequentially extract high-level feature information from the input data; the decoder path includes several sequentially connected decoder blocks, which are used to recover feature maps of different sizes for the input features; encoder blocks and decoder blocks of the same scale are connected through residual connections to fuse different levels of information; the final decoder block outputs a high-level feature map that fuses multi-level and multi-scale feature information; the regressor module is used to perform regression calculations based on the high-level feature map to obtain the physical field data of each node in the meridian grid.

[0039] The graph encoding after dimensionality enhancement is input into the encoder path of the graph information extraction network module based on the encoder-decoder structure, and high-level feature information is extracted in sequence through three sequentially connected encoder blocks; the resulting feature map output by the encoder path is input into the decoder path, and the feature map size is restored through three sequentially connected decoder blocks; in addition, the decoder block and encoder block results of the same size are fused with different levels of information through residual connections; finally, a multi-level and multi-scale fused high-level feature map is obtained; the high-level feature map is input into the regressor module to complete the regression calculation from high-level features to the physical field parameters of the meridian plane nodes.

[0040] The encoder block includes a spatial-spectral regularized graph neural operator unit, a GELU activation function, and a graph pooling layer; preferably, the graph pooling layer is a gPooling layer, and the node selection rate of each layer is set to 0.8; the decoder block includes an on-graph pooling layer, a GELU activation function, and a spatial-spectral regularized graph neural operator unit; preferably, the on-graph pooling layer is a gUnPooling layer; Figure 5 As shown in the figure, the spatial-spectral regularized graph neural operator unit includes a spectral branch, a spatial branch, a splicing mapping layer, a residual direct connection branch and a GELU function layer; the output result of the spectral branch is vector-concatenated with the output result of the spatial branch, mapped back to the input dimension through the splicing mapping layer, and then superimposed with the result of the residual direct connection branch, and the output of the spatial-spectral regularized graph neural operator unit is obtained through the nonlinear activation of the GELU function; the spatial-spectral regularized graph neural operator unit serves as the basic graph information aggregation layer of the blade meridional flow field prediction model, and mines the implicit differential correlation relationship within the physical data as data features during the data feature extraction process.

[0041] The spectral branch includes a kernel integration block and a first hybrid regularization block; the kernel integration block includes a nonlinear path, a linear path, and a first adder; the nonlinear path includes a graph Fourier transform layer, a kernel integration layer, and a graph inverse Fourier transform layer connected in sequence; the linear path includes a linear transformation layer; the transformation process of the nonlinear path can be summarized as follows:

[0042] .

[0043] in, y represents the output result of the nonlinear path, D represents the degree matrix of the input graph, I is the identity matrix, is the regularized graph Laplacian matrix, Q is the eigenvector matrix of the graph Laplacian matrix, is the eigenvalue matrix of the graph Laplace matrix, and the Fourier transform of the characteristic matrix is obtained by accomplish; K is the graph integral kernel, is a diagonal filter learned from data, which transforms the kernel Directly parameterize in the spectral domain, Represents element-by-element multiplication; matrix A is the adjacency matrix of the graph, Represents a node in the graph i and j Whether there is a connection relationship between them, if yes, it is 1, if not, it is 0.

[0044] The output end of the nonlinear path and the output end of the linear path are respectively connected to an input end of the first adder, and the output end of the first adder is connected to the first hybrid regularization block; the first hybrid regularization block includes a global regularization path, a local regularization path and a second adder; the output end of the global regularization path and the output end of the local regularization path are respectively connected to an input end of the second adder; the output of the second adder is comprehensively regularized by a trainable scaling coefficient and a translation coefficient; wherein, as a preferred embodiment, the global regularization path includes a GraphNorm layer, and the local regularization path includes a LayerNorm layer; the calculation process of the hybrid regularization block can be summarized as follows:

[0045] .

[0046] in, h is the encoded graph node information matrix, and Represent the output results of the LayerNorm layer and the GraphNorm layer respectively; and represents a trainable gating parameter with and , Same matrix dimensions; and represents the trainable scaling and translation parameter matrix, and are the trainable scaling parameter matrix and translation parameter matrix in the LayerNorm layer, and The trainable scaling parameter matrix and translation parameter matrix in the GraphNorm layer dynamically transform the two regularization results to obtain the final hybrid regularization result; and is the feature mean vector and standard deviation vector calculated for each graph node according to the feature dimension in the LayerNorm layer, Indicates the i The node in j The value of the feature dimension, and The calculation batches are j The feature mean and feature standard deviation of each dimension; n is the number of samples in the batch; and are the mean vector and standard deviation vector calculated according to the graph dimension in the GraphNorm layer, and The calculation batches are j The feature mean and feature standard deviation of each dimension; n is the number of samples in the batch; is a learnable gating parameter, For each feature dimension j The specific parameter value of is used to control the degree of retention of mean information.

[0047] The spatial branch includes a graph convolution layer and a second hybrid regularization block; the output of the graph convolution layer is connected to the input of the second hybrid regularization block; the structure of the second hybrid regularization block is exactly the same as that of the first hybrid regularization block. Preferably, the graph convolution layer in the spatial branch is a SAGEConv layer.

[0048] The regressor module includes a second fully connected layer, a RELU function layer, and a third fully connected layer. The second fully connected layer reduces the feature dimension of the input reconstruction image encoding by half, performs nonlinear activation through the RELU function layer, and then regresses through the third fully connected layer to obtain the output result. In one embodiment, the third fully connected layer regresses the output result with a feature dimension of 6, namely the pressure, temperature, density, and velocity components in three coordinate directions of each node in the meridian flow field.

[0049] B3. Based on the flow field data sample set, the blade meridian flow field prediction model is trained by adopting the model training strategy of explicit coding of the flow mechanism to obtain a trained blade meridian flow field prediction model.

[0050] The physical field data output by the model prediction and the physical field data corresponding to the sample labels are jointly input into the model loss function explicitly encoded by the mechanism, and the network parameters of the blade meridional flow field prediction model are updated according to the loss function value until the model converges.

[0051] Specifically, the model loss function of the blade meridian flow field prediction model includes a numerical deviation loss term, a control equation residual loss term, and a boundary condition residual loss term; the parameters of each layer of the blade meridian flow field prediction model are optimized and updated through the back propagation process until the model converges. The numerical deviation loss term is the loss value obtained by normalizing the absolute error between the physical field data output by the model prediction and the physical field data corresponding to the sample label; the control equation residual loss term is the loss value determined by using the meridian velocity gradient equation in the turbine flow analysis as the control equation, performing residual calculations based on the physical field data output by the model prediction and the physical field data corresponding to the sample label, and determining the residual value based on the deviation between the two residual values; the boundary condition residual loss term is the satisfaction of the physical field data output by the model prediction with respect to the boundary conditions given by the input sample. In this embodiment, the model loss function of the blade meridian flow field prediction model is shown below:

[0052] .

[0053] in, TotalLoss is the model loss value of the blade meridian flow field prediction model, l reg is the value of the numerical deviation loss term, l res is the value of the residual loss term of the control equation, l bc is the value of the boundary condition residual loss term, w 1. w 2 and w 3 are the weights of the numerical deviation loss term, the control equation residual loss term, and the boundary condition residual loss term, respectively.

[0054] The numerical deviation loss term can be calculated according to the following formula;

[0055] .

[0056] in, d is the total number of physics parameter values, l is the label of the physical field parameter value, n is the total number of nodes in the meridian mesh, i is the node label, and They represent the physical field data output by the model prediction and the physical field data corresponding to the sample label. i On the node l physical field parameter values, | | is the absolute value symbol, and Respectively represent the first l The maximum and minimum values of the physical field parameters.

[0057] The residual loss term of the control equation can be calculated according to the following formula:

[0058] .

[0059] in, The physical field data output by the model prediction i Substitute the physical field parameter values at each node into the control equation of the turbine flow analysis mechanism to obtain the residual value. The physical field data corresponding to the sample label i Substitute the physical field parameter values at each node into the control equation of the turbine flow analysis mechanism to obtain the residual value. A 、 B 、 C are the coefficients of the governing equations for the turbine flow analysis mechanism, w i For the i The relative velocity of the gas relative to the rotor blades in the flow field at each node is, q For the i The streamline where the node is located.

[0060] The governing equation for the turbine flow analysis mechanism is shown below:

[0061] .

[0062] in, w is the relative velocity of the gas relative to the rotor blades in the flow field, It represents the derivative of gas velocity with respect to the quasi-orthogonal direction of the streamline; , , It represents the derivative of each coordinate in the cylindrical coordinate system with respect to the quasi-orthogonal line direction of the streamline; It represents the flow direction angle, which is the angle between the relative velocity and the meridian streamline at that point; represents the flow inclination angle, which is the meridional relative velocity w m The angle between the meridian and the axis; the meridional relative velocity is defined as the projection velocity of the gas relative velocity on the meridian plane. is the meridional relative velocity wm Derivatives about the direction of the meridional streamlines; is the curvature of the streamline at each point, r is the radial coordinate in the rotor coordinate system; is the rotor angular velocity; It represents the derivatives of the total enthalpy, angular momentum and entropy of the inlet section with respect to the quasi-orthogonal direction of the streamline. T For temperature.

[0063] The boundary condition residual loss term is the degree to which the physical field data output by the model satisfies the boundary conditions given by the input sample. Specifically, it includes the average of the absolute deviations between the calculated inlet total pressure and inlet total temperature at the predicted meridian inlet boundary, and the static pressure at the outlet boundary and the given inlet total pressure, inlet total temperature, and outlet static pressure boundary conditions. The three are superimposed as the boundary condition residual loss term.

[0064] During the training process of the blade meridian flow field prediction model, each time a batch of training data undergoes the forward process, the numerical deviation loss is calculated based on the forward prediction output results and the physical field data corresponding to the label. l reg , residual loss of the control equation l res and boundary condition residual loss l bc and uses the training loss to update the network weights during the back propagation process until the blade meridian flow field prediction model training converges.

[0065] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0066] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A blade meridian flow field prediction method based on explicit and implicit mechanism mining coding, characterized by: include: Obtain the meridian plane mesh and design operating condition parameters of the blade flow passage of an axial flow compressor described using a graph data structure; Using the trained blade meridian flow field prediction model, according to the meridian plane mesh of the blade flow passage and the design operating condition parameters, the physical field data of each node of the meridian plane mesh of the blade flow passage is predicted; The trained blade meridian flow field prediction model is a model obtained by training the blade meridian flow field prediction model using a model training strategy with explicit coding of the flow mechanism; the model training strategy with explicit coding of the flow mechanism is to explicitly encode the turbine flow analysis mechanism into the model loss function, and optimize the blade meridian flow field prediction model through back propagation based on the model loss function during the training process; the blade meridian flow field prediction model is a graph neural network model based on an implicit mechanism mining mechanism; the implicit mechanism mining mechanism is to mine the implicit differential correlation relationship within the physical data as data in the process of data feature extraction of the graph neural network model. The blade meridian flow field prediction model includes a first fully connected layer, a graph information extraction network module and a regressor module; the first fully connected layer is used to transform the features of each node from the input dimension to the model calculation dimension through a fully connected transformation; the graph information extraction network module is an encoder-decoder structure, including an encoder path and a decoder path; the encoder path includes a plurality of sequentially connected encoder blocks, and the plurality of encoder blocks are used to sequentially extract high-level feature information from the input data; the decoder path includes a plurality of sequentially connected decoder blocks, and the plurality of decoder blocks are used to restore feature maps of different sizes for the input features; The encoder block and decoder block of the same scale fuse different levels of information through residual connections; the final decoder block outputs a high-level feature map that fuses multi-level and multi-scale feature information; the regressor module is used to perform regression calculations based on the high-level feature map to obtain the physical field data of each node of the meridian plane grid.

2. The blade meridian flow field prediction method based on explicit and implicit mechanism mining coding according to claim 1 is characterized in that: Before using the trained blade meridian flow field prediction model to predict the physical field data of each node on the meridian plane of the blade flow passage according to the meridian plane grid and design operating condition parameters of the blade flow passage, the blade meridian flow field prediction method based on explicit and implicit mechanism mining coding further includes: A flow field data sample set is constructed by computational fluid dynamics simulation; any flow field data sample in the flow field data sample set includes a meridian grid of a blade flow passage described by a graph data structure, design operating condition parameters, and physical field data of each node of the meridian grid; A graph neural network model based on implicit mechanism mining was constructed to obtain a blade meridian flow field prediction model; According to the flow field data sample set, the blade meridian flow field prediction model is trained by adopting a model training strategy of explicit coding of the flow mechanism to obtain a trained blade meridian flow field prediction model.

3. The blade meridian flow field prediction method based on explicit and implicit mechanism mining coding according to claim 2 is characterized in that: Through computational fluid dynamics simulation, a flow field data sample set is constructed, including: The geometric model of the blade flow passage on the axial flow compressor is divided into computational grids to obtain the meridian plane grid of the blade flow passage; The design operating parameters of the blade flow passage are sampled several times by using the Latin hypercube random sampling method to obtain several groups of design operating parameters; For any set of design operating condition parameters, a fluid dynamics calculation is performed based on the meridian plane grid and the design operating condition parameters to obtain physical field data on each node grid of the meridian plane grid; The geometric model of the blade flow channel is calculated using the circumferential averaging method to obtain the physical field data of each node of the meridian grid; The meridian plane grid, design operating parameters and physical field data of each node of the meridian plane grid of the blade flow channel are converted into a form described by a graph data structure and used as a flow field data sample; the flow field data sample includes sample features and sample labels; the sample features are the meridian plane grid and design operating parameters described by the graph data structure; the sample labels are the meridian plane grid and physical field data of each node of the meridian plane grid described by the graph data structure.

4. The blade meridian flow field prediction method based on explicit and implicit mechanism mining coding according to any one of claims 1 to 3 is characterized in that: The design operating parameters of the blade flow channel include rotor speed, inlet total pressure, inlet total temperature and outlet static pressure; the physical field data of each node of the meridian plane grid include pressure, density, temperature and velocity.

5. The blade meridian flow field prediction method based on explicit and implicit mechanism mining coding according to claim 1 is characterized in that: The encoder block includes a spatial-spectral regularization graph neural operator unit, a GELU activation function and a graph pooling layer; the decoder block includes a graph pooling layer, a GELU activation function and a spatial-spectral regularization graph neural operator unit; the spatial-spectral regularization graph neural operator unit includes a spectral branch, a spatial branch, a splicing mapping layer, a residual direct connection branch and a GELU function layer; the output result of the spectral branch is vector-spliced with the output result of the spatial branch, mapped back to the input dimension through the splicing mapping layer, and then superimposed with the result of the residual direct connection branch, and the output of the spatial-spectral regularization graph neural operator unit is obtained by nonlinear activation of the GELU function; the spatial-spectral regularization graph neural operator unit serves as the basic graph information aggregation layer of the blade meridional flow field prediction model, and mines the implicit differential correlation relationship within the physical data as a data feature during the data feature extraction process.

6. The blade meridian flow field prediction method based on explicit and implicit mechanism mining coding according to claim 5 is characterized in that: The spectral branch includes a kernel integration block and a first hybrid regularization block; the kernel integration block includes a nonlinear path, a linear path and a first adder; the nonlinear path includes a graph Fourier transform layer, a kernel integration layer and a graph inverse Fourier transform layer connected in sequence; the linear path includes a linear transformation layer; the output end of the nonlinear path and the output end of the linear path are respectively connected to an input end of the first adder, and the output end of the first adder is connected to the first hybrid regularization block; The first hybrid regularization block includes a global regularization path, a local regularization path, and a second adder; the output end of the global regularization path and the output end of the local regularization path are respectively connected to an input end of the second adder; the output of the second adder is comprehensively regularized by a trainable scaling coefficient and a translation coefficient; the spatial branch includes a graph convolution layer and a second hybrid regularization block; the output end of the graph convolution layer is connected to the input end of the second hybrid regularization block; The structure of the second hybrid regularization block is exactly the same as that of the first hybrid regularization block.

7. The blade meridian flow field prediction method based on explicit and implicit mechanism mining coding according to claim 2 is characterized in that: The regressor module includes a second fully connected layer, a RELU function layer and a third fully connected layer; the second fully connected layer halves the feature dimension of the input reconstructed image encoding, and after nonlinear activation by the RELU function layer, regresses through the third fully connected layer to obtain the output result.

8. The blade meridian flow field prediction method based on explicit and implicit mechanism mining coding according to claim 1 is characterized in that: The model loss function of the blade meridional flow field prediction model includes a numerical deviation loss term, a control equation residual loss term and a boundary condition residual loss term; the numerical deviation loss term is the loss value obtained by normalizing the absolute error between the physical field data output by the model prediction and the physical field data corresponding to the sample label; the control equation residual loss term is a loss value determined by using the meridional velocity gradient equation in the turbine flow analysis as the control equation, performing residual calculations based on the physical field data output by the model prediction and the physical field data corresponding to the sample label, and determining the loss value based on the deviation of the residual values of the two; the boundary condition residual loss term is the satisfaction of the physical field data output by the model prediction with the boundary conditions given by the input sample.

9. The blade meridian flow field prediction method based on explicit and implicit mechanism mining coding according to claim 1 is characterized in that: The model loss function of the blade meridian flow field prediction model is shown in the following formula: ； in, TotalLoss is the model loss value of the blade meridian flow field prediction model, l reg is the value of the numerical deviation loss term, l res is the value of the residual loss term of the control equation, l bc is the value of the boundary condition residual loss term, w 1. w 2 and w 3 are the weights of the numerical deviation loss term, the control equation residual loss term, and the boundary condition residual loss term, respectively.

Images