A gas turbine turbine blade root wheel groove stress field reconstruction method and system

Abstract

The application discloses a gas turbine turbine blade root wheel groove stress field reconstruction method and system, and relates to the technical field of stress field reconstruction. The method comprises the following steps: collecting multiple sets of design variables, multiple sets of measuring point data and graph data corresponding to multiple finite element models under different profile design schemes; performing finite element calculation on the finite element models to obtain multiple first stress fields and multiple second stress fields; training a convolutional neural network by taking the multiple sets of design variables and the multiple sets of measuring point data as inputs and taking the multiple first stress fields as outputs; and training a graph convolutional neural network by taking the multiple sets of design variables, the multiple sets of measuring point data and multiple sets of unstructured grid data as inputs and taking the multiple second stress fields as outputs. The application does not need time-consuming finite element calculation in the application stage, can realize real-time simulation, and can process the most suitable data form for different network structures, thereby improving the stress field reconstruction efficiency under the root wheel groove profile design.

Description

Technical Field

[0001] This invention relates to the field of stress field reconstruction technology, and in particular to a method and system for reconstructing the stress field of the blade root groove of a gas turbine blade. Background Technology

[0002] Gas turbines, as core equipment in modern energy and power systems, are widely used in power generation, aviation propulsion, and marine power. Turbine blades are one of the key components of a gas turbine, and their performance directly affects the turbine's efficiency, reliability, and service life. Turbine blades operate in extreme environments of high temperature, high pressure, and high speed, enduring complex aerodynamic, thermal, and centrifugal loads. Therefore, the design and manufacture of turbine blades must meet extremely high technical requirements.

[0003] The blade root groove is a critical component connecting the turbine blade and the turbine disk, and its profile design has a decisive impact on the stress distribution at the blade root. The complex loads borne by the turbine blade during operation are transferred to the turbine disk through the blade root groove. If the blade root groove profile is poorly designed, it can lead to uneven stress distribution and stress concentration. Stress concentration not only significantly reduces the fatigue life of the blade but can also cause cracks or even fracture, seriously affecting the safety and reliability of the gas turbine. Therefore, one of the core objectives of blade root groove profile design is to optimize the profile shape to achieve a more uniform stress distribution, reduce stress peaks, and thus improve the blade's fatigue resistance and overall strength. Furthermore, a reasonable stress distribution can also reduce localized plastic deformation and creep damage, extending the blade's service life.

[0004] In practical design, profile design typically requires methods such as finite element analysis (FEM) to accurately obtain the stress distribution under different profile design schemes, and then iteratively optimize based on this. Although FEM calculations offer high accuracy, profile design usually involves a large number of FEM calculations, which are time-consuming, especially in complex 3D models and multi-stress field coupled analyses, where calculations can take hours or even days. Furthermore, profile design optimization often requires multiple iterations, each of which necessitates recalculating the FEM, further increasing the design cycle and computational cost. Therefore, profile design requires a significant amount of FEM calculations, greatly increasing the time cost of reconstructing the stress field under different profiles of the blade root groove. Summary of the Invention

[0005] Based on the shortcomings of the existing technology, the present invention provides a method and system for reconstructing the stress field of the blade root groove of a gas turbine blade, which solves the problem that the existing profile design requires a large number of finite element calculations, which greatly increases the time cost of reconstructing the stress field under different profiles of the blade root groove.

[0006] The present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for reconstructing the stress field of the blade root groove of a gas turbine blade, comprising the following steps:

[0008] Multiple finite element models of the blade root groove of a gas turbine blade under different profile design schemes were constructed, and multiple sets of design variables, multiple sets of measurement point data and graphical data were collected for the multiple finite element models. In the finite element models, the profile region of the blade root groove was meshed with a structured mesh, while the internal region of the blade root was meshed with an unstructured mesh.

[0009] Finite element calculations were performed on multiple finite element models to obtain multiple first stress fields corresponding to structured meshes and multiple second stress fields corresponding to unstructured meshes.

[0010] A stress field reconstruction model is constructed, which includes a parallel convolutional neural network and a graph convolutional neural network (GMM). The convolutional neural network is trained with multiple sets of design variables and multiple sets of measurement point data as inputs and multiple first stress fields as outputs. The graph convolutional neural network is trained with multiple sets of design variables, multiple sets of measurement point data and graph data as inputs and multiple second stress fields as outputs.

[0011] The design variables, measurement point data, and graphical data of the turbine root groove of the gas turbine blade are acquired in real time and input into the trained stress field reconstruction model to obtain the reconstructed stress field of the blade root groove.

[0012] Preferably, the convolutional neural network includes one input layer, three upsampling layers and one output layer. The input layer includes two fully connected layers, and each upsampling layer includes one interpolation layer, one batch normalization layer and two first convolutional layers. The output layer is one second convolutional layer.

[0013] The training of the convolutional neural network specifically includes the following steps:

[0014] The design variables and measurement point data are combined to obtain the input set;

[0015] The input set is linearly transformed and reshaped by the input layer to obtain the feature map tensor;

[0016] The feature map is obtained by interpolating, batch normalizing and convolving once through an upsampling layer on the feature map tensor.

[0017] The first stress field after reconstruction is obtained by performing a second convolution on the output feature map through the output layer.

[0018] Preferably, the graph data is as follows:

[0019] D = {A', B', X};

[0020] In the formula, D represents the graph data, A' and B' represent the adjacency matrix and degree matrix of the unstructured grid, and X represents the information matrix of the unstructured grid nodes.

[0021] Preferably, the GMM graph convolutional neural network includes one input layer, five intermediate layers and one output layer. The input layer and the intermediate layers both use Gaussian biased linear units as activation functions, while the output layer does not use an activation function.

[0022] The training of the graph convolutional neural network specifically includes the following steps:

[0023] The design variables, measurement point data, and graphical data are weighted through the input layer;

[0024] By adding nonlinearity to the weighted data through an intermediate layer, the reconstructed second stress field is obtained;

[0025] The reconstructed second stress field is output through the output layer.

[0026] Preferably, when training convolutional neural networks and graph convolutional neural networks, the network parameters are optimized by minimizing a sub-loss function, which is specifically shown below:

[0027] L T =L F +L S +L gx +L gy ;

[0028] In the formula, L T For the sub-loss function, L F Let L be the loss function of the global stress field. S The loss function for the maximum field value, L gx and L gy The loss function is the global stress field gradient in the x and y directions.

[0029] Preferably, the stress field reconstruction model is trained as a whole using a total loss function, which is specifically shown below:

[0030] L C =L T,1 +L T,2 +L′;

[0031] In the formula, L C Let L be the total loss function. T,1 With L T,2 L is the sub-loss function of the convolutional neural network and the GMM graph convolutional neural network, and L′ is the loss function composed of the prediction values ​​of the convolutional neural network and the GMM graph convolutional neural network for the common nodes of the structured grid and the unstructured grid.

[0032] Preferably, the measuring point data includes temperature data and displacement data.

[0033] In a second aspect, the present invention provides a device for reconstructing the stress field of the blade root groove of a gas turbine blade, comprising:

[0034] The data acquisition module is used to construct multiple finite element models of the turbine root groove of a gas turbine blade under different profile design schemes, and to acquire multiple sets of design variables, multiple sets of measurement point data and graphical data corresponding to multiple finite element models. In the finite element model, the profile region of the blade root groove adopts a structured mesh, while the internal region of the blade root adopts an unstructured mesh.

[0035] The calculation module is used to perform finite element calculations on multiple finite element models to obtain multiple first stress fields corresponding to the structured mesh and multiple second stress fields corresponding to the unstructured mesh.

[0036] The training module is used to construct a stress field reconstruction model, which includes a parallel convolutional neural network and a graph convolutional neural network (GMM). The convolutional neural network is trained with multiple sets of design variables and multiple sets of measurement point data as inputs and multiple first stress fields as outputs. The graph convolutional neural network is trained with multiple sets of design variables, multiple sets of measurement point data and graph data as inputs and multiple second stress fields as outputs.

[0037] The reconstruction module is used to acquire the design variables, measurement point data and graphical data of the blade root groove of the gas turbine in real time, and input them into the trained stress field reconstruction model to obtain the reconstructed stress field of the blade root groove.

[0038] Compared with the prior art, the above-mentioned at least one technical solution adopted by the present invention can achieve the following beneficial effects:

[0039] This invention first collected multiple sets of design variables, measurement point data, and graphical data under different profile design schemes, and performed finite element analysis (FEM) calculations on different profile design schemes to obtain multiple first stress fields corresponding to structured meshes and multiple second stress fields corresponding to unstructured meshes. This invention employs a parallel dual-branch design, utilizing a CNN branch to efficiently process structured mesh data and a GCN branch to accurately process unstructured mesh data, proposing a stress field reconstruction model, and training the model using corresponding training data. This method eliminates the need for time-consuming finite element calculations in the application stage, enabling real-time simulation. Furthermore, it can handle the data formats most suited to different network structures, improving the stress field reconstruction efficiency under blade root groove profile design. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This invention provides a stress field reconstruction model based on a graph-combined convolutional neural network.

[0042] Figure 2 The unstructured mesh and its corresponding undirected graph of this invention;

[0043] Figure 3 The graph shows the loss function variation curves of GCCNN during training under different graph convolution models of this invention.

[0044] in, Figure 3 (a): The loss function value on the training set. Figure 3 (b): The loss function value on the test set;

[0045] Figure 4 This refers to the field reconstruction deviation of GCCNN under different graph convolution models in this invention;

[0046] in, Figure 4 (a): Maximum absolute deviation of the field. Figure 4 (b): Average absolute deviation plot;

[0047] Figure 5 This is a comparison diagram of the field reconstruction deviation between GCCNN and GCN in the mesh encryption layer of this invention;

[0048] in, Figure 5 (a): Maximum absolute deviation of the field. Figure 5 (b): Average absolute deviation plot;

[0049] Figure 6 The graph shows the loss function change during the training process of GCCNN of this invention, along with a magnified view of a portion thereof.

[0050] in, Figure 6 (a): Loss function change curve. Figure 6 (b): Enlarged view of the part;

[0051] Figure 7 Here are the turbine blade root groove profiles of six typical samples of the present invention;

[0052] in, Figure 7 (a): Leaf root type line, Figure 7 (b): Groove profile;

[0053] Figure 8 The displacement field reconstruction result of the GCCNN of this invention;

[0054] in, Figure 8 of (a): u x Displacement field, Figure 8 (b): u y Displacement field, Figure 8 (c): u at each point on the blade root groove line x The distribution curve, Figure 8 (d): u at each point of the blade root groove profile y The distribution curve;

[0055] Figure 9 For the GCCNN strain ε of the present invention x Stress σ x Component field reconstruction results;

[0056] in, Figure 9 of (a): u x Strain field, Figure 9 (b): σ x Stress field, Figure 9 (c): ε at each point on the impeller root groove line x The distribution curve, Figure 9 (d): σ at each point of the blade root groove profile x The distribution curve;

[0057] Figure 10 For the GCCNN strain ε of the present invention y Stress σ y Component field reconstruction results;

[0058] in, Figure 10 (a): ε y Strain field, Figure 10 (b): σ y Stress field, Figure 10 (c): ε at each point on the impeller root groove line y The distribution curve, Figure 10 (d): σ at each point of the blade root groove profile y The distribution curve;

[0059] Figure 11 The GCCNN strain γ of the present invention xy Stress τ xy Component field reconstruction results;

[0060] in, Figure 11 (a): γ xy Strain field, Figure 11 (b): τxy Stress field, Figure 11 (c): γ at each point on the impeller root groove line xy The distribution curve, Figure 11 (d): τ at each point of the blade root groove profile xy The distribution curve;

[0061] Figure 12 This is the equivalent stress reconstruction result of sample T1 in this invention;

[0062] in, Figure 12 (a): Equivalent force field Figure 12 (b): Distribution curve of equivalent stress at various points on the blade root groove line;

[0063] Figure 13 This is the equivalent stress reconstruction result of sample T2 in this invention;

[0064] in, Figure 13 (a): Equivalent force field Figure 13 (b): Distribution curve of equivalent stress at various points on the blade root groove line;

[0065] Figure 14 This is the equivalent stress reconstruction result of sample T2 in this invention;

[0066] in, Figure 14 (a): Equivalent force field Figure 14 (b): Distribution curve of equivalent stress at various points on the blade root groove line;

[0067] Figure 15 This is the equivalent stress reconstruction result of sample T2 in this invention;

[0068] in, Figure 15 (a): Equivalent force field Figure 15 (b): Distribution curve of equivalent stress at various points on the blade root groove line;

[0069] Figure 16 This is the equivalent stress reconstruction result of sample T2 in this invention;

[0070] in, Figure 16 (a): Equivalent force field Figure 16 (b): Distribution curve of equivalent stress at various points on the blade root groove line;

[0071] Figure 17 This is the equivalent stress reconstruction result of sample T2 in this invention;

[0072] in, Figure 17 (a): Equivalent force field Figure 17 (b): Distribution curve of equivalent stress at various points on the blade root groove line;

[0073] Figure 18 For the evaluation of the prediction accuracy of the maximum equivalent stress of the turbine blade root groove of the present invention;

[0074] in, Figure 18 (a): Predicted maximum equivalent stress. Figure 18 (b): Distribution of relative deviation in maximum equivalent stress prediction;

[0075] Figure 19 This is a schematic diagram showing the location of the profile parameters of the present invention;

[0076] in, Figure 19 (a): Leaf root profile parameters, Figure 19 (b): Groove profile parameters. Detailed Implementation

[0077] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] In finite element analysis and computational mechanics, meshes are a fundamental tool for discretizing computational domains. Based on their geometric arrangement, they can be divided into structured meshes and unstructured meshes. Structured meshes consist of regularly arranged elements (such as quadrilaterals or hexahedrons), with well-defined topological relationships between mesh nodes. They are computationally efficient and accurate but struggle to handle complex geometries, making them suitable for convolution operations. Unstructured meshes consist of irregular elements (such as triangles or tetrahedrons), with connections between mesh nodes described by adjacency lists (not coordinate indices). They can accurately fit complex shapes but are computationally expensive, making them suitable for graph convolution operations. Therefore, this invention proposes a method for reconstructing the stress field of the turbine root groove of a gas turbine blade, comprising the following steps:

[0079] S1: Construct multiple finite element models of the turbine root groove of a gas turbine blade under different profile design schemes, and collect multiple sets of design variables, multiple sets of measurement point data and graph data corresponding to multiple finite element models; in the finite element model, the profile region of the blade root groove adopts a structured mesh, and the internal region of the blade root adopts an unstructured mesh.

[0080] A large amount of design variable data within a certain range is generated using the Latin hypercube method. The required variables are then used to parameterize the profile, and the shape of the profile can be determined through these design variables.

[0081] Let the generated sample points be {x} (1) ,x (2),...,x (N)}, where x is the sampling parameter sample, N is the number of sample points, and each point d represents a dimension, where each dimension represents a parameter type. Ensure proper projection over each dimension i. satisfy:

[0082]

[0083] in, It is the i-th dimension sample point in the k-th sample. It is the lower bound of the i-th dimension, that is, the minimum value of the sample points, Δx i It is the length of the i-th dimensional subspace, calculated as follows: p i (k) It is a randomly shuffled range index, p i (k) For each distinct k, it is used to divide the i-th dimension into subintervals, and the interval is defined as: Each sub-interval corresponds to an index, from 1 to N. That is, when p i (k) When p = 1, it is the first subinterval; when p i (k) When the interval is N, it is the last sub-interval. All N intervals across all dimensions are combined to generate N sample points.

[0084] The data preparation phase is mainly responsible for parsing the input information and constructing regular grid data suitable for CNNs and graph structure data suitable for GCNs.

[0085] (1) Design variable x d The Latin hypercube method was used to analyze the design variable x under various scenarios. d ∈R m Perform sampling.

[0086] (2) Measurement point data x s : It can directly read the measurement point data x s ∈R n The measurement data is data generated by measuring sensors at key locations, specifically including temperature data and vibration displacement data.

[0087] (3) Graph data D: used for GCN calculation.

[0088] S2: Perform finite element calculations on finite element models under different profile design schemes to obtain multiple first stress fields corresponding to structured meshes and multiple second stress fields corresponding to unstructured meshes.

[0089] S3: Construct a stress field reconstruction model, which includes a parallel convolutional neural network and a graph convolutional neural network (GMM). The convolutional neural network is trained using multiple sets of design variables and multiple sets of measurement point data as inputs, and multiple first stress fields as outputs. Similarly, the graph convolutional neural network is trained using multiple sets of design variables, multiple sets of measurement point data, and graph data as inputs, and multiple second stress fields as outputs.

[0090] To combine the advantages of convolutional neural networks (CNNs) for processing structured data and graph convolutional neural networks (GCNNs) for processing unstructured data, this invention combines these two types of neural networks to establish a Graph Combined Convolutional Neural Network (GCCNN) for processing finite element mesh data for strength analysis containing hybrid structures, thereby achieving stress field reconstruction. The overall framework of the model is as follows: Figure 1 As shown.

[0091] GCCNN consists of a convolutional neural network (CNN) and a graph convolutional neural network (GCNN) connected in parallel. The CNN is used for stress field reconstruction of the outer encrypted layer composed of a structured mesh, while the GCNN is used for stress field reconstruction of the inner unstructured mesh. The CNN mainly includes several convolutional layers, pooling layers, fully connected layers, batch normalization layers, and activation functions. Its input is a low-order tensor composed of design variables or local measurement point data. The GCNN consists of several graph convolutional layers and activation functions. Its input is graph data composed of design variables, local measurement point data, and node connection relationships. The output of both is the calculated stress field of the corresponding finite element mesh.

[0092] The input to the convolutional neural network is the design variable x. d ∈R m and measurement point data x s ∈R n The set X is as follows:

[0093]

[0094] Where n is the dimension of the measurement points and m is the dimension of the design variables.

[0095] The network consists of an input layer (fully connected layer), an upsampling layer, and an output layer.

[0096] The input layer consists of two fully connected layers:

[0097] Z (1) =W (1) X+b (1) ;

[0098] Z (2) =W (2) X+b(2) ;

[0099] Among them, W (1) and W (2) These are the weight matrices of the fully connected layer, b (1) and b (2) Z is the bias term. (1) Z is the linear transformation output of the first fully connected layer. (2) This is the linear transformation output of the second fully connected layer.

[0100] Then, for Z (2) Reshape the shape to match the spatial dimensions that CNNs can handle:

[0101] X (0) =reshape(Z) (2) (,H0,W0,C0);

[0102] Where H0, W0, and C0 are the initialized feature dimensions, and X... (0) It is a feature map tensor.

[0103] In convolutional neural networks, high-order tensor data and vectors can be equivalently converted to each other after shape reshaping. If the problem to be solved is in the form of dimensionality reduction at the data structure level, such as evaluating flow performance indicators from flow field results or evaluating the stress on the bearing surface from stress field, the fully connected layer mainly undertakes the task of dimensionality reduction of the vector after the high-order tensor data is reshaped. When dealing with problems of increased dimensionality, such as predicting the stress field from design parameters, the input low-dimensional vector is increased in dimensionality by the fully connected layer to form a high-dimensional vector, and then the high-order tensor can be generated by shape reshaping. These are the two important application scenarios of the fully connected layer in convolutional neural networks.

[0104] Each upsampling layer consists of one interpolation layer, one batch normalization layer, and two convolutional layers. The convolutional kernels of the convolutional layers are 3×3 in size, with a stride of 1, and modified linear units are used as activation functions.

[0105] The interpolation layer upsamples the input tensor, that is, it uses the data of neighboring elements to perform interpolation operations, thereby increasing the size of the input tensor in terms of length and width. The interpolation formula is as follows:

[0106]

[0107] Where I(·) is the interpolation operator, and bicubic interpolation is used here. This method uses 16 adjacent elements in the input tensor for interpolation, and because it incorporates more elements, it can obtain a smoother computation result. In the first upsampling layer, l = 0. Then proceed to the next step, outputting X. (1)Then in the second upsampling layer, Repeat this process.

[0108] For each point

[0109]

[0110] Where X(i,j) are integer coordinates of points on the original image, i and j refer to the integer coordinates of the discrete grid, and the weights w(i,j) are calculated using a cubic convolution kernel:

[0111]

[0112] Where t = xi represents the offset of a pixel.

[0113] During the training of convolutional neural networks, if the distribution of each batch of data differs, the neural network learns these differences in each iteration, which significantly increases the training cost. To accelerate network convergence and improve its generalization ability, batch normalization layers are applied to the construction of deep neural networks. By introducing two parameters to be learned, the batch normalization layer performs a linear normalization (minimum-maximum normalization) on the input tensor across different channels, i.e.:

[0114]

[0115] in, This is the normalized feature map. X is the interpolated input feature map. max X represents the maximum value in the current batch. min This is the minimum value in the current batch.

[0116] Unlike fully connected layers, each node in a convolutional layer is only connected to a subset of nodes in the previous layer. Furthermore, as the number of layers increases, the receptive field of the network gradually expands, allowing deeper networks to learn non-local features. These are the two main characteristics of convolutional neural networks: local connectivity and parameter sharing. Therefore, convolutional operations can better handle the temporal and spatial features of input data.

[0117] This contains two 3×3 convolutional layers and the ReLU activation function:

[0118] First layer:

[0119]

[0120] Second layer:

[0121]

[0122] Here, * represents the convolution operator.

[0123] Assuming there are L upsampling layers in total, the final output of the upsampling layers is X. (L) X (L) =X (l+k) .

[0124] The output layer is a single convolutional layer with a kernel size of 1×1 and a stride of 1. No activation function is used. The mathematical expression is as follows:

[0125] U J =W out *X (L) +b out ;

[0126] Among them, W out ∈R 1×1×C'×1 The weights are 1×1 convolution kernel parameters, c∈{1,2,...,C'} are channel indices, and X... (L) ∈R H′×W′×C′ b is the output feature map of the last convolution of the upsampling layer. out U is the bias term of a 1×1 convolution. J The stress field results are predicted by the convolutional neural network.

[0127] CNNs utilize regular grid data and extract and reconstruct stress field information through upsampling and convolution operations.

[0128] (1) The input design variables are transformed into an initial feature map format suitable for CNN by increasing the dimensionality through a fully connected layer;

[0129] (2) Interpolation upsampling is used to restore the spatial resolution of the feature map layer by layer.

[0130] (3) Use batch normalization to ensure feature stability and adaptability to data of different scales.

[0131] (4) Deep features are extracted through convolutional layers and the ReLU activation function, with each layer matching higher resolution information.

[0132] (5) Finally, stress field predictions are generated on a regular grid through convolutional layers.

[0133] GCN uses a graph convolutional model called GMM, which includes one input layer, five intermediate layers and one output layer. The input layer and intermediate layers use Gaussian biased linear units as activation functions, while the output layer does not use an activation function.

[0134] First, let's introduce the method for constructing the adjacency matrix:

[0135] Typically, a graph can be described by a pair G = {V, E}, where V is the set of vertices, E is the set of edges, and vertices are represented as v. i∈V, the edge is represented as e i,j =(v i ,v j )∈E, v is called i ,v j Vertices that are neighbors of each other. For example... Figure 2 The image shows an example of an unstructured mesh and its corresponding undirected graph. The nodes of the mesh correspond to the vertices of the graph, and the edges of the mesh correspond to the edges of the graph.

[0136] When describing the connectivity between vertices, an adjacency matrix A and a degree matrix B can be used. Adjacency matrix A stores information about whether vertices are connected to each other, while degree matrix B, a diagonal matrix, stores the number of neighboring vertices of a given vertex, calculated as follows:

[0137]

[0138] Because mesh nodes in the finite element method are connected through elements, the calculation of connection relationships differs from that in a graph. Not only are nodes at the ends of an edge neighbors, but all nodes within the same element are also neighbors. Taking node p8 and its corresponding vertex v8 as an example, the adjacency matrix element a of vertex v8 can be calculated. 8,j , degree matrix element b 8,8 And the adjacency matrix element a′ of node p8 8,j′ , degree matrix element b′ 8,8 for:

[0139] a 8,j =1,j=3,7,9,13;b 8,8 =4

[0140] a′ 8,j′ =1,j′=2,3,5,7,9,10,11,12,13,14; b′ 8,8 =10;

[0141] Therefore, in graph theory, unstructured grid data (graph data) can be described as:

[0142] D = {A', B', X};

[0143] In the formula, A' and B' are the adjacency matrix and degree matrix of the unstructured mesh, and X is the information matrix of the unstructured mesh nodes, which includes the spatial coordinates of the nodes and the results of the finite element analysis.

[0144] Then, the calculation formulas for all graph convolutional layers (GMM-GCN) will be introduced:

[0145] Graph Convolutional Models (GMMs) are a special form of Gaussian Convolutional Networks (GCNs) that uses a Gaussian mixture model for weighted information propagation. Its general form is as follows:

[0146]

[0147] In the formula, K is the number of Gaussian mixture components (i.e., using K different adjacency matrices A'). k ), γ k A' represents the weight coefficients (learnable parameters) in a Gaussian mixture model. k Let be the normalized adjacency matrix of the k-th Gaussian component. Let b be the graph convolution weight matrix corresponding to the k-th Gaussian component. (l) This is the bias term. The formula for calculating the normalized adjacency matrix is:

[0148]

[0149] In the formula, B' k Let be the degree matrix of the k-th Gaussian component.

[0150] The input to the convolutional neural network is the design variable x. d ∈R m Measurement point data x s ∈R n The set of graph data D:

[0151]

[0152] The convolution expression for the first layer graph is:

[0153]

[0154] The activation function is a Gaussian-biased linear unit:

[0155]

[0156] In the formula, σ 2 This is an adjustable parameter that controls the scale of the Gaussian activation.

[0157] Intermediate layer:

[0158] Each hidden layer is computed using GMM-GCN:

[0159]

[0160] Each layer uses Gaussian biased linear units as activation functions, giving information propagation a nonlinear characteristic.

[0161] Output layer:

[0162]

[0163] The output layer does not use an activation function, that is:

[0164] X(7) =Z (7) ;

[0165] Therefore, the final stress field prediction result is:

[0166] U T =X (7) ;

[0167] GCN propagates information through adjacency matrices to predict stress fields on irregular meshes (such as finite element meshes).

[0168] (1) GMM-GCN (Gaussian Mixture Graph Convolution) is used for information propagation, and multiple neighboring matrices are used to capture information at different distances.

[0169] (2) Perform graph convolution calculations layer by layer.

[0170] (3) The Gaussian bias activation function is used to increase the nonlinear expression ability and improve the prediction accuracy.

[0171] (4) Finally, the stress field prediction results on the irregular mesh are obtained.

[0172] Loss calculation and optimization.

[0173] (1) Taking into account the loss function L of the global stress field f The loss function L for the maximum field value s The loss function L for the global stress field gradient in the x and y directions gx ,L gy The loss function of the convolutional neural network is obtained as L. T,1 .

[0174] (2) Taking into account the loss function L of the global stress field f The loss function L for the maximum field value s The loss function L for the global stress field gradient in the x and y directions gx ,L gy The loss function of the graph convolutional neural network is obtained as L. T,2 .

[0175] (3) Taking into account the loss function L of the convolutional neural network T,1 The loss function L of a convolutional neural network. T,2 The total loss function L of GCCNN is obtained by combining the loss function L′ of the predictions of the two networks for the common nodes of the structured and unstructured grids. C This allows for the parallel connection of convolutional neural networks and graph convolutional neural networks.

[0176] Loss function:

[0177] In this invention, all tasks are regression tasks, and the Mean Square Error (MSE) is mainly used as the loss function, expressed as:

[0178]

[0179] In the formula, N is the sample size of the training samples in the dataset. with U i Let M be the output and actual value of the neural network for the i-th sample. i Let be the number of nodes in the i-th sample.

[0180] Considering the focus of structural strength calculation, the MSE of the maximum field value and the MSE of the global stress field gradient are additionally added to the loss function. The total loss function is:

[0181] L T =L F +L S +L gx +L gy ;

[0182] In the formula, L f The loss function for the global stress field is expressed as:

[0183]

[0184] L s The loss function for the maximum value of the field is expressed as:

[0185]

[0186] in, maxU represents the maximum output value of the neural network for the i-th sample. i The maximum actual value for the i-th sample;

[0187] L gx ,L gy Let be the loss function for the global stress field gradient in the x and y directions, expressed as:

[0188]

[0189] in, and Let be the global stress field gradient of the neural network output in the x and y directions for the i-th sample. and Let be the actual global stress field gradient in the x and y directions for the i-th sample;

[0190] In GCCNN, a combined loss function is constructed to achieve parallel operation of convolutional neural networks and graph convolutional neural networks, namely:

[0191] L C =L T,1 +L T,2 +L′;

[0192] In the formula, L T,1 With L T,2 For convolutional neural networks and graph convolutional neural networks in equation L T The calculated total network loss function, L′, is composed of the predictions from both networks for common nodes in the structured and unstructured grids, and its expression is:

[0193]

[0194] in, and These are the predictions from the convolutional neural network and the graph convolutional neural network at the common node, respectively.

[0195] The GCCNN is trained by backpropagation of the combined loss function during the optimization process.

[0196] Compared to standalone convolutional neural networks and graph convolutional neural networks, GCCNN combines loss functions to allow two networks to be trained simultaneously. This not only effectively improves the accuracy and smoothness of prediction results in common regions, but also reduces the possibility of overfitting to some extent by increasing the complexity of the model.

[0197] Compared to standalone convolutional neural networks and graph convolutional neural networks, GCCNN combines loss functions to allow two networks to be trained simultaneously. This not only effectively improves the accuracy and smoothness of prediction results in common regions, but also reduces the possibility of overfitting to some extent by increasing the complexity of the model.

[0198] S4: Real-time acquisition of design variables, measurement point data, and graphical data of the turbine blade root grooves, and inputting them into the trained stress field reconstruction model to obtain the reconstructed stress field of the blade root grooves. Example

[0199] To verify the accuracy of the GCCNN method in reconstructing the stress field of the blade root groove of a gas turbine blade, this invention conducted simulation verification, which mainly included verification of the selection of graph convolutional layers, establishment of the stress field reconstruction model, an example of the blade root groove profile, examples of the reconstruction of displacement, strain and stress components, and analysis of the equivalent stress reconstruction results.

[0200] Verification of the selection of convolutional layers.

[0201] In the preceding description, this invention directly specifies the graph convolutional layer as Gaussian Mixture Graph Convolution (GMM). To select the optimal graph convolutional model for the field reconstruction task, this invention compares the performance of four models—frequency domain graph convolution ChebNet, spatial domain graph convolution Graph Sample and Aggregate (GraphSAGE), Gaussian Mixture Model (GMM), and Spline Convolution (SplineConv)—on stress and strain field reconstruction tasks under the same parameter settings. The loss function variation curves of all models during training are shown below. Figure 3 As shown, the leaf root and wheel groove dataset is used as the training and testing datasets for the model. The curve represents the loss function value of the model on the training set, and the scatter plot represents the loss function value of the model on the test set.

[0202] After training, the predictive performance of all models on the test set for displacement, strain, and stress was further compared, using the maximum absolute deviation and the mean absolute deviation of the field as indicators. The results are as follows: Figure 4 As shown, consistent with the loss function of the model on the validation set after training, the frequency domain graph convolutional model Chebnet performed the worst, while the spatial domain graph convolutional model GMM performed the best. Furthermore, even though SplineConv exhibited significant overfitting during training, its performance still outperformed other graph convolutional models besides GMM. In conclusion, this invention selects GMM as the graph convolutional model for GCCNN.

[0203] Since structured data can be viewed as a simplified form of unstructured data, and structured mesh data can also be described using graphs, the stress field reconstruction task of the blade root groove can be completed entirely using only GCN. Therefore, the next step is to evaluate the stress field reconstruction performance of GCCNN and GCN in the structured mesh-refined region at the geometric boundary of the blade root groove. The field reconstruction deviations of displacement, strain, and stress on the test set are as follows: Figure 5 As shown, for each parameter, the maximum absolute field deviation and the mean absolute field deviation of the GCCNN model are generally smaller than those of the GCN model alone. Therefore, the GCCNN method proposed in this invention has higher prediction accuracy and better performance than the GCN method alone.

[0204] Establishment of the stress field reconstruction model.

[0205] The CNN consists of one input layer, three upsampling layers, and one output layer. Neither the input nor output layers use activation functions, while the upsampling layers use modified linear units (MRUs) as activation functions. The GCN uses a Gaussian-biased linear unit (GMM) model, consisting of seven graph convolutional layers: one input layer, one output layer, and five intermediate layers. Both the input and intermediate layers use GMMs as activation functions, while the output layer does not use an activation function. Table 1 provides the detailed parameters of the model.

[0206] It should be noted that the 10×247 design of the CNN upsampling layer 3 and output layer size in Table 1 is to correspond to the number of mesh nodes densified at the boundary of the turbine blade root groove; the input size of GCN is 14, which includes the x and y coordinates of the nodes and 12 design variables; the output size is 8, which includes 2 displacement components, 3 strain components and 3 stress components of the blade root groove.

[0207] Table 1 Model parameters of GCCNN

[0208]

[0209] By sampling within the range of design parameters for the turbine blade root groove using the Latin hypersol method, a dataset of turbine blade root grooves satisfying the constraints was established. The dataset has a sample size of 6000, divided into a training set (70%, 4200 samples) and a test set (30%, 1800 samples) in a 7:3 ratio for building the stress field reconstruction model. The range of sampled values ​​is shown in Table 2.

[0210] Table 2. Range of values ​​for parametric design variables of turbine blade root grooves

[0211]

[0212] The schematic diagram of the profile parameter location is as follows: Figure 19As shown in the figure. To avoid the blade root shape being too wide or too long due to unreasonable values ​​of design variables, constraints were imposed on the relevant parameters of the model during the research process. These constraints are the overall width of the profile L1 / mm and the overall height of the profile L2, i.e., the parameters in the box in the figure. The overall width of the profile should be less than the width of the blade root intermediate body, and the increase in the overall height of the profile should not exceed 20%. The workstation configuration for training and testing the model was an Intel Core i7-12700K CPU and an Nvidia RTX A6000 (48GB) GPU. The turbine blade root groove dataset was used as the training and testing dataset for the model. The convolutional neural network adopted the Adam optimization algorithm with an initial learning rate of 0.001, and the graph convolutional neural network adopted the Adam optimization algorithm with an initial learning rate of 0.005. Both used the MSE loss function. The batch sample size was 64, and an adaptive learning rate descent algorithm based on the loss function was used. The loss function change curve of the model during training is shown in the figure. Figure 6 As shown, during training, the loss function value of GCCNN generally showed a decreasing trend. As can be seen from the enlarged graph, towards the end of training, after more than 550 iterations, the rate of decrease in the loss function slowed further, and the loss function value on the test set gradually became significantly higher than that on the training set. At this point, the loss function value had decreased sufficiently, so training ended after 600 iterations. The average loss function value of the model in the last 100 iterations on both the training and test sets was 2.59 × 10⁻⁶. -5 and 2.87×10 -5 .

[0213] Table 3 further compares the computational overhead of GCCNN and the finite element method in calculating the stress field of turbine blade root grooves. Compared with finite element analysis, the model proposed in this invention can achieve rapid reconstruction of the stress field of turbine blade root grooves after deployment. As can be seen from the table, the average computation time under GPU acceleration is reduced from 12s to 0.16s, and the average computation speed is improved by 75 times.

[0214] Table 3 Comparison of average computation time for reconstructed models and finite element analyses.

[0215]

[0216] Example of blade root groove profile.

[0217] Table 4 presents six representative samples from the turbine blade root groove test sample set. Figure 7 The geometry of the leaf root and wheel groove profiles of these six samples is shown, which is a typical structure of the leaf root and wheel groove profile of a four-toothed fir tree.

[0218] Table 4 Six typical design schemes

[0219]

[0220] Examples of reconstruction of displacement, strain, and stress components.

[0221] Let's take sample T1 as an example for explanation. Figure 8 The comparison between finite element analysis and reconstruction results of the displacement field in the x and y directions is presented. Under the action of centrifugal force and equivalent tension, the u-axis of the blade root groove... x The displacement field exhibits a numerically symmetrical distribution along the centerline, while the leaf root teeth show a contracting trend along the x-direction. Most of the region to the left of the centerline shows a symmetrical distribution. x ≥0, most of the right-hand region u x ≤0, while the groove is the opposite, with the groove opening widening along the x-direction, and the centerline to the left of u x ≤0, right side u x ≥0. u of the blade root groove y The displacement field is also symmetrically distributed along the centerline, and the u in the leaf root region y It is significantly larger than the wheel groove area. Regarding reconstruction accuracy, u x and u y The reconstruction deviations of the displacement field are all within 5%, and u y The overall relative prediction bias is much smaller than u x .from Figure 8 (c) and Figure 8 As can be seen from (d), the finite element analysis results and the reconstruction results of the displacement on the blade root groove profile are in good agreement; since the maximum value is often the focus in strength analysis, its relative deviation is given, and the displacement u on the blade root profile is shown. x u y The relative deviation of the maximum value is 0.40% and 0.07%, respectively, on the grooved line. x u y The relative deviation of the maximum value is 1.00% and 0.05%.

[0222] Figure 9 Furthermore, a comparison between the finite element analysis results and the reconstruction results of the normal strain stress field and shear stress field in the x and y directions is presented. The results both demonstrate that GCCNN can reconstruct the three stress and strain component fields with high accuracy, and the distributions of the stress and strain fields remain consistent. Figure 9 Figures (a) and (b) show the reconstruction results of normal strain and normal stress in the x-direction. The field distribution still maintains a highly symmetrical form, with the maximum ε at the leaf root. x σ x The value appears at the concave fillet below the non-load-bearing surface of the first pair of teeth, where the maximum ε of the groove is... x σ x The value appears at the concave rounded corner at the bottom. Figure 9The finite element analysis results and reconstruction results of the displacement on the blade root groove in (c) and (d) are in good agreement, and the ε on the blade root groove is... x σ x The relative prediction deviation of the maximum value is 2.13% and 1.60%, respectively, on the wheel groove line ε x σ x The relative prediction bias for the maximum value was 0.08% and 0.42%.

[0223] Figure 10 The reconstruction results of normal strain and normal stress in the y-direction are shown, and the field distribution still maintains a highly symmetrical form. As shown in the figure, the maximum ε at the leaf root... y σ y The value appears at the concave fillet above the bearing surface of the first pair of teeth, on the wheel groove line ε y σ y Multiple peaks of similar size appear, located at the concave fillet below the bearing surface of each pair of teeth. Figure 10 The finite element analysis results and reconstruction results of the displacement on the blade root groove in (c) and (d) are in good agreement, and the ε on the blade root groove is... y σ y The relative prediction deviation of the maximum value is 1.22% and 0.99%, respectively, on the wheel groove line ε y σ y The relative prediction bias for the maximum value was 0.43% and 0.40%.

[0224] Figure 11 The reconstruction results of shear strain and shear stress in the xy plane are shown. The field distribution still maintains a highly symmetrical form, with the maximum γ at the leaf root. xy τ xy The value is located at the concave fillet above the bearing surface of the second pair of teeth, with the maximum γ value in the wheel groove. xy τ xy The value is located at the concave fillet at the bottom of the profile. Figure 11 The finite element analysis results and reconstruction results of the displacement on the blade root groove in (c) and (d) are in good agreement, and the γ on the blade root groove is consistent with the finite element analysis results. xy τ xy The relative prediction deviation of the maximum value is 0.16% and 0.26%, respectively, on the γ-axis of the wheel groove line. xy τ xy The relative prediction bias for the maximum value was 0.62% and 0.12%.

[0225] In summary, the three strain and stress components of the turbine blade root groove exhibit multiple peaks along the blade profile, all located in the concave fillet region of the root groove. Therefore, parameters for this region were considered in the parametric modeling. Since the centrifugal force and equivalent tensile force are mainly along the y-direction, ε... y σ yOverall, the stress component is larger than the other two strain and stress components. The reconstruction results show that the maximum absolute deviation of the stress component reconstruction is less than 36 MPa, and the maximum relative deviation on the blade root groove profile is 2.13%, which further proves that GCCNN has very high accuracy in the reconstruction of the stress field of the blade root groove of a gas turbine blade.

[0226] Analysis of equivalent stress reconstruction results.

[0227] This section will present a comparison of the equivalent stress field results of six typical samples of turbine blade root grooves under finite element analysis and reconstruction models, and then randomly select three batches of samples to evaluate the prediction accuracy of the maximum equivalent stress. Figure 12 The finite element analysis and reconstruction results of the equivalent stress field of sample T1 are presented. As shown in the figure, the maximum equivalent stress of T1 is the smallest among these six samples. The tooth width control parameters, contact surface tooth width, and contact surface inclination angle of this sample are significantly higher than those of the other five samples. The maximum absolute deviation of the reconstructed field is 36 MPa. The maximum equivalent stress of the blade root groove is 742 MPa, which occurs at the concave fillet above the bearing surface of the first pair of teeth at the blade root. The reconstruction deviation at this location is 0.56%. The finite element analysis results and reconstruction results of the equivalent stress along the blade root groove profile are in good agreement. Figures 13-17 The equivalent stress reconstruction results for samples T2 to T6 are further presented.

[0228] Compared to sample T1, T2 has smaller tooth width control parameters, contact surface width, and contact surface inclination angle, while the concave fillet radius of the blade root groove is larger. The maximum absolute deviation of the reconstructed field is 35 MPa, and the maximum equivalent stress of the blade root groove is 800 MPa, which occurs at the concave fillet above the bearing surface of the first pair of teeth at the blade root. The reconstruction deviation at this location is 0.89%. The finite element analysis results of the equivalent stress along the blade root groove profile are in good agreement with the reconstruction results.

[0229] The tooth width control parameters, contact surface inclination angle, and concave radius of the blade root groove of sample T3 are smaller than those of T1 and T2. The maximum absolute deviation of the reconstructed field is 48 MPa, and the maximum equivalent stress of the blade root groove is 895 MPa, which occurs at the concave fillet at the bottom of the groove. The reconstruction deviation at this point is 0.08%. The finite element analysis results of the equivalent stress along the blade root groove profile are in good agreement with the reconstruction results.

[0230] The stress level of sample T4 is significantly higher than that of the previous three samples, and its contact surface width, concave fillet radius, and convex fillet radius of the blade root groove are further reduced. The maximum absolute deviation of the reconstructed field is 53 MPa, and the maximum equivalent stress of the blade root groove is 990 MPa, which occurs at the concave fillet above the bearing surface of the first pair of blade roots. The reconstruction deviation at this location is 1.08%. The finite element analysis results of the equivalent stress along the blade root groove profile are in good agreement with the reconstruction results.

[0231] The throat width of sample T5 is comparable to that of T2 and T4. The concave and convex arc radii of the blade root groove are both greater than those of T4. The contact surface inclination angle is the smallest among the six samples. The maximum absolute deviation of the reconstructed field is 50 MPa. The maximum equivalent stress of the blade root groove is 831 MPa, which occurs at the concave fillet below the bearing surface of the first pair of teeth in the groove. The reconstruction deviation at this location is 1.11%. The finite element analysis results of the equivalent stress along the blade root groove profile are in good agreement with the reconstruction results.

[0232] Considering stress level as the design objective, among the six samples, the design scheme of T6 is only better than T4. Its blade root and wheel groove concave fillet radius and convex fillet radius are both smaller, and its tooth width control parameter is also the smallest among the six samples.

[0233] like Figure 17 As shown, the maximum absolute deviation of the reconstructed field is 49 MPa, and the maximum equivalent stress of the blade root groove is 907 MPa, which occurs at the concave fillet above the bearing surface of the first pair of teeth of the blade root. The reconstruction deviation at this point is 0.63%. The finite element analysis results of the equivalent stress along the blade root groove profile are in good agreement with the reconstruction results.

[0234] In the structural design of blade root groove profiles, the maximum equivalent stress is often the most important indicator, so it is necessary to evaluate the accuracy of the model in predicting the maximum equivalent stress of the blade root groove. Figure 18 The results of the model’s maximum equivalent stress prediction accuracy evaluation for three random batches (batch sample size is 64) in the turbine blade root groove test dataset are presented. Figure 18 (a) is the maximum equivalent stress scatter plot, where the horizontal axis of the stress scatter plot is the result of the finite element analysis and the vertical axis is the result of the reconstruction model. When the stress scatter plot is closer to the straight line y = x, it indicates that the prediction accuracy of the reconstruction model is higher. As can be seen from the figure, the maximum equivalent stress prediction deviation of the reconstruction model of the stress field of the turbine blade root groove is within 5%. Figure 18 (b) shows the relative deviation distribution of the maximum equivalent stress prediction results. It can also be found that the maximum equivalent stress prediction deviation of the blade root groove in the random batch sample is within 5%, and the median deviation is 0.46%.

[0235] The deviation analysis results of the above 6 samples and the maximum equivalent stress show that the stress field reconstruction model based on the GCCNN turbine blade root groove can accurately reconstruct the displacement, strain and stress field of different design parameter profiles in the test dataset, and accurately predict the maximum equivalent stress of the blade root groove.

[0236] Based on the same concept, the present invention also provides a gas turbine blade root groove stress field reconstruction system, including an acquisition module, a calculation module, a training module and a reconstruction module.

[0237] The data acquisition module is used to construct multiple finite element models of the turbine root groove of a gas turbine blade under different profile design schemes, and to acquire multiple sets of design variables, multiple sets of measurement point data and graphical data corresponding to multiple finite element models. In the finite element model, the profile region of the blade root groove adopts a structured mesh, while the internal region of the blade root adopts an unstructured mesh.

[0238] The calculation module is used to perform finite element calculations on multiple finite element models to obtain multiple first stress fields corresponding to the structured mesh and multiple second stress fields corresponding to the unstructured mesh.

[0239] The training module is used to construct a stress field reconstruction model, which includes a parallel convolutional neural network and a graph convolutional neural network (GMM). The convolutional neural network is trained with multiple sets of design variables and multiple sets of measurement point data as inputs and multiple first stress fields as outputs. The graph convolutional neural network is trained with multiple sets of design variables, multiple sets of measurement point data and graph data as inputs and multiple second stress fields as outputs.

[0240] The reconstruction module is used to acquire the design variables, measurement point data and graphical data of the blade root groove of the gas turbine in real time, and input them into the trained stress field reconstruction model to obtain the reconstructed stress field of the blade root groove.

[0241] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0242] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method of reconstructing a stress field in a gas turbine turbine blade root wheel slot, characterized by, The method comprises the following steps: A plurality of finite element models of a gas turbine turbine blade root wheel groove under different profile design schemes are constructed, and a plurality of groups of design variables, a plurality of groups of measurement point data and graph data corresponding to the plurality of finite element models are collected; In the finite element model, a structured grid is used in the root wheel groove profile area, and an unstructured grid is used in the internal area of the root; A plurality of first stress fields corresponding to the structured grid and a plurality of second stress fields corresponding to the unstructured grid are obtained by performing finite element calculation on the plurality of finite element models; A stress field reconstruction model is constructed, and the stress field reconstruction model comprises a convolutional neural network and a GMM graph convolutional neural network in parallel; The convolutional neural network is trained by taking the plurality of groups of design variables and the plurality of groups of measurement point data as inputs and taking the plurality of first stress fields as outputs; The graph convolutional neural network is trained by taking the plurality of groups of design variables, the plurality of groups of measurement point data and the graph data as inputs and taking the plurality of second stress fields as outputs; The design variables, the measurement point data and the graph data of the gas turbine turbine blade root wheel groove are obtained in real time, and are input into the trained stress field reconstruction model to obtain a reconstructed root wheel groove stress field.

2. A method of reconstructing a stress field in a gas turbine turbine blade root wheel slot according to claim 1, characterized in that The convolutional neural network comprises one input layer, three up-sampling layers and one output layer, the input layer comprises two fully connected layers, each up-sampling layer comprises one interpolation layer, one batch normalization layer and two first convolutional layers, and the output layer is one second convolutional layer; The convolutional neural network is trained by taking the plurality of groups of design variables and the plurality of groups of measurement point data as inputs and taking the plurality of first stress fields as outputs; The design variables and the measurement point data are combined to obtain an input set; The input set is linearly transformed and reshaped by the input layer to obtain a feature map tensor; The feature map tensor is interpolated, batch-normalized and once-convoluted by the up-sampling layer to obtain an output feature map; The output feature map is twice-convoluted by the output layer to obtain the reconstructed first stress field.

3. A method of reconstructing a stress field in a gas turbine turbine blade root wheel slot according to claim 1, wherein The graph data is specifically as follows: D={A',B',X}; In the formula, D is the graph data, A' and B' are the adjacency matrix and the degree matrix of the unstructured grid, and X is the information matrix of the unstructured grid node.

4. A method of reconstructing a stress field in a gas turbine turbine blade root wheel slot according to claim 3, characterized in that, The graph convolutional neural network adopts the GMM graph convolutional neural network and comprises one input layer, five intermediate layers and one output layer, the input layer and the intermediate layers adopt Gaussian bias linear units as activation functions, and the output layer does not use an activation function; The graph convolutional neural network is trained by taking the plurality of groups of design variables, the plurality of groups of measurement point data and the graph data as inputs and taking the plurality of second stress fields as outputs; When the convolutional neural network and the graph convolutional neural network are trained, the network parameters are optimized by minimizing a sub-loss function, and the sub-loss function is specifically as follows: The stress field reconstruction model is overall trained by a total loss function, and the total loss function is specifically as follows: The measurement point data comprises temperature data and displacement data.

5. A method of reconstructing a stress field in a gas turbine turbine blade root wheel slot according to claim 1, wherein, The method comprises the following steps: L T = L F + L S + L gx + L gy ; where L T is a sub-loss function, L F is a loss function for the global stress field, L S is a loss function for the field maximum, L gx and L gy are loss functions for the global stress field gradients in the x and y directions.

6. A method of reconstructing a stress field in a gas turbine turbine blade root wheel slot according to claim 1, wherein, ​ L C = L T,1 + L T,2 + L' In the formula, L C is the total loss function, L T,1 is the loss function of the convolutional neural network, L T,2 is the loss function of the graph convolutional neural network, and L' is the loss function of the convolutional neural network and the graph convolutional neural network for the prediction values of the common nodes of the structured grid and the unstructured grid.

7. A method of reconstructing a stress field in a gas turbine turbine blade root wheel slot according to claim 1, wherein ​ 8. A gas turbine turbine blade root wheel slot stress field reconstruction system, characterized by, ​ The collecting module is configured to construct a plurality of finite element models of a turbine blade root wheel groove of a gas turbine under different profile design schemes, and collect a plurality of groups of design variables, a plurality of groups of measurement point data and graph data corresponding to the plurality of finite element models; In the finite element model, a structured grid is used in the profile area of the root wheel groove, and an unstructured grid is used in the internal area of the root; The computing module is configured to perform finite element calculation on the plurality of finite element models to obtain a plurality of first stress fields corresponding to the structured grid and a plurality of second stress fields corresponding to the unstructured grid; The training module is configured to construct a stress field reconstruction model, wherein the stress field reconstruction model comprises a convolutional neural network and a GMM graph convolutional neural network in parallel; The convolutional neural network is trained by taking the plurality of groups of design variables and the plurality of groups of measurement point data as inputs and taking the plurality of first stress fields as outputs; The graph convolutional neural network is trained by taking the plurality of groups of design variables, the plurality of groups of measurement point data and the graph data as inputs and taking the plurality of second stress fields as outputs; The reconstruction module is configured to acquire design variables, measurement point data and graph data of the root wheel groove of the turbine blade of the gas turbine in real time, input the design variables, the measurement point data and the graph data into the trained stress field reconstruction model, and obtain a reconstructed root wheel groove stress field.

Images