A method and system for predicting parameters of a gas turbine turbine blade

Abstract

The application discloses a gas turbine turbine blade parameter prediction method and system, and relates to the technical field of gas turbine strength and aerodynamic design, and comprises the following steps: obtaining a generator and a discriminator, iteratively processing the generator and the discriminator, and obtaining an optimal generator and an optimal discriminator; resampling sampling parameters of a turbine blade model, inputting the resampled data into the optimal generator, and obtaining optimal blade parameter data; training a third graph convolutional neural network through the resampled data and the optimal blade parameter data, and obtaining a physical field prediction network; and predicting the gas turbine turbine blade parameters through the physical field prediction network. The application generates high-quality parameter training data based on high-quality generators and discriminators, can greatly reduce the finite element calculation data volume, greatly reduces the finite element calculation cost, improves the data acquisition efficiency, and speeds up the construction of the physical field prediction network under the premise of ensuring the prediction accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gas turbine strength and aerodynamic design, in particular to a gas turbine turbine blade parameter prediction method and system. BACKGROUND

[0002] As an important component of a gas turbine, the structure and performance of a turbine blade directly affect the working efficiency and service life of the entire equipment. When the turbine blade is in a harsh working environment of high temperature, high pressure and high stress, it needs to withstand complex thermal mechanical loads. Therefore, the design, manufacturing and performance evaluation of the turbine blade are particularly important. In order to ensure the safety and reliability of the equipment and prolong its service life, it is necessary to accurately predict the key parameters of the turbine blade, which is of great significance to optimize the design, improve the efficiency of the gas turbine and reduce the operating cost.

[0003] At present, the prediction method of the turbine blade parameters of the gas turbine mainly adopts artificial intelligence and neural network technology. These technologies are based on a large amount of experimental data and simulation data, and a deep neural network model is trained to quickly and efficiently complete the prediction of the key performance parameters of the blade. Compared with traditional physical models, the neural network model has strong nonlinear fitting ability and can capture the parameter variation law under complex working conditions. In addition, these artificial intelligence-based methods can also combine actual operation data to realize online prediction and adaptive optimization, saving time and computing resources while improving the flexibility and applicability of prediction.

[0004] However, there are still some challenges in training the neural network to predict the turbine blade parameters of the gas turbine. The high performance of the neural network model depends on a large amount of high-quality parameter training data, and these data are currently mainly obtained based on finite element calculation, which is a computationally intensive process that requires a large amount of computing resources and time, thereby making the generation of a large amount of data a bottleneck problem in the prediction process. Large-scale computing requirements not only reduce efficiency, but also limit the rapid iteration and optimization of the neural network model. SUMMARY

[0005] Based on the defects of the prior art, the present application provides a gas turbine turbine blade parameter prediction method and system, which solves the problem that the current data is mainly obtained based on finite element calculation, which is a computationally intensive process that requires a large amount of computing resources and time, thereby making the generation of a large amount of data a bottleneck problem in the prediction process.

[0006] The present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a gas turbine turbine blade parameter prediction method, comprising the following steps:

[0008] A turbine blade model of a gas turbine is constructed, and sampling parameters are specified;

[0009] The sampling parameters of the turbine blade model are sampled once under different conditions to obtain first training data; the turbine blade model is subjected to blade thermal fluid-structure coupling calculation to obtain real data of the blade parameters; the first graph convolutional neural network is trained with the first training data as input and the real data as output to obtain a generator;

[0010] The sampling parameters of the turbine blade model are sampled twice to obtain second training data, which is input into the generator to obtain fake data; the second graph convolutional neural network is trained with the first training data and the real data and the second training data and the fake data to obtain a discriminator;

[0011] The sampling parameters of the turbine blade model are sampled three times to obtain third training data, which is input into the generator to obtain generated data; the third training data and the generated data are input into the discriminator to obtain a probability value; the loss function value of the discriminator is determined based on the probability value, and the generator and the discriminator are iterated through the loss function value to obtain an optimal generator and an optimal discriminator;

[0012] The sampling parameters of the turbine blade model are resampled, and the resampled data is input into the optimal generator to obtain optimal blade parameter data; the third graph convolutional neural network is trained with the resampled data as input and the optimal blade parameter data as output to obtain a physical field prediction network; the turbine blade parameters of the gas turbine are predicted through the physical field prediction network.

[0013] Preferably, the turbine blade model is obtained by modeling the solid domain and the fluid domain of the turbine blade of the gas turbine; the sampling parameters include working condition parameters, processing parameters, and material parameters.

[0014] Preferably, the real data of the blade parameters is obtained by performing blade thermal fluid-structure coupling calculation on the turbine blade model through an aerodynamic batch processing file, and the blade parameters include pressure field and temperature field.

[0015] Preferably, the first graph convolutional neural network is trained with the first training data as input and the real data as output to obtain the generator, including the following steps:

[0016] The turbine blade model is subjected to grid division;

[0017] An adjacency matrix is established through the turbine blade model after grid division;

[0018] The neighborhood node set of each node in the first graph convolutional neural network is obtained through the adjacency matrix;

[0019] For each node, aggregate neighborhood features from its neighborhood nodes to obtain an aggregated feature vector of the current node; the neighborhood features include the first training data and the real data;

[0020] Combine the aggregated feature vector of the current node with the self-feature of the node to obtain an updated feature;

[0021] The first graph convolutional neural network with the updated node feature is used as a generator;

[0022] The output of the generator is as follows:

[0023] Y recon =G(X,Z=0; θ1);

[0024] In the formula, Y is the output of the generator G, recon is the reconstruction, θ1 is the learnable parameter of the generator, X is the first training data, and Z is the noise;

[0025] The reconstruction loss of the generator is as follows:

[0026]

[0027] In the formula, L recon is the reconstruction loss, N is the number of sample points, is the square of the two-norm, Y real，i is the i-th real data, is the i-th generated data;

[0028] The generator parameters are updated by gradient descent.

[0029] Preferably, the second graph convolutional neural network is trained by the first training data and the real data and the second training data and the fake data, comprising the following steps:

[0030] The first training data and the real data are labeled with true label 1, and the second training data and the fake data are labeled with fake label 0;

[0031] The first training data, the real data and the true label 1, and the second training data, the fake data and the fake label 0 are input into the second graph convolutional neural network;

[0032] The node feature H (L) is obtained from the second graph convolutional network, and the global feature H graph is obtained by average pooling:

[0033] H graph =meanpool(H (L) );

[0034] Average pooling is performed on each feature dimension:

[0035]

[0036] where h N(v)，k is the representation of node v in the k-th feature dimension after pooling, |N(v)| is the size of the neighborhood of node v, h u,k is the k-th feature value of node u, d is the number of dimensions, and N(v) is the neighborhood, which contains the set of all nodes directly connected to node v.

[0037] The output probability is calculated using an activation function.

[0038] The trained second graph convolutional neural network is used as a discriminator.

[0039] The discriminator's discrimination output for real data is as follows:

[0040] D(Y real )=D(X,Y real ; θ2).

[0041] where D(Y real ) is the discriminator's output for real data Y real , and θ2 is the discriminator's learnable parameters.

[0042] The discriminator's discrimination output for fake data is as follows:

[0043] D(Y real )=D(G(X, Z=0; θ1); θ2).

[0044] where D(Y recon ) is the discriminator's output for fake data .

[0045] Preferably, the loss function of the generator is as follows:

[0046] L gen =λ recon L recon +λ G L G .

[0047] where L gen is the total generator loss, λ recon and λ G are hyperparameters, and L G is the adversarial loss.

[0048] The loss function of the discriminator is as follows:

[0049]

[0050] where L DFor the loss of the discriminator, D(X i , Y i ) is the output probability of the discriminator, y i is the label of the discriminator, X i is the i-th third sampling parameter, Y i is the i-th generated data.

[0051] The loss function value is used to iterate the generator and the discriminator, and specifically includes the following steps:

[0052] The loss function value is compared with a target threshold value;

[0053] When the loss function value is greater than the target threshold value, the generator and the discriminator are iterated based on the Adam optimizer, otherwise the optimal learning parameters of the generator and the discriminator are output.

[0054] Preferably, the third graph convolutional neural network is trained, specifically including the following steps:

[0055] The resampled data, the optimal blade parameter data, the first training data, and the real data of the blade parameter are used as a training set, wherein the training set is as follows:

[0056] (X+X, Y+Y real )→(X new , Y new );

[0057] In the formula, is the resampled data, X is the first training data, is the optimal blade parameter data, Y real is the real data, X new is the input of the training set, Y new is the output of the training set;

[0058] The training is performed by the following formula:

[0059] Y new = F(X new ; θ3);

[0060] In the formula, is the physical field prediction network, and θ3 is the learnable parameter of the physical field prediction network.

[0061] In a second aspect, the present application provides a gas turbine turbine blade parameter prediction system, comprising:

[0062] A construction module is used to construct a turbine blade model of a gas turbine and specify a sampling parameter;

[0063] The generator training module is configured to sample the sampling parameters of the turbine blade model under different conditions to obtain first training data, perform blade thermal fluid-structure coupling calculation on the turbine blade model to obtain real data of the blade parameters, train the first graph convolutional neural network by taking the first training data as input and the real data as output, and obtain a generator;

[0064] The discriminator training module is configured to perform secondary sampling on the sampling parameters of the turbine blade model to obtain second training data, input the second training data into the generator to obtain fake data, and train the second graph convolutional neural network by taking the first training data and the real data and the second training data and the fake data as input to obtain a discriminator.

[0065] The iteration module is configured to perform tertiary sampling on the sampling parameters of the turbine blade model to obtain third training data, input the third training data into the generator to obtain generated data, input the third training data and the generated data into the discriminator to obtain a probability value, determine a loss function value of the discriminator based on the probability value, and iteratively train the generator and the discriminator based on the loss function value to obtain an optimal generator and an optimal discriminator.

[0066] The prediction module is configured to resample the sampling parameters of the turbine blade model, input the resampled data into the optimal generator to obtain optimal blade parameter data, train a third graph convolutional neural network by taking the resampled data as input and the optimal blade parameter data as output, and obtain a physical field prediction network, and predict the turbine blade parameters of the gas turbine by using the physical field prediction network.

[0067] Compared with the prior art, the above at least one technical solution of the present application can achieve the following beneficial effects:

[0068] The application firstly samples the parameter data of the turbine blade model under different conditions for multiple times, and simultaneously performs the blade thermal fluid-solid coupling calculation corresponding to the sampling data to obtain the real data of the blade parameters. The first graph convolutional neural network is trained by the first sampling parameter data and the blade parameter data to obtain a generator. The second graph convolutional neural network is trained by the second sampling data and the data generated by the generator to obtain a discriminator. The constructed generator and discriminator are further optimized by the third sampling data to obtain an optimal generator and an optimal discriminator. Finally, the required data amount of the physical field prediction network is quickly generated by the optimal generator and the optimal discriminator. In the whole training process of the generator and the discriminator, only a small amount of calculation of the blade thermal fluid-solid coupling is required, and the high-quality generator and discriminator are generated based on the calculation results of the thermal fluid-solid coupling and the sampling data. The high-quality parameter training data is generated based on the high-quality generator and discriminator, which can greatly reduce the finite element calculation data amount, greatly reduce the finite element calculation cost, improve the data acquisition efficiency, and accelerate the construction of the physical field prediction network under the premise of ensuring the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0070] Figure 1 A flow chart of a gas turbine turbine blade parameter prediction method of the present application;

[0071] Figure 2 A physical field prediction network construction process schematic diagram of the present application;

[0072] Figure 3 A generator construction process schematic diagram of the present application;

[0073] Figure 4 A discriminator construction process schematic diagram of the present application. DETAILED DESCRIPTION

[0074] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only constitute some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0075] REFERENCE Figure 1The embodiment of the application provides a gas turbine turbine blade parameter prediction method, and specifically comprises the following steps:

[0076] S1: modeling a gas turbine turbine blade solid and fluid to obtain a turbine blade model, including a turbine blade solid domain and a flow channel region through which fluid flows through the turbine blade, and performing grid division on the turbine blade model to obtain a plurality of grid nodes.

[0077] S2: establishing a boundary condition.

[0078] Under different conditions, the turbine blade model is sampled and blade thermal fluid-structure coupling calculation is performed to obtain a plurality of sample points and corresponding boundary conditions (blade parameter data), that is, first training data.

[0079] The sampling parameters are generally related parameters capable of determining blade parameter changes and can be set in finite element calculation. The sampling parameters and the variation range of the parameters are specified to determine the boundary conditions of calculation. The variation range is given according to the actual situation. The sampling parameters include working condition parameters, processing parameters and material parameters. The working condition parameters include main flow gas inlet total temperature, main flow gas inlet total pressure, single flow channel outlet flow, cooling gas inlet total temperature, single flow channel cooling gas flow and rotating speed, etc. The processing parameters include blade inlet absolute flow angle. The material parameters include blade material specific heat capacity, blade material thermal conductivity, blade material density, blade material linear expansion coefficient and blade material elastic modulus, etc.

[0080] The Latin hypercube method is used to sample the sampling parameters under different conditions (different working condition conditions, different processing conditions and different material conditions).

[0081] Supposing that the generated sample points are {x (1) ,x (2) ,...,x (N)}, x is a sampling parameter sample, N is the number of sample points, and each point d is the dimension, and one dimension represents one parameter type. It is ensured that the projection of each dimension i satisfies:

[0082]

[0083] wherein, is the i-th dimension sample point in the k-th sample, is the lower limit of the i-th dimension, that is, the minimum value of the sample point, Δx i is the length of the i-th dimension subspace, and the calculation method is as follows: is a randomly shuffled interval index, different k are not repeated with each other, and are used to divide the i-th dimension subinterval, and the interval is defined as: Each sub-interval corresponds to an index, from 1 to N. That is, when , it is the first sub-interval; when , it is the last sub-interval. Combining the N groups of intervals of all dimensions generates N sample points.

[0084] The node attribute matrix is established. The blade thermal fluid-solid coupling calculation of the turbine blade model is carried out by using the aerodynamic batch processing file. The blade thermal fluid-solid coupling calculation is a finite element calculation, and the node attribute matrix, i.e. the gas turbine turbine blade parameter data, including pressure field and temperature field data, etc. is calculated.

[0085] The adjacency matrix is established. For unstructured grids, the connection relationship can be expressed as G:

[0086] G={N,E} (2);

[0087] Wherein, N is the set of grid nodes, each node represents a calculation unit, which is input into the graph convolutional neural network to represent the point on the adjacency matrix, the node of the adjacency matrix corresponds to the grid node, and E is the connection relationship between the grid nodes.

[0088] The input definition of the adjacency matrix A in the graph convolution network is:

[0089]

[0090] S3: Reconstruction training:

[0091] Referring to Figure 3 , the above established training set is used to train the training set, including boundary conditions, node attribute matrix and adjacency matrix. The graph convolutional neural network (GCN) in the generator is trained. The GCN network adopts SAGEConv convolutional layer, and the aggregation and update process can be expressed as two steps:

[0092] First step: neighborhood feature aggregation.

[0093] The neighborhood node set N(v) of node v is obtained through the adjacency matrix, N(v)={u|A νu =1}.

[0094] The neighborhood feature is aggregated from the neighborhood N(v) of node v:

[0095]

[0096] Wherein, represents the feature vector related to node v and all its neighborhood nodes N(v) or represents denotes the value or feature generated at the l-th layer associated with node u, including boundary conditions and node attributes, u is a node adjacent to node v, and the aggregation function AGG describes how to aggregate the features of the neighborhood N(v), which can be implemented by the mean aggregation as shown below:

[0097]

[0098] Second step: center node feature update.

[0099] The aggregated neighborhood features are combined with the own features of the center node to generate the updated features. The expression is as follows:

[0100]

[0101] wherein, denotes the linear transformation weight of the node's own features, denotes the linear transformation weight of the neighborhood aggregated information, denotes the feature representation of the l-th layer node v, and the nonlinear activation function σ(·) is GeLU, which is mathematically defined as:

[0102]

[0103] In this stage, the input noise Z is zero, and the output is generated only depending on the sample point X:

[0104] Y recon = G(X, Z = 0; θ1) (8);

[0105] wherein Y denotes the output of the generator G, wherein θ1 is the learnable parameter of the generator, and recon denotes reconstruction.

[0106] The reconstruction loss is expressed by the mean square error (MSE):

[0107]

[0108] wherein, denotes the square of the two-norm.

[0109] The generator parameter θ1 is updated by gradient descent:

[0110]

[0111] wherein η1 is the generator learning rate.

[0112] Through the training of the reconstruction loss, the generator learns to map the parameter X to the temperature field Y (1) , the pressure field Y (2)The basic mapping relationship is learned from the sampling results and the adjacency matrix to the temperature field and the pressure field.

[0113] S4: The Latin hypercube method and the given range are used for secondary sampling of the sampling parameters to generate a new parameter set X. The corresponding pressure field and temperature field parameters are obtained by inputting into the generator. fake , which can better approach the task target. This data is called fake data.

[0114] S5: In the present application, the discriminator is also composed of a graph convolutional neural network, and the main purpose is to determine the authenticity probability of the data. The data Y generated by the generator is used as the training set. fake with label 0 (fake data), and the training set data Y real with label 1 (true data).

[0115] D(X,Y real )→1 (11);

[0116] D(X,Y fake )→0 (12);

[0117] S6: Refer to Figure 4 , the discriminator is trained. The network of the discriminator still adopts the SAGEConv convolutional layer, which is a binary classification model. The activation function adopts GeLU. The output of the final layer is a single probability value, which is used to represent the authenticity of the input data. The last layer node feature H (L) is generated from the graph convolutional network, and the graph feature H graph is obtained by average pooling:

[0118] H graph =meanpool(H (L) ) (13);

[0119] The average pooling is specific to each feature dimension:

[0120]

[0121] Where |N(v)| represents the neighborhood size of node v, and h u,k represents the k-th feature value of node u.

[0122] The final layer adopts the activation function (Sigmoid) to calculate the output probability:

[0123] D(G,X)=σ(W out H graph +b out ) (15);

[0124] where: W out is the output weight, b out is the bias.

[0125] The activation function Sigmoid is expressed as:

[0126]

[0127] where x e R is an arbitrary real number input, e -x is the natural exponential function. The Sigmoid function maps the input from -∞ to ∞ to a continuous value in the interval (0, 1).

[0128] The discriminative output of the discriminator for real data is shown as follows:

[0129] D(Y real ) = D(X, Y real ; θ2) (17);

[0130] where D(Y real ) is the output of the discriminator for real data Y real , and θ2 is the learnable parameter of the discriminator;

[0131] The discriminative output of the discriminator for fake data is shown as follows:

[0132] D(Y recon ) = D(G(X, Z = 0; θ1); θ2) (18);

[0133] where D(Y recon ) is the output of the discriminator for fake data .

[0134] S7: Determine the loss function. The loss function L D (θ2) of the discriminator is:

[0135] The loss function L D of the discriminator is:

[0136]

[0137] where D(X i , Y i ) represents the output probability of the discriminator for sample i, and y i is the label of the discriminator, with 1 for real samples and 0 for fake samples.

[0138] The discriminator parameter θ2 is updated by gradient descent as follows:

[0139]

[0140] wherein η2 is the discriminator learning rate.

[0141] The convergence condition is:

[0142] wherein, is a hyperparameter, representing the loss function target value. Different loss functions have different target values as the convergence condition.

[0143] The adversarial loss L G of the generator is:

[0144]

[0145] The total generator loss L gen is:

[0146] L gen = λ recon L recon + λ G L G (22);

[0147] wherein λ recon and λ G are hyperparameters, respectively controlling the weights of the reconstruction loss and the adversarial loss.

[0148] The optimization objectives of the generator and the discriminator are to minimize their own loss functions L gen and L D , respectively. Both are trained using the Adam optimizer:

[0149]

[0150] S8: After the above training steps, the learnable parameters θ1 and θ2 of the generator and the discriminator are constantly updated, gradually improving their abilities. Until the quality of the samples generated by the generator is high enough, and the adversarial network reaches an approximately converged state.

[0151] S9: Fix the learnable parameters θ1 of the generator, and use the trained network to create a number of data sets, called high-quality fake data (X, Y).

[0152] S10: Refer to Figure 2 , combine the high-quality fake data set (X, Y) with the original training set (X, Y real ) to form the input of the graph prediction network, and train the physical field prediction network.

[0153] The training set is represented as:

[0154] (X+X, Y+Y real )→(X new , Y new ) (25);

[0155] The training mode is represented as:

[0156] Y new = F(X new ; θ3) (26);

[0157] Wherein, θ3 is a learnable parameter of the physical field prediction network.

[0158] The network adopts SAGEConv convolutional layer, and adopts GELU activation function in each layer except the last layer, and adopts SAGEConv output graph convolution to predict the physical field in the last layer.

[0159] The loss function adopts MSELoss:

[0160]

[0161] Wherein, Y new,i represents the target value of the i-th sample, Y i represents the predicted value of the i-th sample.

[0162] The Adam optimizer is used to optimize the loss function to minimize the numerical value.

[0163] S11: After determining the good parameter θ3, the physical field prediction network F is obtained, which has a prediction function.

[0164] Based on the same idea, the application also provides a gas turbine turbine blade parameter prediction system, comprising a construction module, a generator training module, a discriminator training module, an iteration module and a prediction module.

[0165] The construction module is used to construct a turbine blade model of a gas turbine and specify sampling parameters.

[0166] The generator training module is used to sample the sampling parameters of the turbine blade model under different conditions to obtain first training data; perform blade thermal fluid-structure coupling calculation on the turbine blade model to obtain real data of the blade parameters; train the first graph convolutional neural network with the first training data as input and the real data as output to obtain the generator.

[0167] The discriminator training module is used to perform secondary sampling on the sampling parameters of the turbine blade model to obtain second training data, input the second training data into the generator to obtain fake data, and train the second graph convolutional neural network through the first training data and the real data and the second training data and the fake data to obtain the discriminator.

[0168] The iteration module is configured to perform three times sampling on the sampling parameters of the turbine blade model to obtain third training data, input the third training data into the generator to obtain generated data, input the third training data and the generated data into the discriminator to obtain a probability value, determine a loss function value of the discriminator based on the probability value, and perform iteration on the generator and the discriminator through the loss function value to obtain an optimal generator and an optimal discriminator.

[0169] The prediction module is configured to perform resampling on the sampling parameters of the turbine blade model, input the resampling data into the optimal generator to obtain optimal blade parameter data, train the third graph convolutional neural network with the resampling data as input and the optimal blade parameter data as output to obtain a physical field prediction network, and predict the parameters of the turbine blade of the gas turbine through the physical field prediction network.

[0170] Although preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to these embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including all the preferred embodiments and all the changes and modifications falling within the scope of the present application.

[0171] Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and changes of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and changes.

Claims

1. A method for predicting parameters of a gas turbine blade, characterized in that, Includes the following steps: Construct a turbine blade model of a gas turbine and specify the sampling parameters; The sampling parameters of the turbine blade model were sampled once under different conditions to obtain the first training data; Perform thermal-fluid-structure interaction calculations on the turbine blade model to obtain the actual data of the blade parameters; Using the first training data as input and the real data as output, the first graph convolutional neural network is trained to obtain the generator; The sampling parameters of the turbine blade model are sampled a second time to obtain the second training data. The second training data is then input into the generator to obtain the fake data. The second graph convolutional neural network is trained using the first training data and real data, as well as the second training data and fake data, to obtain a discriminator; The sampling parameters of the turbine blade model are sampled three times to obtain the third training data. The third training data is then input into the generator to obtain the generated data. The third training data and the generated data are input into the discriminator to obtain probability values; the loss function value of the discriminator is determined based on the probability values, and the generator and discriminator are iterated through the loss function value to obtain the optimal generator and the optimal discriminator; The sampling parameters of the turbine blade model are resampled, and the resampled data is input into the optimal generator to obtain the optimal blade parameter data; Using resampled data as input and optimal blade parameter data as output, the convolutional neural network in the third graph is trained to obtain the physical field prediction network; the physical field prediction network is used to predict the parameters of the gas turbine blades. The process of training a first graph convolutional neural network using first training data as input and real data as output to obtain a generator includes the following steps: Mesh the turbine blade model; An adjacency matrix is ​​established using the turbine blade model after mesh generation; Obtain the set of neighboring nodes for each node in the first graph convolutional neural network using the adjacency matrix; For each node, neighborhood features are aggregated from its neighboring nodes to obtain the aggregated feature vector of the current node; the neighborhood features include the first training data and the real data; The aggregated feature vector of the current node is combined with the node's own features to obtain the updated features; The first graph convolutional neural network, updated with node features, is used as the generator. The generator's output is shown below: Y recon = G(X, Z = 0; θ1); In the formula, Y is the output of generator G, recon is the reconstruction, θ1 is the learnable parameter of generator, X is the first training data, and Z is noise; The generator's reconstruction loss is shown below: In the formula, L recon This is the reconstruction loss, where N is the number of sample points. It is the square of the L2 norm, Y real，i It is the i-th real data. It is the i-th generated data; The generator parameters are updated using gradient descent.

2. The method for predicting parameters of a gas turbine blade as described in claim 1, characterized in that, The solid and fluid domains of the gas turbine blades are modeled to obtain the turbine blade model; the sampling parameters include operating parameters, processing parameters and material parameters.

3. The method for predicting parameters of a gas turbine blade as described in claim 1, characterized in that, By performing thermal-fluid-structure interaction calculations on the turbine blade model using a pneumatic batch processing file, the actual data of the blade parameters, including the pressure field and temperature field, are obtained.

4. The method for predicting parameters of a gas turbine blade as described in claim 1, characterized in that, The training of the second graph convolutional neural network using first training data and real data, and second training data and fake data, includes the following steps: Create a real label 1 for the first training data and real data, and create a fake label 0 for the second training data and fake data; The first training data, real data, and true label 1, as well as the second training data, fake data, and pseudo label 0, are input into the second graph convolutional neural network. Obtain node features H from the second graph convolutional network (L) Obtain the global graph feature H by average pooling graph : H graph =meanpool(H (L) ); Average pooling across each feature dimension: In the formula, h N(v),k It is the representation of node v in the k-th feature dimension after pooling, |N(v)| is the neighborhood size of node v, and h u,k Let d be the feature value of node u in the k-th dimension, d be the number of dimensions, and N(ν) be the neighborhood, which contains the set of all nodes directly connected to node v. The output probability is calculated using an activation function; The trained second graph convolutional neural network is used as the discriminator; The discriminator's output for judging real data is shown below: D(Y real )=D(X,Y real ;θ2); In the formula, D(Y) real () is the discriminator for the real data Y real The output of θ2 is the learnable parameter of the discriminator; The discriminator's output for identifying forged data is shown below: D(Y recon )=D(G(X,Z=0;θ1);θ2); In the formula, D(Y) recon (This is for the discriminator to detect forged data) The output of .

5. The method for predicting parameters of a gas turbine blade as described in claim 4, characterized in that, The loss function of the generator is as follows: L gen =λ recon L recon +λ G L G ; In the formula, L gen For the total generator loss, λ recon and λ G It's a hyperparameter, L G It is about combating losses; The loss function of the discriminator is as follows: In the formula, L D For the discriminator's loss, D(X) i ,Y i y is the output probability of the discriminator. i X is the label for the discriminator. i Y is the i-th third sampling parameter. i Generate data for the i-th data; The iteration of the generator and discriminator using the loss function value specifically includes the following steps: The loss function value is compared with the target threshold; When the loss function value is greater than the target threshold, the generator and discriminator are iterated based on the Adam optimizer; otherwise, the optimal learning parameters of the generator and discriminator are output.

6. The method for predicting parameters of a gas turbine blade as described in claim 1, characterized in that, The training of the third graph convolutional neural network specifically includes the following steps: The resampled data, the optimal blade parameter data, the first training data, and the actual blade parameter data are used as the training set, as shown below: (X+X,Y+Y real )→(X new ,AND new ); In the formula, For resampled data, X is the first training data. For the optimal blade parameter data, Y real For real data, X new The input to the training set, Y new The output of the training set; Training can be performed using the following formula: Y new =F(X new ;θ3); In the formula, θ3 is the physics prediction network, and θ3 is the learnable parameter of the physics prediction network.

7. A gas turbine blade parameter prediction system, characterized in that, include: The building module is used to build a turbine blade model of a gas turbine and specify sampling parameters; The generator training module is used to sample the turbine blade model's parameters under different conditions to obtain the first training data. Perform thermal-fluid-structure interaction calculations on the turbine blade model to obtain the actual data of the blade parameters; Using the first training data as input and the real data as output, the first graph convolutional neural network is trained to obtain the generator; The discriminator training module is used to perform secondary sampling on the sampling parameters of the turbine blade model to obtain second training data. The second training data is then input into the generator to obtain fake data. The second graph convolutional neural network is trained using the first training data and real data, as well as the second training data and fake data, to obtain the discrimination. The iteration module is used to sample the parameters of the turbine blade model three times to obtain the third training data. The third training data is then input into the generator to obtain the generated data. The third training data and the generated data are input into the discriminator to obtain probability values; The loss function value of the discriminator is determined based on the probability value. The generator and discriminator are then iterated through the loss function value to obtain the optimal generator and the optimal discriminator. The prediction module is used to resample the sampled parameters of the turbine blade model and input the resampled data into the optimal generator to obtain the optimal blade parameter data. Using resampled data as input and optimal blade parameter data as output, the convolutional neural network in the third graph is trained to obtain the physical field prediction network; the physical field prediction network is used to predict the parameters of the gas turbine blades. The process of training a first graph convolutional neural network using first training data as input and real data as output to obtain a generator includes the following steps: Mesh the turbine blade model; An adjacency matrix is ​​established using the turbine blade model after mesh generation; Obtain the set of neighboring nodes for each node in the first graph convolutional neural network using the adjacency matrix; For each node, neighborhood features are aggregated from its neighboring nodes to obtain the aggregated feature vector of the current node; the neighborhood features include the first training data and the real data; The aggregated feature vector of the current node is combined with the node's own features to obtain the updated features; The first graph convolutional neural network, updated with node features, is used as the generator. The generator's output is shown below: Y recon =G(X,Z=0;θ1); In the formula, Y is the output of generator G, recon is the reconstruction, θ1 is the learnable parameter of generator, X is the first training data, and Z is noise; The generator's reconstruction loss is shown below: In the formula, L recon This is the reconstruction loss, where N is the number of sample points. It is the square of the L2 norm, Y real，i It is the i-th real data. It is the i-th generated data; The generator parameters are updated using gradient descent.

Images