Generative gas film cooling efficiency two-dimensional distribution prediction method based on prior knowledge embedding

Abstract

The invention provides a generative gas film cooling efficiency two-dimensional distribution prediction method based on prior knowledge embedding, and the method comprises the steps: learning a mapping relation between an input parameter and a coefficient through multiple layers of neurons, carrying out the dimension rising through combining a function equation set, and reducing the learning amount explosion caused by the increase of the information amount due to the dimension rising; and the information extracted by the feature extraction module and the deep information extracted by the image coding module are fused, and then the multi-head attention mechanism is combined, so that the prediction precision of the model on different input parameters is improved. According to the method, dimension rising in the generative learning process is achieved in a priori knowledge embedding mode, so that the huge learning amount and the model complexity needed by the model in the dimension rising learning process are reduced, and the learning amount needs a large number of training samples to complement in a traditional model. Therefore, the method provided by the invention enables the model training difficulty to be lower, enables the model to capture more information under the condition of limited sample quantity, and improves the training precision and the training speed.

Description

Technical Field

[0001] The present invention relates to the field of air film cooling and artificial intelligence technology, and in particular to a generative air film cooling efficiency two-dimensional distribution prediction method based on prior knowledge embedding. Background Art

[0002] Film cooling is a key technology widely used in high-temperature components of aircraft engines, and plays an important role in the thermal protection of combustion chamber walls, turbine blades and nozzles. In the design of film cooling structures, film cooling efficiency and its distribution uniformity are extremely important performance indicators. If the film cooling efficiency is low and the cooling is uneven, it will bring many adverse effects:

[0003] 1: Increase the amount of cooling gas used;

[0004] 2: Uneven film cooling will cause uneven temperature distribution in the combustion chamber, resulting in increased combustion instability, thereby generating more unburned hydrocarbons (UHC) and carbon monoxide (CO), as well as other pollutants (such as NOx);

[0005] 3: Long-term uneven local cooling will accelerate the wear of engine blades and hot end components, resulting in more frequent blade replacement and maintenance, and in severe cases even leading to component ablation and engine failure.

[0006] Therefore, accurate prediction of film cooling efficiency can help designers quickly iterate and optimize cooling solutions, guide the layout and improvement of cooling structures in the early stages of design, thereby improving the cooling methods of high-temperature components and increasing the iterative efficiency and speed of cooling design.

[0007] At present, the methods for predicting two-dimensional film cooling based on one-dimensional vector parameters mainly include empirical fitting methods and artificial intelligence methods.

[0008] The empirical fitting method is to use the fitting empirical correlation to fit the effectiveness and standard deviation on the center line, so as to obtain the two-dimensional distribution results of the effectiveness. However, the analytical form of the two-dimensional distribution equation of the effectiveness introduces many model parameters and customized intermediate variables. The model complexity is too high, and the parameters and intermediate variables need to be calibrated through experimental data, which is difficult to implement and greatly limits the scope of application of this method.

[0009] The artificial intelligence method is to make predictions through artificial neural network models. For example, the Chinese patent application "AI prediction method for two-dimensional distribution of single-hole air film cooling efficiency based on bell curve" with publication number CN112084597A proposed a two-dimensional cooling efficiency prediction method based on bell curve, which builds a fully connected neural network to learn the mapping relationship between the input parameters and the bell curve coefficients. However, this method ignores the spatial correlation of adjacent pixels in the air film distribution, and the method strictly requires that the distribution of cooling efficiency is in the form of a bell curve, which obviously limits the scope of application of this method. For example, the document "Research on prediction of single-hole adiabatic air film cooling efficiency distribution based on machine learning [J]. Advance Technology, 2022, 43 (4): 218-228" uses the characteristics of neural network convolution to introduce the spatial correlation between vector pixels, calculates a large amount of sample data as the network input data set through CFD, and directly builds multiple fully connected layers to expand the length of the vector, and then reconstructs it into a two-dimensional matrix for multiple upsampling, and restores the scale through convolution. In this way, the mapping training and learning between low dimensions and high dimensions is realized. However, the research conclusions of the above two methods show that the network has a strong dependence on the quality and quantity of sample data. The sample data volumes of the above two methods are 1300 and 3900 respectively, and both are CFD calculation samples.

[0010] This leads to two key issues: First, many scholars have shown that the accuracy of CFD calculations of film cooling efficiency is questionable, which puts a test on the quality of the data set. Second, the current two-dimensional generative learning itself requires a large amount of learning, and requires enough samples to achieve accurate predictions. In order to ensure the quality of the data set, it costs a lot to obtain thousands of samples through experimental testing.

[0011] Therefore, how to accurately predict the two-dimensional distribution of film cooling efficiency under the condition of a small number of test samples is a key problem that the present invention needs to solve. Summary of the invention

[0012] Technical issues solved:

[0013] The film cooling efficiency prediction model established by the existing method mainly directly constructs the mapping relationship between one-dimensional input and two-dimensional output through multiple convolutions. Since a large amount of learning is required to generate the two-dimensional film cooling efficiency distribution, this type of model has a great demand for the amount of training data. However, the increase in the amount of data will undoubtedly increase the cost of data acquisition, and the quality of the film cooling efficiency data obtained by the CFD method is questionable, resulting in the data being obtained only through experimental testing, which undoubtedly greatly increases the cost and difficulty of data acquisition. In addition, the existing network model has many convolution kernels and neurons, and many weight parameters need to be trained, which also requires more training time and parameter adjustment cycle.

[0014] Technical solution:

[0015] In order to solve the problems existing in the prior art, the present invention proposes a generative film cooling efficiency two-dimensional distribution prediction method based on prior knowledge embedding, establishes a neural network model combined with prior knowledge embedding, realizes the dimensionality upgrade process from one dimension to two dimensions in the model through prior knowledge, and reduces the amount of training required for the model; and improves the prediction accuracy and generalization ability of the model through image encoding and decoding modules, so as to achieve the purpose of accurately predicting the two-dimensional film cooling efficiency distribution by one-dimensional parameter input even with a small sample size.

[0016] The technical solution of the present invention is:

[0017] The method for predicting the two-dimensional distribution of the generative film cooling efficiency based on prior knowledge embedding comprises the following steps:

[0018] Step 1: Get training samples through experimental testing:

[0019] Each set of sample data includes a film hole parameter vector composed of structural parameters and aerodynamic parameters and a two-dimensional distribution image of film cooling efficiency obtained by experimental testing of the corresponding film hole;

[0020] Step 2: Build the network model:

[0021] The network model input is the air film hole parameter vector, and the output is the air film cooling efficiency two-dimensional distribution image;

[0022] The network model includes a priori knowledge embedding module, a feature extraction module, an image encoding and decoding module, and an output module:

[0023] The input of the prior knowledge embedding module is the air film hole parameter vector, which is mapped to the coefficient A 1 , A 2 , A 3 , B 1 , B 2 , B 3 On; According to coefficient A1 , A 2 , A 3 , B 1 , B 2 , B 3 , using the function

[0024] F(x)=A 1 ·log(A 2 ·x)+A 3

[0025] Q(x)=B 1 (1-x) 2 +B 2 ·2·x·(1-x)+B 3 ·x 2

[0026]

[0027] First, the coefficients F(x) and Q(x) are calculated, and then the two-dimensional distribution cloud map of the film cooling efficiency is restored by dimensionality increase, where x represents the coordinates of each point in the two-dimensional distribution image of the film cooling efficiency along the flow direction, y represents the coordinates of each point in the two-dimensional distribution image of the film cooling efficiency along the span direction, G(y, F(x), Q(x)) is the cooling efficiency value at the (x, y) coordinate point, and after passing through the convolution block, a matrix with the same size as the model output image is obtained as the input of the image encoding module;

[0028] The feature extraction module performs high-dimensional extraction on the input air film hole parameter vector, and reconstructs the extracted high-dimensional features into a two-dimensional matrix, and then performs channel expansion through a convolution block to obtain a matrix with a size consistent with the image encoding output;

[0029] The image encoding and decoding module downsamples the matrix output by the prior knowledge embedding module N times to obtain information at different levels of the image, and the features extracted at each layer are spliced ​​with the information at the same level after decoding; the matrix obtained after the last downsampling is fused with the high-dimensional features extracted by the feature extraction module and flattened into a one-dimensional vector, and then reconstructed through a multi-head attention mechanism to capture the key information in the image and vector features. After obtaining the key information, it is gradually upsampled and restored to the size of the image, which is provided to the output module to obtain the final two-dimensional distribution output of the film cooling efficiency;

[0030] Step 3: Use the sample data obtained in step 1 to train the network model constructed in step 2 to obtain a trained prediction model;

[0031] Step 4: For the film cooling scheme to be predicted, obtain the film hole parameter vector, input the trained prediction model obtained in step 3, and obtain the two-dimensional distribution prediction result of the film cooling efficiency.

[0032] In a further preferred embodiment, the structural parameters include: the hole inclination angle α between the axis of the air film hole and the outlet plane; the expansion angle β of the air film hole expansion section; the forward inclination angle γ of the air film hole expansion section; the inner fold angle δ of the V-shaped hole outlet; the aspect ratio L / d of the air film hole; the ratio of the width to the diameter of the air film hole outlet L w / d; Air film hole inlet and outlet area ratio A out / A in ; The dimensionless pore spacing P / d of the air film pores.

[0033] In a further preferred embodiment, the aerodynamic parameters include: blowing ratio M; density ratio DR; turbulence Tu and acceleration factor k.

[0034] In a further preferred embodiment, the test is carried out by the following test system:

[0035] The main flow is generated from the blower, passes through the butterfly valve and the rectifying cavity to produce a uniform inlet flow at the inlet of the test section; then different turbulence intensities are generated at the outlet of the film hole through turbulence grids of different densities to simulate the effects of different turbulence intensities on cooling; a suction slit is set at the front end of the measuring plate to allow the boundary layer to redevelop there, so that the boundary layer thickness at the outlet of the film hole is close to the local boundary layer thickness of the film hole of the actual engine; at the same time, a trip wire is set behind the slit to promote the transition of the boundary layer from laminar flow to turbulent flow; velocity, temperature and pressure measurement points are set above the film hole outlet to monitor the test time in real time The coolant system uses a high-pressure gas tank and a compressor as the gas source. After passing through the valve and the pressure regulating device, it is heated by the heater so that the temperature difference between the primary and secondary flows does not exceed 0.2K. Then, it enters the cold air cavity after the flow is controlled by the Alicat mass flowmeter and is supplied to the air film hole. The temperature and pressure measuring points are set in the cold air cavity to monitor the coolant state in real time. The surface of the measuring plate is sprayed with PSP pressure-sensitive paint, and the corresponding excitation light source and a scientific-grade CCD camera with a filter are used to collect images during the test. After post-processing the image, a two-dimensional distribution image of the air film cooling efficiency is obtained.

[0036] Beneficial Effects

[0037] The generative film cooling efficiency two-dimensional distribution prediction method based on prior knowledge embedding proposed in the present invention has the following advantages:

[0038] 1. Learn the mapping relationship between input parameters and coefficients through multi-layer neurons, and then combine the function equation group to increase the dimension, so as to reduce the learning volume explosion caused by the increase in the amount of information due to the increase in the dimension;

[0039] 2. The information extracted by the feature extraction module is integrated with the deep information extracted by the image encoding module, and then combined with the multi-head attention mechanism to improve the model's prediction accuracy for different input parameters.

[0040] In summary, the present invention achieves dimensionality increase in the generative learning process by embedding prior knowledge, thereby reducing the huge amount of learning and model complexity required by the model in the dimensionality increase learning process, which requires a large number of training samples to supplement in traditional models. Therefore, the method of the present invention makes model training easier, enables the model to capture more information under the condition of a limited number of samples, and improves training accuracy and training speed.

[0041] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0043] Figure 1 : Training sample measurement system;

[0044] Figure 2 : Schematic diagram of air film pore structure parameters;

[0045] Figure 3 : Generative neural network model based on prior knowledge embedding;

[0046] Figure 4 : Schematic diagram of prior knowledge embedding module;

[0047] Figure 5 : Schematic diagram of feature extraction module;

[0048] Figure 6 : Schematic diagram of image encoding and decoding module;

[0049] Figure 7 : Loss convergence curve of the training process;

[0050] Figure 8 : The physical structure of 4 test samples;

[0051] Fig. 9 : Comparison of the true result, the prediction result of the traditional model and the prediction result of the method of the present invention; (a) the true result of sample 1, (a1) the prediction result of the traditional model for sample 1, (a2) the prediction result of the method of the present invention for sample 1; (b) the true result of sample 2, (b1) the prediction result of the traditional model for sample 2, (b2) the prediction result of the method of the present invention for sample 2; (c) the true result of sample 3, (c1) the prediction result of the traditional model for sample 3, (c2) the prediction result of the method of the present invention for sample 3; (d) the true result of sample 4, (d1) the prediction result of the traditional model for sample 4, (d2) the prediction result of the method of the present invention for sample 4;

[0052] Fig.10 : The difference between the prediction results of the traditional model and the prediction results proposed by the present invention and the real cloud map; (a1) the prediction result of the traditional model for sample 1, (a2) the prediction result of the method of the present invention for sample 1; (b1) the prediction result of the traditional model for sample 2, (b2) the prediction result of the method of the present invention for sample 2; (c1) the prediction result of the traditional model for sample 3, (c2) the prediction result of the method of the present invention for sample 3; (d1) the prediction result of the traditional model for sample 4, (d2) the prediction result of the method of the present invention for sample 4;

[0053] Fig.11 : Prediction errors of the traditional model and the prediction model proposed in this invention on four typical samples. DETAILED DESCRIPTION

[0054] Embodiments of the present invention are described in detail below. The embodiments are exemplary and intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0055] In view of the problem that the film cooling efficiency prediction model established by the existing method requires a large amount of training data, which greatly increases the data acquisition cost, and has a large number of model weight parameters, and the model training time and parameter adjustment cycle are long, this embodiment proposes a generative film cooling efficiency two-dimensional distribution prediction method based on prior knowledge embedding. The prior knowledge is used to realize the dimensionality upgrade process from one dimension to two dimensions in the model, which reduces the required training amount of the model, and improves the prediction accuracy and generalization ability of the model through the image encoding and decoding modules, so as to achieve the purpose of accurately predicting the two-dimensional film cooling efficiency distribution by one-dimensional parameter input even when the sample size is small.

[0056] Specifically include the following:

[0057] (I) Obtain training samples:

[0058] In order to obtain high-quality test data, the training samples used in this embodiment are all experimental test results. Figure 1 The experimental system used for sample acquisition in this embodiment is demonstrated.

[0059] Depend on Figure 1It can be seen that the mainstream is generated from the blower, and passes through the butterfly valve and the rectifying cavity to produce a uniform inlet flow at the entrance of the test section. Then, different turbulences are generated at the outlet of the film hole through turbulence grids of different densities to simulate the influence of different turbulences on cooling. A suction slit is set at the front end of the measuring plate (the test surface in the figure) to allow the boundary layer to redevelop there, so that the boundary layer thickness at the outlet of the film hole is close to the local boundary layer thickness of the film hole of the actual engine. At the same time, a trip wire is set after the slit to promote the transition of the boundary layer from laminar flow to turbulent flow, because the outlet of the film hole of the actual engine is often in a turbulent boundary layer. Speed, temperature and pressure measurement points are set above the outlet of the film hole to monitor the mainstream state during the test in real time. The coolant system uses a high-pressure gas tank and a compressor as the gas source. After passing through the valve and the pressure regulating device, it is heated by the heater so that the temperature difference between the main and secondary flow gases does not exceed 0.2K. Then, after the flow is controlled by the Alicat mass flowmeter, it enters the cold air cavity and is supplied to the film hole. The temperature and pressure measurement points are set in the cold air cavity to monitor the coolant state in real time. The surface of the measuring plate is sprayed with PSP pressure-sensitive paint, and equipped with a corresponding excitation light source and a scientific-grade CCD camera with a filter to collect images during the test. After post-processing the images, a two-dimensional distribution image of the film cooling efficiency is obtained.

[0060] Figure 2 The air film hole structure studied in this embodiment is shown. The air film hole structure parameters mainly include the hole inclination angle α between the air film hole axis and the outlet plane, the expansion angle β of the air film hole expansion section, the forward inclination angle γ of the air film hole expansion section, the inner fold angle δ of the V-shaped hole outlet; the length L of the air film hole, the length L of the expansion section 1 , air film hole outlet width L w ; Air film hole entrance area A in , outlet area A out ; and the hole spacing P between two adjacent air film holes.

[0061] Among them, the length parameter is dimensionless using the aperture d, the inlet and outlet are dimensionless using the area ratio, and the inlet and outlet area ratio and the expansion length L 1 They are not independent of each other. Only one of the two needs to be determined to determine the entire hole structure. In this embodiment, the area ratio is used as one of the input items. Therefore, a total of 8 structural parameters are used as input:

[0062] The inclination angle α between the air film hole axis and the outlet plane:

[0063] The expansion angle β of the expansion section of the air film hole, the hole type studied in this embodiment is a bilaterally symmetrical expansion;

[0064] The forward tilt angle γ of the expansion section of the air film hole;

[0065] The inner angle δ of the V-shaped hole outlet, where δ = 0 for the fan-shaped hole;

[0066] The aspect ratio of the air film hole is L / d;

[0067] The ratio of the width and diameter of the air film hole outlet is L w / d;

[0068] Air film hole inlet and outlet area ratio A out / A in ;

[0069] The dimensionless pore spacing P / d of the air film pores.

[0070] Table 1 shows the input structure parameter range of the sample data in this embodiment:

[0071] Table 1 Structural parameter value range

[0072] Tilt angle α(°) 25，30，40 Expansion angle β(°) 0～15 Forward tilt angle γ(°) 0～15 Internal folding angle δ(°) 0～30 Aspect ratio L / d 4～7 <![CDATA[Ratio of outlet width to diameter L w / d]]> 1～3.5 <![CDATA[Area ratio A out / A in > 1.0～8.6 Hole spacing P / d 3～8

[0073] Table 2 shows the range of aerodynamic parameters of the sample group. This embodiment selects the aerodynamic parameters commonly used in studying the efficiency of air film cooling: blowing ratio M = (ρ c u c ) / (ρ m u m ), density ratio DR = ρ c / ρ m , turbulence Tu=u ′ / u m And the acceleration factor k = ν × (du / dx) / u, which measures the magnitude of the pressure gradient 2 As input. Where ρ represents gas density, u represents gas velocity, and subscripts c and m represent secondary flow and mainstream flow, respectively. ′ represents the pulsating velocity of the mainstream, v represents the dynamic viscosity coefficient of the gas, and x represents the coordinates of each point along the flow direction in the two-dimensional distribution image of film cooling efficiency.

[0074] Table 2 Pneumatic parameter value range

[0075] <![CDATA[主流速度u m (m / s)]]> 30 <![CDATA[Primary and secondary stream temperature T m T c (K)]]> 298±0. Blowing ratio M 0.5～3 Density Ratio DR 1，1.5 Turbulence Tu 2，10 Acceleration factor k <![CDATA[-2.0×10 -6 ~2.0×10 -6 ]]>

[0076] In the final training samples, each set of sample data includes a film hole parameter vector composed of 8 structural parameters and 4 aerodynamic parameters and a two-dimensional distribution image of the film cooling efficiency of the corresponding film hole obtained through experimental testing.

[0077] (II) Building a network model

[0078] Figure 3 The generative network model architecture constructed in this embodiment is demonstrated, which takes a one-dimensional vector consisting of 8 structural parameters constituting the air film holes and 4 main aerodynamic parameters as input and takes a two-dimensional distribution of air film cooling efficiency as output.

[0079] The network model mainly includes a prior knowledge embedding module, a feature extraction module, an image encoding and decoding module, and an output module. The model framework is mainly composed of three branches. The first branch is composed of a prior knowledge embedding module and a feature extraction module, which is used to capture and upgrade the input information. The second branch is the image encoding module, which improves the prediction ability of the overall distribution and local correlation of film cooling efficiency by utilizing the spatial recognition characteristics of convolution. The third is the decoding module, which gradually decodes the fused deep information and restores it to the original two-dimensional size.

[0080] The functions and construction framework of each module are introduced in detail below.

[0081] Prior knowledge embedding module:

[0082] like Figure 4 As shown in Figure 1, the prior knowledge embedding module consists of a coefficient mapping module and a learning function module. The coefficient mapping module includes an extraction layer, a hidden layer, and an output layer. The coefficient mapping module first extracts the features of the one-dimensional vector composed of 12 elements of the input through an extraction layer composed of 16 neurons, and then maps it to the coefficient A through an output layer composed of 16 neurons after passing through two hidden layers composed of 32 neurons respectively. 1 , A 2 , A 3 , B 1 , B 2 , B 3 In the learning function module, the 6 coefficients A learned 1 , A 2 , A 3 , B 1 , B 2 , B 3 By combining the attenuation function (Formula 1F(x)) and the diffusion function equation (Formula 2Q(x)), the coefficients F(x) and Q(x) of the Gaussian function (Formula 3G(y, F(x), Q(x))) are obtained, where x and y represent the coordinate values ​​of each point in the two-dimensional distribution image of film cooling efficiency along the streamwise direction and the spanwise direction, respectively:

[0083] F(x)=A 1 ·log(A 2 ·x)+A 3 (1)

[0084] Q(x)=B 1 (1-x) 2 +B 2 2 x (1-x) + B 3 ·x 2 (2)

[0085]

[0086] The cooling efficiency value G(y, F(x), Q(x)) at each (x, y) coordinate point is calculated by formula 3, where x represents the coordinate of each point in the two-dimensional distribution image of film cooling efficiency along the flow direction, and y represents the coordinate of each point in the two-dimensional distribution image of film cooling efficiency along the span direction. The learned six coefficients A1, A2, A3, B1, B2, and B3 (one-dimensional) are upgraded to restore the two-dimensional distribution cloud map of film cooling efficiency. The two-dimensional distribution cloud map of film cooling efficiency is a 512×512 matrix, which is used as the input of the image encoding module.

[0087] The purpose of the prior knowledge embedding module is to learn the mapping relationship between the input vector and the coefficient vector, and to construct a two-dimensional distribution cloud map of film cooling efficiency in combination with prior knowledge.

[0088] Feature extraction module:

[0089] like Figure 5 As shown in the figure, the feature extraction module first extracts the high-dimensional features of the one-dimensional vector composed of 12 elements of the input by a high-dimensional extraction layer with multiple neurons, and then reconstructs the extracted high-dimensional features into the form of a two-dimensional matrix and expands the channels through a 3×3 convolution block to obtain a matrix with the same size as the image encoding output.

[0090] Specifically, the input one-dimensional vector passes through the high-dimensional extraction layer composed of 512 neurons, 1024 neurons and 4096 neurons in turn, and the high-dimensional features are reconstructed into a 64×64 matrix. After the channel is expanded through a 3×3 convolution block, the expanded 64×64 matrix consistent with the image encoding output is obtained.

[0091] The purpose of the feature extraction module is to extract the high-dimensional features of the one-dimensional vector composed of 12 elements of the input, so as to fuse them into the high-dimensional hierarchical information after image encoding.

[0092] Image encoding and decoding module:

[0093] Figure 6The construction logic of the image encoding and decoding module is demonstrated. The 512×512 matrix corresponding to the two-dimensional distribution image of the film cooling efficiency output by the prior knowledge embedding module is first processed by two 3×3 convolution blocks to obtain a 512×512 first matrix. After one downsampling and two 3×3 convolution blocks, the first matrix is ​​obtained to obtain a 256×256 second matrix. After the second downsampling and two 3×3 convolution blocks, the third matrix is ​​obtained to obtain a 128×128 matrix. After the third downsampling and two 3×3 convolution blocks, the fourth matrix is ​​obtained to obtain a 64×64 matrix. The 64×64 fourth matrix is ​​fused with the high-dimensional features extracted by the feature extraction module and flattened into a one-dimensional vector. It is then reconstructed into a 64×64 matrix through the multi-head attention mechanism. 4; the reconstructed 64×64 fifth matrix passes through a 3×3 convolution block and is concatenated with the 64×64 fourth matrix, and then passes through two 3×3 convolution blocks and the first deconvolution to obtain a 128×128 sixth matrix, the sixth matrix is ​​concatenated with the 128×128 third matrix, and then passes through two 3×3 convolution blocks and the second deconvolution to obtain a 256×256 seventh matrix, the seventh matrix is ​​concatenated with the 256×256 second matrix, and then passes through two 3×3 convolution blocks and the third deconvolution to obtain a 512×512 eighth matrix, and the eighth matrix is ​​concatenated with the 512×512 first matrix and output to the output module.

[0094] After obtaining the corresponding matrix of the two-dimensional distribution image of the film cooling efficiency output by the prior knowledge embedding module, the encoding module obtains information at different levels of the image through three downsamplings, and the features extracted at each level are spliced ​​with the information at the same level after decoding. This embodiment integrates the high-dimensional features extracted by the feature extraction module on the basis of the deep information obtained after the third downsampling, and captures the key information in the image and vector features through the Transformer model with a multi-head attention mechanism. After obtaining the key information, it is gradually upsampled and restored to the size of the image, which is provided to the output module to obtain the final two-dimensional distribution output of the film cooling efficiency.

[0095] The purpose of the image encoding and decoding module is to make up for the defects of inaccurate prediction in the overall space and local range of images obtained only through prior knowledge.

[0096] Output Module:

[0097] The output module passes the output matrix of the image encoding and decoding module through two 3×3 convolution blocks to obtain the final two-dimensional distribution output of the film cooling efficiency.

[0098] (III) Model training

[0099] Since it is very costly to obtain test data through experiments, in this embodiment, a total of 200 groups of sample data are obtained, and this sample group is divided into training set, validation set and test set according to the ratio of 7:2:1. The training is carried out on Nvidia RTX4080Super GPU with 16GB video memory, Intel Core i7 14700K CPU platform and 64GB memory. Python is used as programming language, Adam is used as optimizer, and the learning rate decays according to the natural exponential law with the cumulative number of iterations (Formula 4). The loss function adopts a weighted combination of MSE loss and L1 loss (Formula 5). Figure 7 The loss curve of the model training process is shown. It can be seen that the loss of the trained model on the validation set is at the same level as the loss on the training set. And it is basically stable after 400 rounds of training. The total training time is about 928 minutes, and the prediction dissipation of the test sample is about 0.94ms.

[0100]

[0101]

[0102] where a k is the learning rate of the kth step, ψ is the decay coefficient, which is set to 0.05. c is the time scale coefficient for controlling decay, which is set to 100. i is the true value, is the predicted value, and n is the number of matrix elements.

[0103] (IV) Training and testing results

[0104] For the convenience of presentation, the prediction results of four representative test samples are shown here. Table 3 shows the input structural parameters and aerodynamic parameters of these four samples. Figure 8 The physical structures of the four test samples are shown.

[0105] Table 3 Test sample parameters

[0106]

[0107]

[0108] Fig. 9 and Fig.10 The comparison results of the actual results, the prediction results of the traditional model and the prediction results of this embodiment on these four samples are shown. It can be seen that the model proposed in this embodiment significantly improves the prediction accuracy of the overall film cooling efficiency distribution and the prediction accuracy at the local position of the hole outlet.

[0109] Fig.11The prediction errors of the surface average film cooling efficiency of the four samples by the traditional model and the prediction model proposed by the present invention are shown. It can be seen that the prediction error of the prediction model of the present invention is reduced by 4.3% to 41.2% compared with the original traditional model.

[0110] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent of the present invention.

Claims

1. A generative film cooling efficiency two-dimensional distribution prediction method based on prior knowledge embedding, characterized by: The following steps are involved: Step 1: Get training samples through experimental testing: Each set of sample data includes a film hole parameter vector composed of structural parameters and aerodynamic parameters and a two-dimensional distribution image of film cooling efficiency obtained by experimental testing of the corresponding film hole; Step 2: Build the network model: The network model input is the air film hole parameter vector, and the output is the air film cooling efficiency two-dimensional distribution image; The network model includes a priori knowledge embedding module, a feature extraction module, an image encoding and decoding module, and an output module; The input of the prior knowledge embedding module is the air film hole parameter vector, which is mapped to coefficients A1, A2, A3, B1, B2, and B3; according to the coefficients A1, A2, A3, B1, B2, and B3, the function F(x)=A1·log(A2·x)+A3 Q(x)=B1·(1-x) 2 +B2·2·x·(1-x)+B3·x 2 First, the coefficients F(x) and Q(x) are calculated, and then the two-dimensional distribution cloud map of the film cooling efficiency is restored by dimensionality increase, where x represents the coordinates of each point in the two-dimensional distribution image of the film cooling efficiency along the flow direction, y represents the coordinates of each point in the two-dimensional distribution image of the film cooling efficiency along the span direction, G(y, F(x), Q(x)) is the cooling efficiency value at the (x, y) coordinate point, and after passing through the convolution block, a matrix with the same size as the model output image is obtained as the input of the image encoding module; The feature extraction module performs high-dimensional extraction on the input air film hole parameter vector, and reconstructs the extracted high-dimensional features into a two-dimensional matrix, and then performs channel expansion through a convolution block to obtain a matrix with a size consistent with the image encoding output; The image encoding and decoding module downsamples the matrix output by the prior knowledge embedding module N times to obtain information at different levels of the image, and the features extracted at each layer are spliced ​​with the information at the same level after decoding; the matrix obtained after the last downsampling is fused with the high-dimensional features extracted by the feature extraction module and flattened into a one-dimensional vector, and then reconstructed through a multi-head attention mechanism to capture the key information in the image and vector features. After obtaining the key information, it is gradually upsampled and restored to the size of the image, which is provided to the output module to obtain the final two-dimensional distribution output of the film cooling efficiency; Step 3: Use the sample data obtained in step 1 to train the network model constructed in step 2 to obtain a trained prediction model; Step 4: For the film cooling scheme to be predicted, obtain the film hole parameter vector, input the trained prediction model obtained in step 3, and obtain the two-dimensional distribution prediction result of the film cooling efficiency.

2. According to claim 1, a method for predicting the two-dimensional distribution of film cooling efficiency based on prior knowledge embedding, characterized in that: The structural parameters include: the hole inclination angle α between the axis of the air film hole and the outlet plane; the expansion angle β of the air film hole expansion section; the forward inclination angle γ of the air film hole expansion section; the inner fold angle δ of the V-shaped hole outlet; the aspect ratio L / d of the air film hole; the width and diameter ratio L of the air film hole outlet w / d; Air film hole inlet and outlet area ratio A out / A in ; The dimensionless pore spacing P / d of the air film pores.

3. A method for predicting the two-dimensional distribution of film cooling efficiency based on prior knowledge embedding according to claim 1 or 2, characterized in that: The aerodynamic parameters include: blowing ratio M; density ratio DR; turbulence Tu and acceleration factor k.

4. According to claim 1, a method for predicting the two-dimensional distribution of film cooling efficiency based on prior knowledge embedding, characterized in that: The test is carried out by the following test system: The main flow is generated from the blower, passes through the butterfly valve and the rectifying cavity to produce a uniform inlet flow at the inlet of the test section; then different turbulence intensities are generated at the outlet of the film hole through turbulence grids of different densities to simulate the influence of different turbulence intensities on cooling; a suction split is set at the front end of the measuring plate to allow the boundary layer to redevelop there, so that the boundary layer thickness at the outlet of the film hole is close to the local boundary layer thickness of the film hole of the actual engine; at the same time, a trip wire is set behind the split to promote the transition of the boundary layer from laminar flow to turbulent flow; Velocity, temperature and pressure measuring points are set above the air film hole outlet to monitor the mainstream state during the test in real time; the coolant system uses a high-pressure gas tank and a compressor as the gas source, and is heated by a heater after passing through a valve and a pressure regulating device so that the temperature difference between the primary and secondary flows does not exceed 0.2K, and then enters the cold air cavity after the flow is controlled by an Alicat mass flowmeter and is supplied to the air film hole; temperature and pressure measuring points are set in the cold air cavity to monitor the coolant state in real time; PSP pressure-sensitive paint is sprayed on the surface of the measuring plate, and a corresponding excitation light source and a scientific-grade CCD camera with a filter are used to collect images during the test, and a two-dimensional distribution image of the air film cooling efficiency is obtained after post-processing the image.

5. According to claim 3, a method for predicting the two-dimensional distribution of film cooling efficiency based on prior knowledge embedding, characterized in that: The prior knowledge embedding module is composed of a coefficient mapping module and a learning function module; the coefficient mapping module includes an extraction layer, a hidden layer and an output layer. The coefficient mapping module first extracts the characteristics of the input air film hole parameter vector through an extraction layer composed of 16 neurons, and then maps it to coefficients A1, A2, A3, B1, B2, and B3 through an output layer composed of 16 neurons after passing through two hidden layers composed of 32 neurons respectively; in the learning function module, the cooling efficiency value of each point in the two-dimensional distribution cloud map of air film cooling efficiency is calculated using the learned 6 coefficients A1, A2, A3, B1, B2, and B3.

6. According to claim 1, a method for predicting the two-dimensional distribution of film cooling efficiency based on prior knowledge embedding, characterized in that: In the feature extraction module, the input air film hole parameter vector passes through the high-dimensional extraction layer composed of 512 neurons, 1024 neurons and 4096 neurons in turn, and the high-dimensional features are reconstructed into a 64×64 matrix. After the channel is expanded through a 3×3 convolution block, the expanded 64×64 matrix consistent with the image coding output is obtained.

7. The method for predicting the two-dimensional distribution of film cooling efficiency based on prior knowledge embedding according to claim 1, characterized in that: In the image encoding and decoding module, the 512×512 matrix corresponding to the two-dimensional distribution image of the film cooling efficiency output by the prior knowledge embedding module is first processed by two 3×3 convolution blocks to obtain a 512×512 first matrix. After the first matrix is ​​downsampled once and two 3×3 convolution blocks, a 256×256 second matrix is ​​obtained. After the second downsampling and two 3×3 convolution blocks, a 128×128 third matrix is ​​obtained. After the third downsampling and two 3×3 convolution blocks, a 64×64 fourth matrix is ​​obtained. The 64×64 fourth matrix is ​​fused with the high-dimensional features extracted by the feature extraction module and flattened into a one-dimensional vector. Then, the multi-head attention mechanism is used to obtain the 64×64 fourth matrix. Reconstruct into a fifth matrix of 64×64; after passing through a 3×3 convolution block, the reconstructed fifth matrix of 64×64 is concatenated with the fourth matrix of 64×64, and then passes through two 3×3 convolution blocks and the first deconvolution to obtain a sixth matrix of 128×128, and the sixth matrix is ​​concatenated with the third matrix of 128×128, and then passes through two 3×3 convolution blocks and the second deconvolution to obtain a seventh matrix of 256×256, and the seventh matrix is ​​concatenated with the second matrix of 256×256, and then passes through two 3×3 convolution blocks and the third deconvolution to obtain an eighth matrix of 512×512, and the eighth matrix is ​​concatenated with the first matrix of 512×512 and output to the output module.

Images