Turbine blade air film cooling temperature field reconstruction method based on iterative Fourier neural operator network

Abstract

The invention discloses a turbine blade air film cooling temperature field reconstruction method based on an iterative Fourier neural operator network, and the method comprises the steps: 1, carrying out the frequency domain transformation of an input temperature field and a boundary condition through a Fourier neural operator, dynamically fusing the global features of a frequency domain and the local features of a space domain through an attention mechanism, and carrying out the frequency domain transformation; the modeling capability of a complex flow mode is improved; 2, a multi-scale iterative optimization framework is constructed, a temperature field prediction result is iteratively corrected step by step from low resolution to high resolution, the calculation efficiency is improved, and error accumulation is avoided; and 3, gradient consistency loss is introduced to constrain the consistency between the temperature field space gradient and the real distribution, the reconstruction result is prevented from being over-smooth, and the requirement for the target domain data volume is reduced by using a transfer learning strategy. According to the method, the Fourier neural operator, multi-scale iterative optimization and transfer learning are integrated, the turbine blade air film cooling temperature field can be efficiently reconstructed, and the generalization ability of the model under sparse data can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal fluid data reconstruction and deep learning, and specifically to a method for reconstructing the temperature field of turbine blade film cooling by integrating frequency-domain feature learning and multi-scale iterative optimization, which is applicable to the cooling efficiency analysis and optimal design of high-temperature components such as aeroengines and gas turbines. Background Art

[0002] Turbine blades are the core hot-end components of aeroengines and gas turbines, and are subjected to the impact of high-temperature gas for a long time. To ensure the structural integrity and service life of the blades under extreme working conditions, film cooling technology is widely used for blade thermal protection. This technology injects low-temperature cooling gas into the high-temperature gas flow by opening film holes on the blade surface, forming a cooling gas film covering the blade surface, and significantly reducing the surface temperature. Therefore, high-resolution temperature field data of the film cooling efficiency of turbine blades is crucial for optimizing the performance of high-temperature components. The accurate reconstruction of the temperature field is the key basis for evaluating the cooling efficiency, optimizing the layout of film holes and flow distribution, and directly affects the thermodynamic performance and service life of the blades.

[0003] The acquisition of the traditional temperature field of turbine blade film cooling mainly relies on the following two types of methods: one is the experimental measurement method, which can directly reflect the temperature distribution under actual working conditions, but it is difficult to install sensors in high-temperature, high-pressure, and high-speed rotating environments, the measurement accuracy is limited, and the experimental cost is high, the cycle is long, and it is difficult to cover the entire working condition parameter space; the other is the computational fluid dynamics (CFD) simulation method, which can simulate the temperature field details under complex geometries and flow conditions, but high-fidelity simulation requires ultra-fine grid division, the single simulation takes a long time, and the transient flow coupling solution of the porous film cooling structure consumes huge computing resources. In recent years, the temperature field reconstruction method based on machine learning has gradually become a research hotspot, and its core is to replace or accelerate the traditional numerical simulation through data.

[0004] Although significant progress has been made in the temperature field reconstruction method of turbine blade film cooling, many challenges still remain. One of the main problems is that the film cooling temperature field contains multi-scale flow characteristics and the data is scarce. Existing studies are difficult to capture global and local characteristics simultaneously, and are prone to losing high-frequency temperature fluctuations in high-gradient regions such as the jet impact area and the edge of film holes, resulting in the reconstructed temperature field being too smooth. If the reconstructed temperature fields are incomplete and inaccurate, they cannot accurately reflect the cooling efficiency of the turbine blade film, nor can they guide engineering optimization. Summary of the Invention

[0005] The present invention aims to solve the deficiencies of the above-mentioned existing technologies, and proposes a method for reconstructing the temperature field of turbine blade film cooling based on an iterative Fourier neural operator network, with the expectation of effectively modeling the global flow characteristics and local characteristics of the temperature field, enhancing local high-frequency details, and improving the adaptability in sparse data scenarios, thereby contributing to the efficient and accurate reconstruction of the film cooling temperature field and improving the accuracy of temperature field reconstruction.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for reconstructing the temperature field of turbine blade film cooling based on an iterative Fourier neural operator network is characterized in that it is carried out according to the following steps:

[0008] Step 1: Define the temperature field set and the boundary condition set , let The temperature field of any turbine blade film hole in is denoted as , and let The boundary condition of any turbine blade film hole in represents the height of the temperature field, represents the width of the temperature field;

[0009] Step 2: Use equations (1) and (2) to perform downsampling on the temperature field and the boundary condition of any turbine blade film hole synchronously for times to generate a multi-scale temperature field set and its matching boundary condition set ; thus, the fusion tensor of the th scale is obtained using equation (3); where represents the temperature field of the th scale, represents the boundary condition of the th scale, represents the height of the temperature field of the th scale, represents the width of the temperature field of the th scale, and represents the height of the temperature field of the th scale, represents the width of the temperature field of the th scale; when , let , let ; represents rounding down; represents the The height of the temperature field at the scale, and , representing the width of the temperature field at the scale, and :

[0010] (1)

[0011] (2)

[0012] (3)

[0013] In equations (1) and (2), respectively represent the temperature field and boundary conditions at the scale; is the pixel coordinate after downsampling; represents the normalized bilinear weight kernel at different row index i and column index j; when n = 1, let , let ;

[0014] In equation (3), represents the channel concatenation function;

[0015] Step 3, construct an iterative Fourier neural operator network, and perform multi-scale iterative optimization processing to obtain the finally reconstructed temperature field :

[0016] Step 4, based on the reconstructed temperature field and the true temperature field at its corresponding resolution construct the total loss function , where represents the true temperature field at the scale:

[0017] Step 5, use the gradient descent method to train the temperature field reconstruction model based on the iterative Fourier neural operator network, calculate the total loss function to update the network parameters until the total loss function converges, so as to obtain the trained temperature field reconstruction model and use it for effective reconstruction of the turbine blade film cooling temperature field.

[0018] The feature of the turbine blade film cooling temperature field reconstruction method based on the iterative Fourier neural operator network according to the present invention also lies in that the step 3 includes:

[0019] Step 3.1, initialize ;

[0020] Step 3.2, use equation (4) to obtain the Initial hidden layer features at the scale , denotes the dimension of the hidden layer features:

[0021] (4)

[0022] In equation (4), denotes a linear layer; denotes the fusion tensor at the scale;

[0023] Step 3.3. Perform iterations of the layer Fourier neural operator to obtain the hidden layer features of the scale at the layer: :

[0024] Step 3.4. Use equation (10) to obtain the temperature field reconstructed at the scale:

[0025] (10)

[0026] In equation (10), denotes another linear layer;

[0027] Step 3.5. Use equation (11) to perform one upsampling on to generate the temperature field at the scale:

[0028] (11)

[0029] In equation (11), denotes the pixel coordinates after upsampling; denotes the convolutional kernel to be learned during upsampling; and are the row index and column index of the convolutional kernel ;

[0030] Step 3.6. Use equation (12) to generate the fusion tensor at the scale:

[0031] (12)

[0032] In equation (12), denotes the boundary condition at the scale;

[0033] Step 3.7. Assign to After that, return to step 3.2 and execute sequentially until it reaches the end, thus obtaining the finally reconstructed temperature field ; when it is the case, let .

[0034] Furthermore, the said step 3.3 includes:

[0035] Step 3.3.1, define the current layer number as , and initialize ;

[0036] Step 3.3.2, obtain the frequency domain feature of the th scale and the th layer by using Equation (5): :

[0037] (5)

[0038] In Equation (5), represents the fast Fourier transform, represents the inverse Fourier transform; represents the frequency of the th scale and the th layer; represents 's importance score, and there is:

[0039] (6)

[0040] In Equation (6), represents average pooling; represents a fully connected layer; represents the sigmoid activation function;

[0041] Step 3.3.3, obtain the spatial domain feature of the th scale and the th layer by using Equation (7): :

[0042] (7)

[0043] In Equation (7), represents the weight matrix to be learned;

[0044] Step 3.3.4, calculate the attention weight score of the th scale and the th layer by using Equation (8): :

[0045] (8)

[0046] In formula (8), represents another fully connected layer; represents the channel concatenation operation;

[0047] Step 3.3.5. Obtain the th layer hidden layer feature of the th

[0048] using formula (9):

[0049] In formula (9), represents the dot product operation; represents the Gaussian error linear unit activation function;

[0050] Step 3.3.6. After assigning to , return to Step 3.3.2 and execute sequentially until is reached, so as to obtain the th layer hidden layer feature of the th

[0051] Furthermore, the said Step 4 includes:

[0052] Step 4.1. Construct the main loss using formula (13):

[0053] (13)

[0054] In formula (13), represents the weight coefficient of the th scale, and represents the L2 norm;

[0055] Step 4.2. Construct the gradient consistency loss using formula (14):

[0056] (14)

[0057] In formula (14), represents the spatial gradient;

[0058] Step 4.3. Construct the total loss function using formula (15):

[0059] (15)

[0060] In formula (15), is a hyperparameter.

[0061] An electronic device according to the present invention includes a memory and a processor, characterized in that the memory is used to store a program that supports the processor to execute the method for reconstructing the temperature field of the turbine blade film cooling, and the processor is configured to execute the program stored in the memory.

[0062] A computer-readable storage medium according to the present invention, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is run by a processor, it executes the steps of the method for reconstructing the temperature field of the turbine blade film cooling.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. The present invention introduces a Fourier neural operator network, which transforms the fusion input of the temperature field and boundary conditions into the frequency domain through Fourier transform, and performs a linear transformation in the frequency domain to capture the global flow characteristics of the temperature field; at the same time, through the attention mechanism, the global features in the frequency domain and the local features in the spatial domain are dynamically fused, enhancing the local high-frequency details, thereby improving the model's ability to model complex flow modes and enhancing the detail reconstruction ability of high-gradient regions such as the jet impingement zone and the edge of the film hole.

[0065] 2. The present invention constructs a multi-scale iterative optimization framework through downsampling-upsampling operations, iteratively corrects the temperature field prediction results from low resolution to high resolution level by level, and realizes the reconstruction from coarse to fine; at the same time, by calculating the loss layer by layer, the global consistency from low resolution to high resolution is ensured, thereby improving the computational efficiency of the model and avoiding the error accumulation of single reconstruction.

[0066] 3. The present invention introduces gradient consistency loss into the loss function. By constraining the consistency between the spatial gradient of the temperature field and the true distribution, the over-smoothing phenomenon of the reconstruction result is effectively suppressed, and high-frequency features such as the jet impingement zone are retained; at the same time, using the transfer learning strategy, the general flow features are pre-trained through the CFD data of flat plate film cooling, and then fine-tuned with a small amount of turbine blade data, significantly reducing the demand for the amount of data in the target domain. Description of the Drawings

[0067] Figure 1 It is the overall flowchart of the method of the present invention. Detailed Embodiments

[0068] In this embodiment, a method for reconstructing the temperature field of film cooling of turbine blades based on an iterative Fourier neural operator network is to use the Fourier neural operator network to perform frequency domain transformation to extract the global flow characteristics of the temperature field, and dynamically fuse the frequency domain global characteristics with the spatial domain local characteristics in combination with the attention mechanism to enhance the local high-frequency details. Then, a multi-scale iterative framework is used to achieve coarse-to-fine reconstruction, and a gradient consistency loss is introduced to constrain high-frequency details. At the same time, the dependence on the amount of target domain data is further reduced through a transfer learning strategy. Specifically, Figure 1 As shown, the method is performed in the following steps:

[0069] Step 1: Define the temperature field set and boundary condition set ,make The temperature field of any turbine blade film hole is recorded as ,make The boundary condition of any turbine blade film hole is ;in, represents the temperature field height, Indicates the width of the temperature field.

[0070] Step 2: Use equations (1) and (2) to calculate the temperature field of the film hole of any turbine blade: and boundary conditions Synchronous Downsampling, in this embodiment, set , generate a multi-scale temperature field collection and its matching boundary condition set ; Then, using formula (3), we can get Scaled fused tensor , which is used as the initial input of the multi-scale iterative optimization framework; Indicates The temperature field of the scale, Indicates The boundary conditions of the scale, Indicates The temperature field height of the scale, Indicates The temperature field width of the scale is , Indicates The temperature field height of the scale, Indicates The temperature field width of the scale; when season ,make ; Indicates rounding down; Indicates The temperature field height of the scale, and , representing the temperature field width at the scale, and :

[0071] (1)

[0072] (2)

[0073] (3)

[0074] In equations (1) and (2), respectively represent the temperature field and boundary conditions at the scale; is the pixel coordinate after downsampling; represents the normalized bilinear weight kernel at different row index i and column index j. In this embodiment, equations (1) and (2) adopt the same kernel parameters; when n = 1, let , let ;

[0075] In equation (3), represents the channel concatenation function.

[0076] Step 3, construct an iterative Fourier neural operator network, and perform multi-scale iterative optimization processing on , starting from the scale with the lowest resolution, iteratively correct the temperature field prediction results from low resolution to high resolution level by level, realizing reconstruction from coarse to fine, and obtaining the finally reconstructed temperature field :

[0077] Step 3.1, initialize ;

[0078] Step 3.2, obtain the initial hidden layer feature at the scale using equation (4), represents the dimension of the hidden layer feature. In this embodiment, set :

[0079] (4)

[0080] In equation (4), represents a linear layer that elevates the fusion tensor to a high-dimensional channel space; represents the fusion tensor at the scale.

[0081] Step 3.3, perform on Iteration of the layer Fourier neural operator. In this embodiment, it is set that , to obtain the -scale -layer hidden layer features :

[0082] Step 3.3.1: Define the current layer number as , and initialize ;

[0083] Step 3.3.2: Use Equation (5) to obtain the -scale -layer frequency domain features :

[0084] (5)

[0085] In Equation (5), represents the fast Fourier transform, which converts the input to the frequency domain; represents the inverse Fourier transform, which converts the result back to the spatial domain; represents the -scale -layer frequency; represents 's importance score, enabling the network to autonomously enhance the frequency components strongly related to the turbine blade cooling efficiency (such as the main vortex frequency, boundary layer fluctuation characteristics, etc.), suppressing irrelevant noise, and the importance score can be automatically adjusted at different scales, and there is:

[0086] (6)

[0087] In Equation (6), represents average pooling; represents a fully connected layer. In this embodiment, is a three-layer fully connected layer; represents the sigmoid activation function.

[0088] Step 3.3.3: Use Equation (7) to obtain the -scale -layer spatial domain features :

[0089] (7)

[0090] In Equation (7), represents the weight matrix to be learned, and performs a local linear transformation on to extract local spatial domain features;

[0091] Step 3.3.4: Calculate the -scale Layer attention weight score , dynamic weighted frequency-domain global features and spatial-domain local features:

[0092] (8)

[0093] In formula (8), represents another fully connected layer. In this embodiment, is a three-layer fully connected layer; represents a channel concatenation operation.

[0094] Step 3.3.5. Use formula (9) to obtain the -th scale of the -th layer hidden layer features :

[0095] (9)

[0096] In formula (9), represents a dot product operation; represents the Gaussian error linear unit activation function;

[0097] Step 3.3.6. After assigning to , return to Step 3.3.2 and execute sequentially until is reached, so as to obtain the -th scale of the -th layer hidden layer features .

[0098] Step 3.4. Use formula (10) to obtain the temperature field reconstructed at the scale:

[0099] (10)

[0100] In formula (10), represents another linear layer, projects back to the target dimension, and ensures the network stability through residual connection to alleviate the vanishing gradient;

[0101] Step 3.5. Use formula (11) to perform an upsampling on to generate the temperature field at the scale:

[0102] (11)

[0103] In formula (11), represents the pixel coordinates after upsampling; Denote the convolutional kernel to be learned during the upsampling process; and are the row index and column index of the convolutional kernel .

[0104] Step 3.6, generate the fusion tensor of the th scale using Equation (12): :

[0105] (12)

[0106] In Equation (12), denotes the boundary condition of the th scale.

[0107] Step 3.7, after assigning to , return to Step 3.2 and execute sequentially until ; during the iteration process, the temperature field input is updated with the iteration, and the boundary condition remains unchanged, so as to obtain the finally reconstructed temperature field ; when , let .

[0108] Step 4, construct the total loss function based on the reconstructed temperature field and the true temperature field at its corresponding resolution , where denotes the true temperature field of the th scale:

[0109] Step 4.1, construct the main loss using Equation (13):

[0110] (13)

[0111] In Equation (13), denotes the weight coefficient of the th scale, and , in this embodiment, is set in an exponentially decaying form to emphasize the reconstruction accuracy of the high-resolution scale; denotes the L2 norm.

[0112] Step 4.2, construct the gradient consistency loss using Equation (14) to strengthen the consistency of the local changes in the temperature field and avoid the reconstructed result from being over-smoothed:

[0113] (14)

[0114] In Equation (14), represents the spatial gradient.

[0115] Step 4.3: Construct the total loss function using Equation (15) :

[0116] (15)

[0117] In Equation (15), is a hyperparameter. In this embodiment, is set.

[0118] Step 5: Use the gradient descent method to train the temperature field reconstruction model based on the iterative Fourier neural operator network, and calculate the total loss function to update the network parameters until the total loss function converges, so as to obtain the trained temperature field reconstruction model, which is used for the effective reconstruction of the turbine blade film cooling temperature field. Specifically, first construct a pre-training dataset based on the CFD simulation data of the flat plate film cooling structure, and use the gradient descent method to preliminarily train the temperature field reconstruction model based on the iterative Fourier neural operator network, and calculate the total loss function to update the model parameters until the total loss function converges, so as to obtain the pre-trained temperature field reconstruction model; then, fine-tune the temperature field reconstruction model on the turbine blade film cooling dataset. During fine-tuning, first transfer the pre-trained model parameters to the target model, and then calculate the total loss function to update the model parameters until the total loss function converges, so as to obtain the fine-tuned temperature field reconstruction model, which can not only efficiently reconstruct the turbine blade film cooling temperature field, but also improve the generalization ability of the model in the case of sparse turbine blade film cooling data.

[0119] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0120] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium. When the computer program is run by a processor, it executes the steps of the above method.

Claims

1. A method for reconstructing the temperature field of turbine blade film cooling based on an iterative Fourier neural operator network, characterized in that: The steps are as follows: Step 1: Define the temperature field set and boundary condition set ,make The temperature field of any turbine blade film hole is recorded as ,make The boundary condition of any turbine blade film hole is ;in, represents the height of the temperature field, Indicates the width of the temperature field; Step 2: Use equations (1) and (2) to calculate the temperature field of the film hole of any turbine blade: and boundary conditions Synchronous Downsample to generate a multi-scale temperature field collection and its matching boundary condition set ; Then, using formula (3), we can get Scaled fused tensor ;in, Indicates The temperature field of the scale, Indicates The boundary conditions of the scale, Indicates The temperature field height of the scale, Indicates The temperature field width of the scale is , Indicates The temperature field height of the scale, Indicates The temperature field width of the scale; when season ,make ; Indicates rounding down; Indicates The temperature field height of the scale, and , indicating the The temperature field width of the scale is : (1) (2) (3) In formula (1) and formula (2), Respectively represent Temperature fields and boundary conditions at different scales; is the pixel coordinate after downsampling; represents the normalized bilinear weight kernel under different row index i and column index j; when n=1, let ,make ; In formula (3), represents the channel splicing function; Step 3: Construct an iterative Fourier neural operator network and Perform multi-scale iterative optimization to obtain the final reconstructed temperature field : Step 4: Based on the reconstructed temperature field And the real temperature field at the corresponding resolution Constructing the total loss function ,in, Indicates The real temperature field of the scale: Step 5: Use the gradient descent method to train the temperature field reconstruction model based on the iterative Fourier neural operator network and calculate the total loss function To update the network parameters until the total loss function The training process is continued until convergence, thereby obtaining the trained temperature field reconstruction model, which is used for the effective reconstruction of the temperature field of the turbine blade film cooling.

2. The method for reconstructing the temperature field of turbine blade film cooling based on iterative Fourier neural operator network according to claim 1, characterized in that: The step 3 comprises: Step 3.1: Initialization ; Step 3.2: Use formula (4) to get Initial hidden layer features of scale , Represents the dimension of hidden layer features: (4) In formula (4), represents a linear layer; Indicates Scaled fusion tensor; Step 3.3: conduct The Fourier neural operator is iterated to obtain The scale Hidden layer features : Step 3.4: Use formula (10) to get Temperature field after scale reconstruction : (10) In formula (10), represents another linear layer; Step 3.5: Use formula (11) to Perform an upsampling to generate Temperature field of scale : (11) In formula (11), Represents the pixel coordinates after upsampling; Represents the convolution kernel to be learned during upsampling; and is the convolution kernel The row and column indices of ; Step 3.6: Use formula (12) to generate Scaled fused tensor : (12) In formula (12), Indicates Boundary conditions of scale; Step 3.7: Assign to Then return to step 3.2 and execute sequentially until So far, the final reconstructed temperature field is obtained ;when season .

3. The method for reconstructing the temperature field of turbine blade film cooling based on iterative Fourier neural operator network according to claim 2, characterized in that: The step 3.3 comprises: Step 3.3.1, define the current number of layers as , and initialize ; Step 3.3.2: Use formula (5) to get The scale Layer frequency domain features : (5) In formula (5), represents the fast Fourier transform, represents inverse Fourier transform; Indicates The scale Layer frequency; express The importance score of , and there are: (6) In formula (6), represents average pooling; represents a fully connected layer; Represents the sigmoid activation function; Step 3.3.3: Use formula (7) to get The scale Layer Spatial Characteristics : (7) In formula (7), Represents the weight matrix to be learned; Step 3.3.4: Use formula (8) to calculate the The scale Layer attention weight scores : (8) In formula (8), represents another fully connected layer; Indicates channel splicing operation; Step 3.3.5: Use formula (9) to get The scale Hidden layer features : (9) In formula (9), Represents the dot product operation; represents the Gaussian error linear unit activation function; Step 3.3.6: Assign to Then return to step 3.3.2 and execute sequentially until So far, we get The scale Hidden layer features .

4. The method for reconstructing the temperature field of turbine blade film cooling based on iterative Fourier neural operator network according to claim 3, characterized in that: The step 4 comprises: Step 4.1: Use formula (13) to construct the main loss : (13) In formula (13), Indicates The weight coefficient of the scale, and ; represents the L2 norm; Step 4.2: Use formula (14) to construct gradient consistency loss : (14) In formula (14), represents the spatial gradient; Step 4.3: Use formula (15) to construct the total loss function : (15) In formula (15), is a hyperparameter.

5. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the turbine blade film cooling temperature field reconstruction method as described in any one of claims 1-4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for reconstructing the temperature field of film cooling of turbine blades according to any one of claims 1 to 4 are executed.

Images