Air-cooled turbine cascade cooling efficiency prediction method suitable for sparse data
By combining a multi-scale generator and a discriminator, the problem of accuracy in predicting the cooling efficiency of air-cooled turbines under sparse data is solved, and high-precision reconstruction of the cooling efficiency distribution is achieved, which is suitable for the design of air-cooled turbines with complex configurations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-14
- Publication Date
- 2026-04-07
AI Technical Summary
Existing empirical correlation methods have low generalization ability in air-cooled turbine cooling design, cannot effectively utilize sparse datasets, resulting in inaccurate cooling efficiency predictions and difficulty in applying them to complex air-cooled turbine configurations.
A generative adversarial network (GAN) with multi-scale generators and non-parametric inputs and outputs is used to generate high-resolution cooling efficiency distribution maps through the generators. Combined with a multi-scale discriminator, high-precision prediction is performed, and sparse data is used to predict the cooling efficiency of air-cooled turbine blades.
It achieves high-precision prediction of air-cooled turbine cooling efficiency under sparse data conditions, and can quickly and accurately reconstruct the cooling efficiency distribution of air-cooled turbines, making it suitable for cooling designs with complex geometries.
Smart Images
Figure CN121811071A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of turbine cooling technology, and in particular to a method for predicting the cooling efficiency of air-cooled turbine blades suitable for sparse data. Background Technology
[0002] Complex and precise cooling technologies are employed for the high-temperature components of aero-engines to ensure their long service life. Film cooling (FSM) is a crucial aspect of cooling hot-end components in aero-engines. FSM is a cooling design structure influenced by multiple parameters, and rapid assessment of FSM efficiency is a key focus of cooling design research. Researchers often use empirical correlations to quantitatively describe the impact of parameters on cooling efficiency, but algebraic forms of empirical correlations (see the following references for details) are also important.
[0003] [1] Sellers JP. GASEOUS FILM COOLING WITH MULTIPLE INJECTIONSTATIONS. AIAA Journal. 1963;1(9):2154-6.
[0004] [2] Muska JF, Fish RW, Suo M. The Additive Nature of Film Cooling From Rows of Holes. Journal of Engineering for Power. 1976;98(4):457-63.) has the following obvious shortcomings:
[0005] First, empirical correlations are generally derived from experimental data or theoretical derivations. Their simple functional form and parameter space limit their applicability, resulting in low generalization. Second, empirical correlations are often subject to limitations in data collection and processing methods, leading to the loss of high-dimensional information and low fidelity. For air-cooled turbine cooling design, the experimental data for empirical correlations typically comes from simple geometries such as flat plate cooling. These types of empirical correlations are difficult to directly apply to complex air-cooled turbine configurations. Furthermore, the construction of empirical correlations is often based on a large amount of experimental or simulation data. Due to the complex flow in turbine channels and the complex cooling configuration of air-cooled turbines, the dataset of air-cooled turbine cooling efficiency distribution is often sparse.
[0006] In the development of cooling design, empirical correlations with low generalization and low fidelity cannot provide accurate references for the design environment faced by new cooling forms of air-cooled turbines. Therefore, it is necessary to develop high-precision, high-dimensional, and rapid evaluation and prediction methods for cooling efficiency applicable to sparse datasets. Summary of the Invention
[0007] The purpose of this invention is to provide a method for predicting the cooling efficiency of air-cooled turbine blades using sparse data. For the sparse data problem, a network incorporating a multi-scale generator and discriminator is specifically constructed to alleviate the data volume requirements in the design process. Simultaneously, considering the complex geometric configurations of air-cooled turbines, the network employs non-parametric inputs and outputs to ensure high-fidelity cooling efficiency prediction. This prediction method, based on a small amount of air-cooled turbine cooling efficiency data, inputs the three-dimensional geometry of the air-cooled turbine and outputs a high-resolution distribution of the air-cooled turbine cooling efficiency. This method enables rapid and accurate prediction of air-cooled turbine cooling efficiency, solving the technical problems of low generalization and easy loss of high-dimensional information in current cooling design applications.
[0008] This invention provides the following technical solution:
[0009] A method for predicting the cooling efficiency of air-cooled turbine blade cascades for sparse data includes the following steps:
[0010] S1. Obtain a raw, high-resolution image of the air-cooled turbine blade cascade, the raw image including geometry and cooling efficiency distribution images;
[0011] S2. The original image is input into a first generator and a second generator, respectively. Each of the first, second, and third generators includes a global generator and a local enhancer. The features in the cooling efficiency distribution maps generated by the first and second generators are summed element-wise and used as input to the third generator. The third generator then generates a comprehensive cooling efficiency distribution map. Both the first and second generators generate cooling efficiency distribution maps with lower resolutions compared to the original image. The comprehensive cooling efficiency distribution map generated by the third generator has the same resolution as the original image.
[0012] S3. The comprehensive cooling efficiency distribution map is obtained by CFD calculation and then downsampled at different scales and output to three discriminators at different scales, namely the first discriminator, the second discriminator and the third discriminator. The three discriminators are used to identify the authenticity of the comprehensive cooling efficiency distribution map at different resolutions.
[0013] S4. By initializing network weights, optimizing weights and loss functions, the predicted cooling efficiency of the air-cooled turbine blade cascade is obtained; the loss function includes adversarial loss, feature matching loss and VGG perception loss.
[0014] In the above embodiment, step S2 involves generating low-resolution images (which are cooling efficiency distribution maps, but at different resolutions) using the first and second generators G1 and G2. These two images with different resolutions need to be element-wise summed. For the entire generator, if generator G3 generates a 1028*1028 resolution image, then the resolutions of the images generated by G1 and G2 decrease exponentially to 256*256 (1028 / 4) and 512*512 (1028 / 2).
[0015] The three generators are laid out in a coarse-to-fine structure, with the main purpose of efficiently and stably generating cooling efficiency distribution maps at different resolution scales.
[0016] According to some possible implementations, in step S2, the generation process of the first, second, or third generator all includes the following steps;
[0017] S2.1 The original image first reaches the global generator, which is a U-Net with skip connections removed and uses residual blocks. It is responsible for understanding the global layout and semantic information of the image and generating a coarse-grained image with a lower resolution. The detailed hierarchical structure includes convolutional downsampling, residual blocks, transposed convolutional upsampling, and an output layer.
[0018] The convolutional downsampling encoder part has three downsampling layers, each implemented by a StridedConvolution. The convolutional kernels are 3x3 or 4x4. Each convolutional layer is followed by the ReLU activation function and instance normalization is used.
[0019] For a 2048x1024 output, six residual blocks are stacked, each containing two convolutional layers with the following structure: [Conv3x3 -> InstanceNorm -> ReLU] and [Conv3x3 -> InstanceNorm]. Finally, the input of the block is added to the output of the second convolutional layer, i.e., a skip connection. ReLU is used after the first convolution, and no activation function is added after the second one.
[0020] The transposed convolution upsampling and downsampling layers are symmetrical, with 3 upsampling layers. Each upsampling layer is implemented through a Transposed Convolution with a stride of 2. The convolution kernel is 3x3 or 4x4. The ReLU activation function is used after each transposed convolution layer.
[0021] The output layer is the final Conv3x3 layer, which uses the Tanh activation function to compress the output values to the range of [-1, 1] to match the normalized real image value range;
[0022] S2.2 The local enhancer receives the intermediate layer features from the global generator and the downsampled input semantic map, adding details and textures to the coarse-grained image to improve the image resolution to a higher level.
[0023] The detailed hierarchical structure includes input processing, convolutional layers and residual blocks, and an output layer. The input processing of the local enhancer concatenates the intermediate features of the global generator with the downsampled semantic input map along the channel dimension; the activation function is ReLU; the output layer uses Conv3x3 and Tanh activation functions to output high-resolution residual details; the residual details are added to the upsampled result of the global generator to obtain the final refined output.
[0024] In the above implementation, the global generator in step S2.2 ensures the correctness of the object layout and structure, while the local enhancer is responsible for making the details more realistic. If a higher resolution image (such as 4096×2048) needs to be generated, more local enhancers can be stacked. The detailed hierarchical structure is similar to the latter half of the global generator, also containing multiple convolutional layers and residual blocks, but the network capacity (number of channels) may be less because it only needs to process local details.
[0025] According to some possible implementations, the discriminator in step S3 is PatchGAN, which distinguishes between real and fake patches in each local region of the image, and finally takes the average value of all patch responses as the final discrimination result.
[0026] According to some possible implementations, each of the discriminators includes an input layer, a convolutional downsampling layer, and an output layer;
[0027] The input layer receives either an image or a conditional image combined with a target image.
[0028] The convolutional downsampling consists of 5 convolutional layers. The first 4 layers are strided convolutions with a stride of 2 to achieve downsampling, and the convolutional kernels are 4x4. Except for the last layer, each convolutional layer is followed by a LeakyReLU activation function with a negative slope of 0.2. Except for the first and last layers, the intermediate layers use spectral normalization, which stabilizes the discriminator training by constraining the spectral norm of the weight matrix of each convolutional layer. Starting from an initial 64 layers, the number of channels doubles with each downsampling.
[0029] The last convolutional layer of the output layer uses a convolution with a stride of 1 and no activation function; it outputs an N x N matrix, where each element represents the probability that a specific region in the input image is a true patch.
[0030] According to some possible implementations, step S4 includes the following steps:
[0031] S4.1 Before training the network, pre-training is performed, with the input and output being an image of the air-cooled turbine blade cascade and a cloud map of the cooling efficiency distribution, respectively. After the pre-trained network is initialized, the hyperparameter values are assigned to the generator and discriminator architecture of the prediction model network to ensure its training stability.
[0032] S4.2 Initialize network weights. The weights of the convolutional layers are randomly initialized from a normal distribution with a mean of 0 and a standard deviation of 0.02. The bias term is initialized to 0. The Adam optimizer is used, with the initial learning rate set to 0.0002, the momentum parameter β1 set to 0.5, and β2 set to 0.999.
[0033] S4.3 The loss function includes the adversarial loss inherent in GAN networks, the feature matching loss, and the VGG perceptual loss; the adversarial loss is the core of GAN training, which makes the generator and discriminator compete with each other; the feature matching loss is the perceptual difference loss between the features of the real image and the generated image calculated on the intermediate feature layer of the discriminator; the VGG perceptual loss is the perceptual difference loss between the real image and the generated image calculated on its feature layer by the pre-trained VGG network.
[0034] In the above implementation, the loss function is multi-scale to ensure prediction quality. The loss function consists of multiple parts that collectively guide the optimization of the generator and discriminator. These include the adversarial loss inherent in the GAN network (the core of GAN training, where the generator and discriminator compete against each other), feature matching loss (to address training instability, the high-precision prediction model calculates the perceptual difference loss between features of the real and generated images on the intermediate feature layer of the discriminator, which helps stabilize the training process and improve generation quality), and VGG perceptual loss (the pre-trained VGG network calculates the perceptual difference loss between the real and generated images on its feature layer). This ensures that the generated images are consistent with the real images in visual content and perception, rather than just pixel-level matching, contributing to the generation of more realistic textures.
[0035] Compared with the prior art, the present invention has the following beneficial effects:
[0036] The present invention provides a method for predicting the cooling efficiency of air-cooled turbine blades for sparse data. This method is a non-parametric turbine loss prediction method that can quickly build a non-parametric turbine loss prediction model.
[0037] The prediction method provides image-to-image output and output methods, and extracts the geometric features of the air-cooled turbine cooling configuration through generative adversarial networks to achieve rapid reconstruction of the air-cooled turbine cooling efficiency.
[0038] The prediction method also provides a multi-scale generator and discriminator architecture designed for sparse data. To ensure the prediction accuracy of the network on sparse datasets, the generator adopts a coarse-to-fine structure, primarily aimed at efficiently and stably generating high-resolution images of air-cooled turbine cooling efficiency. Simultaneously, the high-precision prediction method for air-cooled turbine cooling efficiency employs a multi-scale discriminator to distinguish between genuine and fake high-resolution generated images. Attached Figure Description
[0039] Figure 1 The diagram shows the prediction network architecture of the air-cooled turbine blade cooling efficiency prediction method for sparse data provided in this embodiment of the invention.
[0040] Figure 2 This is a schematic diagram of the VGG pre-trained network architecture provided in an embodiment of the present invention.
[0041] Figure 3 The prediction results provided in the training and test sets for embodiments of the present invention.
[0042] Figure 4 This is a comparison chart of prediction errors between prediction results and simulation calculation results on the test set, provided as an embodiment of the present invention. Detailed Implementation
[0043] The present invention will now be described in detail with reference to embodiments and accompanying drawings. However, it should be understood that the embodiments and drawings are for illustrative purposes only and do not constitute any limitation on the scope of protection of the present invention. All reasonable modifications and combinations included within the scope of the inventive spirit of the present invention fall within the scope of protection of the present invention.
[0044] The present invention will be further described below with reference to the accompanying drawings.
[0045] Example 1
[0046] The flowchart of the air-cooled turbine blade cooling efficiency prediction method for sparse data provided in this embodiment is as follows: Figure 1 and Figure 2 As shown in the figure, the main framework and processing steps of the high-precision prediction model for air-cooled turbine cooling efficiency are illustrated. Figure 2 This refers to the model's prediction results on the training and test sets. Before testing, samples are selected in the sample space using the hypercube sampling method and randomly divided into training and test sets. The network is trained on the training set and then makes predictions on the test set. The prediction results are compared with the CFD calculation results to obtain the prediction accuracy on the training set.
[0047] like Figure 1This method is based on Generative Adversarial Network (GAN). To address the sparse sample problem, it employs multi-scale generators and discriminators. The input to the prediction process is an image of the air-cooled turbine blade cascade. The cooling efficiency feature maps generated by the two generators at different resolution scales are element-wise summed and used as input to the high-resolution generator. Finally, the output is fed to the discriminator with the same architecture but different sampling rates to reconstruct a high-resolution and high-precision air-cooling efficiency distribution cloud map. This architecture can effectively alleviate the data requirements of neural networks.
[0048] To ensure high-fidelity representation of the cooling efficiency of the air-cooled turbine surface, both the network input and output use high-resolution (1024*1024 pixels) images. The high-resolution image of the air-cooled turbine geometry serves as the network input. For this cooling configuration, low-resolution generators G1 and G2 generate cooling efficiency distributions with resolutions lower than the original image. These generated distributions are then element-wise summed before being used as input to generator G3. Generator G3 then generates a cooling efficiency distribution with the same resolution as the original image. This approach ensures that the generator captures image details at different image scales. The high-resolution images generated by the multi-scale generators and the high-resolution images calculated by CFD (Computational Fluid Dynamics, in this patent, directly refer to conventional computational fluid dynamics simulation software) are downsampled at different scales and output to discriminators D1, D2, and D3 at different scales. Discriminators D1, D2, and D3 then determine whether the high-resolution images generated by the generators are genuine at different resolutions. (The evaluation criterion used here is SSIM, image similarity, which is whether it is consistent with the CFD calculation results.)
[0049] The operating mechanism of this method can be summarized in the following steps:
[0050] 1. Obtain original high-resolution images of the air-cooled turbine blade cascade, the original images including geometry and cooling efficiency distribution images;
[0051] 2. Generator network that executes a coarse-to-fine rendering strategy:
[0052] All three generators employ the same coarse-to-fine structure, primarily aimed at efficiently and stably generating high-resolution images. It typically comprises two sub-networks: a global generator and a local enhancer.
[0053] A. Global Generator: Its structure is a U-Net with skip connections removed (or more like an encoder-decoder structure) and uses residual blocks. It is responsible for understanding the global layout and semantic information of the image, generating a coarse-grained image at a lower resolution (e.g., 1024×512). The detailed hierarchical structure includes convolutional downsampling (encoder part), residual blocks (intermediate transform part), transposed convolutional upsampling (decoder part), and the output layer.
[0054] Convolutional downsampling (encoder part) typically has 3 downsampling layers, each implemented by a strided convolution (convolution with stride of 2). The convolution kernel is usually 3x3 or 4x4. Each convolutional layer is followed by a ReLU activation function and instance normalization is used.
[0055] For a 2048x1024 output, six residual blocks are stacked, each containing two convolutional layers with the following structures: [Conv3x3 -> InstanceNorm -> ReLU] and [Conv3x3 -> InstanceNorm]. Finally, the block's input is added to the output of the second convolutional layer (a skip connection). ReLU is applied after the first convolution, but no activation function is applied after the second.
[0056] The transposed convolutional upsampling (decoder part) is symmetrical to the downsampling layers and typically consists of three upsampling layers. Each upsampling layer is implemented by a Transposed Convolution (sometimes called deconvolution) with a stride of 2. The convolutional kernels are usually 3x3 or 4x4, and each transposed convolutional layer is followed by a ReLU activation function.
[0057] The output layer is a final Conv3x3 layer that uses the Tanh activation function to compress the output values to the range of [-1,1] to match the normalized range of the real image.
[0058] B. The local enhancer receives intermediate layer features (not the final output) from the global generator and a downsampled input semantic map. It focuses on adding details and textures to the coarse-grained image, upscaling the image resolution to higher levels (e.g., 2048×1024). In this way, the global generator ensures the correctness of object layout and structure, while the local enhancer is responsible for making the details more realistic. If generating even higher resolution images (e.g., 4096×2048) is required, more local enhancers can be stacked.
[0059] The detailed hierarchical structure includes input processing, convolutional layers and residual blocks, and an output layer. The input processing of the local enhancer concatenates the intermediate features of the global generator with the downsampled semantic input map along the channel dimension. Its structure is similar to the latter half of the global generator, also containing multiple convolutional layers and residual blocks, but the network capacity (number of channels) may be less because it only needs to process local details. The activation function is ReLU. The output layer also uses Conv3x3 and Tanh activation functions, outputting high-resolution residual details. This output is added to the upsampled result of the global generator to obtain the final refined output.
[0060] 3. Multi-scale discriminator networks:
[0061] To distinguish between genuine and fake high-resolution generated images, a multi-scale discriminator is employed in the high-precision prediction of air-cooled turbine cooling efficiency. It comprises three discriminators with identical structures but different processing scales (D1, D2, D3):
[0062] D1 judges at the original scale (e.g., 1024×512), focusing on the realism of details.
[0063] D2 is used for identification at a 1 / 2 downsampling scale, taking into account both details and local structures.
[0064] D3 performs discrimination at a 1 / 4 downsampling scale, focusing more on the global consistency of the image and whether the overall layout is reasonable.
[0065] Each discriminator is a PatchGAN, which does not directly classify the entire image as real or fake. Instead, it performs a real / fake judgment on each local region (patch) in the image and finally takes the average of all patch responses (output matrix) as the final judgment result. This effectively guides the generator to generate high-frequency information rich in detail.
[0066] The detailed hierarchical structure of a single discriminator includes an input layer, convolutional downsampling, and an output layer.
[0067] The input layer receives an image (which may be a real image or a generated image), or a concatenation of a conditional image (semantic graph) and a target image.
[0068] Convolutional downsampling typically involves five convolutional layers. The first four layers are strided convolutions (convolutions with a stride of 2) to achieve downsampling, and the kernels are usually 4x4. Except for the last layer, each convolutional layer uses the LeakyReLU activation function, with a negative slope typically of 0.2. Except for the first and last layers, the intermediate layers use spectral normalization. A significant improvement in prediction networks is the use of spectral normalization instead of batch normalization commonly used in other GANs. Spectral normalization stabilizes the discriminator's training by constraining the spectral norm (i.e., the largest singular value) of the weight matrix of each convolutional layer, thereby improving the generator's performance. Starting with an initial 64 layers, the number of channels doubles with each downsampling iteration (64 -> 128 -> 256 -> 512).
[0069] The last convolutional layer in the output layer uses a stride of 1 and has no activation function. Instead of a single true / false scalar, it outputs an N x N matrix (e.g., 32x32). Each element in this matrix represents the probability that a specific region (patch) in the input image is true. This structure allows the discriminator to focus on the authenticity of local structures in the image.
[0070] 4. Before network training, pre-training is performed, with the input and output being an image of the air-cooled turbine blade cascade and a cloud map of cooling efficiency distribution, respectively. After the pre-trained network is initialized, the hyperparameter values are assigned to the generator and discriminator architectures of the prediction model network to ensure its training stability.
[0071] Key parameters and initialization:
[0072] Weight initialization: The weights of the convolutional layer are usually randomly initialized from a normal distribution with a mean of 0 and a standard deviation of 0.02. The bias term is initialized to 0. The Adam optimizer is used, and the initial learning rate is usually set to 0.0002. The momentum parameter β1 is usually set to 0.5 and β2 is set to 0.999.
[0073] The generator effectively constructs image content and details through a coarse-to-fine design and instance normalization, while the discriminator accurately distinguishes between real and fake images from different perspectives and stabilizes the training process through multi-scale and spectral normalized PatchGAN. The entire network works collaboratively through carefully selected initialization, optimizers, and loss functions (adversarial loss, feature matching loss, and VGG perceptual loss).
[0074] Multi-scale loss functions ensure prediction quality:
[0075] The loss function of the high-precision prediction model for air-cooled turbine cooling efficiency consists of multiple parts that collectively guide the optimization of the generator and discriminator:
[0076] Adversarial loss is the core of GAN training, allowing the generator and discriminator to compete against each other.
[0077] Feature matching loss: To address the issue of training instability, the high-precision prediction model for air-cooled turbine cooling efficiency calculates the L1 loss between features of the real and generated images at the intermediate feature layer of the discriminator. This helps stabilize the training process and improve generation quality.
[0078] Content loss: A pre-trained VGG network is used to compute the perceptual difference (L1 loss) between the real and generated images at its feature layers. This ensures that the generated images are consistent with the real images in terms of visual content and perception, rather than just pixel-level matching, which helps to generate more realistic textures.
[0079] The overall loss function is a weighted sum of these losses, which together guide the generator to produce high-resolution images that are both realistic and rich in detail.
[0080] Figure 3 and Figure 4 The test results of the prediction method provided in this embodiment on the training set and the test set are shown respectively. The number of samples in the training set and the test set is less than 200, which is much smaller than the dataset size required for conventional deep learning training; and as can be seen from the figure, the average accuracy of the predicted images on the training set is 99.4%, and the average accuracy of the predicted images on the test set is 99.2%.
[0081] The above embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A method for predicting the cooling efficiency of air-cooled turbine blades suitable for sparse data, characterized in that: Includes the following steps: S1. Obtain a raw, high-resolution image of the air-cooled turbine blade cascade, the raw image including geometry and cooling efficiency distribution images; S2. The original image is input into a first generator and a second generator, respectively. Each of the first, second, and third generators includes a global generator and a local enhancer. The features in the cooling efficiency distribution maps generated by the first and second generators are summed element-wise and used as input to the third generator. The third generator then generates a comprehensive cooling efficiency distribution map. Both the first and second generators generate cooling efficiency distribution maps with lower resolutions compared to the original image. The comprehensive cooling efficiency distribution map generated by the third generator has the same resolution as the original image. S3. The comprehensive cooling efficiency distribution map is obtained by CFD calculation and then downsampled at different scales and output to three discriminators at different scales, namely the first discriminator, the second discriminator and the third discriminator. The three discriminators are used to identify the authenticity of the comprehensive cooling efficiency distribution map at different resolutions. S4. By initializing network weights, optimizing weights and loss functions, the predicted cooling efficiency of the air-cooled turbine blade cascade is obtained; the loss function includes adversarial loss, feature matching loss and VGG perception loss.
2. The method for predicting the cooling efficiency of air-cooled turbine blades for sparse data according to claim 1, characterized in that: In step S2, the generation process of the first, second, or third generator all includes the following steps; S2.1 The original image first reaches the global generator, which is a U-Net with skip connections removed and uses residual blocks. It is responsible for understanding the global layout and semantic information of the image and generating a coarse-grained image with a lower resolution. The detailed hierarchical structure includes convolutional downsampling, residual blocks, transposed convolutional upsampling, and an output layer. The convolutional downsampling encoder part has 3 downsampling layers, each implemented by a strided convolution. The convolutional kernel is 3x3 or 4x4. Each convolutional layer is followed by the ReLU activation function and instance normalization is used. For a 2048x1024 output, six residual blocks are stacked, each containing two convolutional layers with the following structure: [Conv3x3 -> InstanceNorm -> ReLU] and [Conv3x3 -> InstanceNorm]. Finally, the input of the block is added to the output of the second convolutional layer, i.e., a skip connection. ReLU is used after the first convolution, and no activation function is added after the second one. The transposed convolution upsampling and downsampling layers are symmetrical, with 3 upsampling layers. Each upsampling layer is implemented through a Transposed Convolution with a stride of 2. The convolution kernel is 3x3 or 4x4. The ReLU activation function is used after each transposed convolution layer. The output layer is the final Conv3x3 layer, which uses the Tanh activation function to compress the output values to the range of [-1, 1] to match the normalized real image value range; S2.2 The local enhancer receives the intermediate layer features from the global generator and the downsampled input semantic map, adding details and textures to the coarse-grained image to improve the image resolution to a higher level. The detailed hierarchical structure includes input processing, convolutional layers and residual blocks, and an output layer. The input processing of the local enhancer concatenates the intermediate features of the global generator with the downsampled semantic input map along the channel dimension; the activation function is ReLU; the output layer uses Conv3x3 and Tanh activation functions to output high-resolution residual details; the residual details are added to the upsampled result of the global generator to obtain the final refined output.
3. The method for predicting the cooling efficiency of air-cooled turbine blades for sparse data according to claim 2, characterized in that: The discriminator mentioned in step S3 is PatchGAN, which distinguishes between real and fake patches in each local region of the image, and finally takes the average value of all patch responses as the final discrimination result.
4. The method for predicting the cooling efficiency of air-cooled turbine blades for sparse data according to claim 3, characterized in that: Each of the discriminators includes an input layer, a convolutional downsampling layer, and an output layer; The input layer receives either an image or a conditional image combined with a target image. The convolutional downsampling consists of 5 convolutional layers. The first 4 layers are strided convolutions with a stride of 2 to achieve downsampling, and the convolutional kernels are 4x4. Except for the last layer, each convolutional layer is followed by a LeakyReLU activation function with a negative slope of 0.
2. Except for the first and last layers, the intermediate layers use spectral normalization, which stabilizes the discriminator training by constraining the spectral norm of the weight matrix of each convolutional layer. Starting from an initial 64 layers, the number of channels doubles with each downsampling. The last convolutional layer of the output layer uses a convolution with a stride of 1 and no activation function; it outputs an N x N matrix, where each element represents the probability that a specific region in the input image is a true patch.
5. The method for predicting the cooling efficiency of air-cooled turbine blades for sparse data according to claim 4, characterized in that: Step S4 includes the following steps: S4.1 Before training the network, pre-training is performed, with the input and output being an image of the air-cooled turbine blade cascade and a cloud map of the cooling efficiency distribution, respectively. After the pre-trained network is initialized, the hyperparameter values are assigned to the generator and discriminator architecture of the prediction model network to ensure its training stability. S4.2 Initialize network weights. The weights of the convolutional layers are randomly initialized from a normal distribution with a mean of 0 and a standard deviation of 0.
02. The bias term is initialized to 0. The Adam optimizer is used, with the initial learning rate set to 0.0002, the momentum parameter β1 set to 0.5, and β2 set to 0.
999. S4.3 The loss function includes the adversarial loss inherent in GAN networks, the feature matching loss, and the VGG perceptual loss; the adversarial loss is the core of GAN training, which makes the generator and discriminator compete with each other; the feature matching loss is the perceptual difference loss between the features of the real image and the generated image calculated on the intermediate feature layer of the discriminator; the VGG perceptual loss is the perceptual difference loss between the real image and the generated image calculated on its feature layer by the pre-trained VGG network.