Turbine blade surface field rapid prediction method based on multi-channel self-attention network

Abstract

The invention discloses a turbine blade surface field rapid prediction method based on a multi-channel self-attention network, and relates to the technical field of data prediction. The technical problems that in the prior art, an existing turbine blade surface field prediction technology is long in consumed time, high in calculation cost and insufficient in precision, and obvious temperature and pressure distribution nonuniformity exists are solved. According to the technical scheme, the turbine blade surface field rapid prediction model building method based on the multi-channel self-attention network comprises the steps that turbine blade surface field data are collected; collecting a preset multi-channel self-attention network architecture; performing pretreatment; performing feature extraction and modeling on the turbine blade surface field according to the processed multi-channel self-attention network architecture; and according to the feature extraction and modeling results, establishing a'mapping relation from boundary conditions to a temperature field and a pressure field on the surface of the turbine blade '. The method is suitable for predicting the temperature field and the pressure field on the surface of the turbine blade.

Description

Technical Field

[0001] The invention relates to the technical field of data prediction, and in particular to rapid prediction of turbine blade surface fields. Background Art

[0002] As an important part of modern energy conversion equipment, gas turbines play a key role in aviation power, ship propulsion, ground power generation and other industrial fields. Its core components include compressors, combustion chambers and turbines. The turbine part converts the internal energy of high-temperature and high-pressure gas output from the combustion chamber into mechanical work, which directly determines the overall performance of the gas turbine. Therefore, in order to improve the efficiency of gas turbines and reduce pollution emissions, the design and optimization of turbine blades is one of the important research directions in the field of gas turbines.

[0003] In the prior art, the design and optimization of the surface field of turbine blades mainly rely on numerical simulation and physical experiments. Numerical simulation (such as CFD, computational fluid dynamics) technology can simulate the aerodynamic profile of turbine blades to evaluate their flow field characteristics, such as temperature field and pressure field. This method can provide relatively accurate prediction results in the early stages of design, but its computing resource requirements are huge, especially when simulating high Reynolds numbers or three-dimensional complex flow fields. The time and hardware costs of numerical simulation are often bottlenecks in the design process. On the other hand, physical experiments are usually expensive and time-consuming due to the complex equipment installation and processing involved, and can usually only be carried out in the final evaluation stage to verify the actual performance of the design.

[0004] In recent years, the application of machine learning in the field of fluid dynamics has gradually attracted attention. Scholars have tried to replace some numerical simulations by building proxy models to accelerate the design and optimization process. For example, the rapid prediction of flow fields through machine learning methods can be divided into the following main routes:

[0005] 1. Machine learning method based on turbulence model correction: predict anisotropic turbulence characteristics through machine learning or linear algebra methods, and correct the existing turbulence model to accelerate numerical calculations. This type of method can improve the convergence of numerical simulations, but it still needs to call a numerical solver in complex flow field predictions, which has high computational costs.

[0006] 2. Flow field prediction for initializing numerical simulation: Use machine learning to construct a rough estimate of the flow field as the initial condition for the numerical simulation to reduce the number of iterations and accelerate convergence. Although this method speeds up the calculation process, the prediction results are limited in accuracy and cannot be directly used for turbine blade design.

[0007] 3. End-to-end flow field prediction: Establish an end-to-end mapping model from boundary conditions to flow field characteristics through machine learning methods. This method has significant advantages in response speed, but most existing studies focus on low Reynolds number flows (Re<500) or relatively simple two-dimensional and three-dimensional flow patterns. Its application in complex three-dimensional turbulent flow fields and high Reynolds number conditions is limited, and it is difficult to accurately predict the complex temperature and pressure distribution on the surface of turbine blades.

[0008] In summary, the existing turbine blade surface field prediction technology has a significant trade-off between efficiency and accuracy. Especially for complex turbine blade flow field design, high-precision numerical simulation calculations are time-consuming and computationally expensive, while existing machine learning methods often lack prediction accuracy when dealing with complex wave structures (such as shock waves and expansion waves) and flow fields under high Reynolds number conditions. The surface of turbine blades in complex flow fields usually exhibits complex three-dimensional flow characteristics and has significant temperature and pressure distribution inhomogeneities. These characteristics make it difficult for traditional flow field prediction methods to meet the requirements of modern gas turbine performance design.

[0009] Therefore, how to quickly and accurately predict the temperature and pressure fields on the surface of turbine blades to meet the design requirements of the new generation of gas turbines is a major challenge facing the current technical field. Summary of the invention

[0010] In order to solve the technical problems existing in the prior art, that the existing turbine blade surface field prediction technology is time-consuming and has high computational cost, insufficient accuracy, and significant temperature and pressure distribution unevenness, the technical solution provided by the present invention aims to reduce computing resources and time costs, improve the prediction accuracy of complex turbine blade surface flow fields, and thus accelerate the overall process of turbine design and optimization:

[0011] A method for establishing a fast prediction model of turbine blade surface field based on a multi-channel self-attention network, comprising:

[0012] Steps for collecting turbine blade surface field data;

[0013] Steps to collect preset multi-channel self-attention network architecture;

[0014] A step of preprocessing the preset multi-channel self-attention network architecture;

[0015] The step of extracting and modeling the turbine blade surface field according to the processed multi-channel self-attention network architecture;

[0016] According to the feature extraction and modeling results, a step of establishing a "mapping relationship from boundary conditions to temperature field and pressure field on the surface of turbine blades" is performed.

[0017] Furthermore, a preferred embodiment is provided, wherein the preprocessing specifically comprises: configuring a convolutional layer, a residual Swin Transformer block, and a PixelShuffle upsampling operator in each path of the multi-channel self-attention network architecture.

[0018] Furthermore, a preferred embodiment is provided, wherein the convolution layer is used for feature integration, the residual SwinTransformer block is used to calculate the correlation between different regions in the enhanced flow field matrix, and the PixelShuffle upsampling operator is used to reconstruct the low-resolution feature map into a high-resolution output.

[0019] Furthermore, a preferred implementation is provided, which establishes a "mapping relationship from boundary conditions to temperature field and pressure field on the surface of turbine blades" through supervised learning.

[0020] Further, a preferred embodiment is provided, a device for establishing a fast prediction model of a turbine blade surface field based on a multi-channel self-attention network, comprising:

[0021] A module for collecting surface field data of turbine blades;

[0022] Collect modules of preset multi-channel self-attention network architecture;

[0023] A module for preprocessing the preset multi-channel self-attention network architecture;

[0024] A module for feature extraction and modeling of the turbine blade surface field according to the processed multi-channel self-attention network architecture;

[0025] According to the feature extraction and modeling results, a module of "mapping relationship from boundary conditions to temperature field and pressure field on the surface of turbine blades" is established.

[0026] Furthermore, a preferred embodiment is provided, a method for rapid prediction of turbine blade surface field based on a multi-channel self-attention network, comprising: a step of predicting the surface temperature field and pressure field of the blade to be measured through the mapping relationship from the boundary conditions to the surface temperature field and pressure field of the turbine blade.

[0027] Furthermore, a preferred embodiment is provided, a device for rapid prediction of turbine blade surface field based on a multi-channel self-attention network, comprising: a module for predicting the surface temperature field and pressure field of the blade to be measured through the mapping relationship from the boundary conditions to the surface temperature field and pressure field of the turbine blade.

[0028] Furthermore, a preferred embodiment is provided, in which a computer storage medium is used to store a computer program, and when the computer reads the computer program, the computer executes the described method.

[0029] Furthermore, a preferred embodiment is provided, in which a computer comprises a processor and a storage medium, and when the processor reads a computer program stored in the storage medium, the computer executes the method described.

[0030] Furthermore, a preferred embodiment is provided, a computer program product, as a computer program, which implements the method when the computer program is executed.

[0031] Compared with the prior art, the technical solution provided by the present invention is beneficial in that:

[0032] By introducing a multi-path self-attention network architecture, the accuracy and efficiency problems in complex flow field prediction are effectively solved. By adopting a multi-path feature extraction method, input data of different scales are assigned to independent paths to capture the physical field characteristics of turbine blade surfaces of different granularities. Compared with the existing single-path model, the multi-path architecture significantly improves the ability to capture flow field features, especially for complex flow fields with multi-scale features, and its prediction accuracy has been effectively improved.

[0033] The residual Swin Transformer block is used in flow field feature extraction. By introducing window-based attention calculation and sliding window mechanism, the model's ability to understand the correlation between flow fields in different regions is enhanced. Compared with traditional convolutional neural networks (CNN), Swin Transformer can capture long-range dependencies more efficiently, especially when predicting complex flow fields under high Reynolds numbers. The sliding window enables the model to share information between different sub-regions, thereby improving the accuracy of the overall prediction.

[0034] The combination of the convolutional layer and the PixelShuffle upsampling operator enables efficient reconstruction of low-resolution feature maps, thereby obtaining high-resolution prediction results. PixelShuffle avoids the common checkerboard artifact problem in traditional upsampling methods, making the predicted turbine blade surface temperature and pressure fields smoother and more natural. Compared with other upsampling methods such as bilinear interpolation or deconvolution, PixelShuffle performs more stably and accurately, especially when dealing with subtle features on the turbine blade surface, its reconstruction effect is better.

[0035] The network is trained using supervised learning and the Adam optimizer is used to optimize the network parameters, which significantly improves the convergence speed and prediction performance of the model. Compared with the simple fully connected network that has not been optimized, this solution can reach the global optimum faster during the training process and avoid falling into the local optimum problem, so that the mean square error and relative error on different test sets are kept at a low level. This training strategy effectively improves the generalization performance of the model, enabling it to still give high-quality prediction results when dealing with unseen boundary conditions.

[0036] The model performance was systematically verified using a variety of evaluation metrics (such as mean square error, peak signal-to-noise ratio, and structural similarity), showing the model's reconstruction quality and prediction accuracy for complex flow fields. The differences in these metrics between the training set and the test set were minimal, indicating that the model performed well in generalization. Compared with traditional CFD-based prediction methods, this scheme not only significantly shortened the calculation time, but also showed similar or even better results in reconstruction quality, especially when dealing with high-gradient areas, significantly reducing the probability of outliers.

[0037] It is suitable for use in the prediction of temperature and pressure fields on the surface of turbine blades. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a schematic diagram of the flow field prediction network structure;

[0039] Figure 2 Schematic diagram of sliding window self-attention;

[0040] Figure 3 Comparison cloud diagram of the model predicted surface field and the reference surface field at low Mach number;

[0041] Where (a) is the case where the isentropic Mach number is 0.4, and (b) is the case where the isentropic Mach number is 1.0;

[0042] Figure 4 Comparison cloud diagram of the model predicted surface field and the reference surface field at high Mach number;

[0043] Where (a) is the case where the isentropic Mach number is 1.4, and (b) is the case where the isentropic Mach number is 1.6;

[0044] Figure 5 Comparison cloud diagram of model predicted surface field and reference surface field;

[0045] Figure 6 Schematic diagram of data processing method. DETAILED DESCRIPTION

[0046] In order to make the advantages and benefits of the technical solution provided by the present invention more clearly reflected, the technical solution provided by the present invention is now further described in detail with reference to the accompanying drawings, specifically:

[0047] Embodiment 1: This embodiment provides a method for establishing a rapid prediction model of a turbine blade surface field based on a multi-channel self-attention network, including:

[0048] Steps for collecting turbine blade surface field data;

[0049] Steps to collect preset multi-channel self-attention network architecture;

[0050] A step of preprocessing the preset multi-channel self-attention network architecture;

[0051] The step of extracting and modeling the turbine blade surface field according to the processed multi-channel self-attention network architecture;

[0052] According to the feature extraction and modeling results, a step of establishing a "mapping relationship from boundary conditions to temperature field and pressure field on the surface of turbine blades" is performed.

[0053] Implementation method 2. This implementation method is a further limitation of the method for establishing a fast prediction model of a turbine blade surface field based on a multi-channel self-attention network provided in implementation method 1. The preprocessing is specifically as follows: a convolutional layer, a residual Swin Transformer block and a PixelShuffle upsampling operator are configured in each path of the multi-channel self-attention network architecture.

[0054] Implementation method three: This implementation method is a further limitation of the method for establishing a fast prediction model of a turbine blade surface field based on a multi-channel self-attention network provided in implementation method two. The convolution layer is used for feature integration, the residual Swin Transformer block is used to calculate the correlation between different regions in the enhanced flow field matrix, and the PixelShuffle upsampling operator is used to reconstruct the low-resolution feature map into a high-resolution output.

[0055] Implementation method 4. This implementation method is a further limitation of the method for establishing a rapid prediction model of the turbine blade surface field based on a multi-channel self-attention network provided in implementation method 1, and establishes a "mapping relationship from boundary conditions to the turbine blade surface temperature field and pressure field" through supervised learning.

[0056] Embodiment 5: This embodiment provides a device for establishing a fast prediction model of a turbine blade surface field based on a multi-channel self-attention network, comprising:

[0057] A module for collecting surface field data of turbine blades;

[0058] Collect modules of preset multi-channel self-attention network architecture;

[0059] A module for preprocessing the preset multi-channel self-attention network architecture;

[0060] A module for feature extraction and modeling of the turbine blade surface field according to the processed multi-channel self-attention network architecture;

[0061] According to the feature extraction and modeling results, a module of "mapping relationship from boundary conditions to temperature field and pressure field on the surface of turbine blades" is established.

[0062] Implementation method six: This implementation method provides a method for rapid prediction of turbine blade surface field based on a multi-channel self-attention network, including: predicting the surface temperature field and pressure field of the blade to be measured through the mapping relationship from boundary conditions to the surface temperature field and pressure field of the turbine blade provided in implementation method one.

[0063] This implementation is intended to be applied in the preliminary design stage of gas turbines and aircraft engines, replacing numerical simulation methods to improve design efficiency and reduce costs. The overall solution is based on machine learning and uses a multi-channel self-attention network to predict the surface temperature and pressure fields of turbine blades by inputting boundary conditions.

[0064] Data preparation: Turbine blade surface field data is collected and preprocessed to form training and test data sets.

[0065] Flow field prediction network construction: Build a multi-channel self-attention network and define its architecture and composition, including feature extraction, convolution calculation, and upsampling steps.

[0066] Model training: A specific training algorithm is used to optimize the network parameters to ensure that the model can accurately predict the temperature and pressure fields on the blade surface.

[0067] Performance verification and evaluation: Evaluate model performance through the test set to ensure the accuracy and stability of the prediction results.

[0068] Flow field prediction application: By using the trained model for prediction, you only need to provide normalized boundary conditions to quickly obtain the surface temperature field and pressure field of the turbine blade.

[0069] 1. Data preparation

[0070] Sampling point arrangement: The turbine blade surface is cut into 128 sections along the span direction, and 256 sampling points are evenly arranged on the suction side and pressure side along the blade arc. Starting from the trailing edge of the pressure side, across the leading edge to the trailing edge of the suction side, the three-dimensional blade surface is unfolded into a matrixed two-dimensional plane.

[0071] Data structuring: Each complete blade surface field is constructed into an array of shape 2×128×512, storing the pressure (P) and temperature (T) data of the blade surface.

[0072] Data preprocessing: Normalize the data based on statistical characteristics to eliminate the impact of data scale and improve model training results.

[0073] 2. Construction of flow field prediction network

[0074] Network architecture overview: The network is a multi-path structure consisting of three branches, each of which processes data of different sizes to capture features of different scales. The boundary condition is expanded to 32×128, and after two downsampling layers, the three paths are 32×128, 16×64, and 8×32 respectively.

[0075] Feature extraction and modeling: Each path includes a convolutional layer, a residual Swin Transformer block (RSTB), and an upsampling operator PixelShuffle. The convolutional layer is responsible for feature integration, RSTB is used to extract deep features, and PixelShuffle is used for high-resolution reconstruction to avoid checkerboard artifacts.

[0076] Sliding window self-attention mechanism: The sliding window method is introduced to increase the correlation between sub-matrices and improve the modeling ability of the model. The circular shift strategy is used to reduce the computational complexity and improve the utilization efficiency of the flow field matrix information.

[0077] 3. Model training

[0078] Training algorithm selection: Adam optimizer is used to train the model and optimize the hyperparameters in the network. The training goal is to minimize the mean square error (MSE) to ensure that the error between the predicted value and the true value is minimized.

[0079] Dataset division: The dataset is divided into a training set and a test set for model parameter optimization and performance verification, respectively. Indicators such as mean square error and relative error are used to evaluate the performance of the model during the training process.

[0080] 4. Performance verification and evaluation

[0081] Evaluation indicators: Mean square error (MSE), relative error (RE), gradient error (GL), peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and other indicators are used to comprehensively evaluate the prediction accuracy of the model under different flow field conditions.

[0082] Performance analysis: The performance of the model on the test set is close to that of the training set, indicating that the model has good generalization ability. The relative error is less than 15%, the average error is about 0.07%, and the SSIM is close to 1, indicating that the reconstruction quality is high.

[0083] 5. Flow field prediction application

[0084] Boundary condition input: The normalized boundary conditions are input into the model, and the model can output the surface temperature field and pressure field of the turbine blade within 0.1 seconds.

[0085] Prediction accuracy: The model has good reconstruction quality for high-gradient areas, which minimizes the possibility of outliers and can provide reliable data support for the design and optimization of turbine blades.

[0086] Implementation method seven. This implementation method provides a device for rapid prediction of turbine blade surface field based on a multi-channel self-attention network, including: a module for predicting the surface temperature field and pressure field of the blade to be measured through the mapping relationship from boundary conditions to the surface temperature field and pressure field of the turbine blade provided in implementation method one.

[0087] Embodiment 8: This embodiment provides a computer storage medium for storing a computer program. When the computer reads the computer program, the computer executes the method provided in embodiment 1.

[0088] Embodiment 9: This embodiment provides a computer, including a processor and a storage medium. When the processor reads a computer program stored in the storage medium, the computer executes the method provided in embodiment 1.

[0089] Embodiment 10: This embodiment provides a computer program product, which is a computer program. When the computer program is executed, the method provided in embodiment 1 is implemented.

[0090] Implementation Method XI: Combination Figure 1-6 This embodiment is described in further detail for the technical solution provided above, specifically:

[0091] 1. Data Preparation

[0092] Collect and preprocess turbine blade surface data to prepare for network model training.

[0093] Detailed description:

[0094] 1. Sampling point arrangement: The turbine blade surface is evenly cut into 128 sections along the span direction, and 256 sampling points are evenly arranged along the blade arc line in each section, including the suction side and the pressure side. The specific operation is to start from the trailing edge of the pressure side, cross the leading edge line to the trailing edge of the suction side, so as to unfold the three-dimensional blade surface into a matrixed two-dimensional plane.

[0095] 2. Data structuring: Each complete blade surface field is constructed into an array of shape 2×128×512 to store the pressure (P) and temperature (T) information of the blade surface. These arrays will serve as the input data of the model.

[0096] 3. Data normalization: Normalize the original data to eliminate the scale differences of different physical quantities and improve the stability and efficiency of training. The specific normalization method is based on the mean (μ) and standard deviation (σ) of the data, so that the data normalization range is controlled between -1 and 1.

[0097] 2. Construction of flow field prediction network

[0098] A multi-channel self-attention network is built to efficiently extract and predict the surface temperature and pressure fields of turbine blades.

[0099] Detailed description:

[0100] 1. Multi-path feature extraction: The normalized data is input into a multi-path self-attention network. The network contains three paths, each with a different data size (32×128, 16×64, and 8×32). It is first expanded to an input size of 32×128 by replicating the boundary conditions, and then the input data is adjusted to the specific size of different paths through the downsampling layer.

[0101] 2. Downsampling and convolution processing: Downsampling is achieved by using a convolution operation with a step size of 2 to compress the input data and extract features. The convolution layer is used to integrate the features of the input data and enhance the ability to capture features at different granularities.

[0102] 3. Residual Swin Transformer Block (RSTB): After the features in each path are processed by convolution, they enter the residual Swin Transformer block (RSTB), which contains multiple cascaded Swin Transformer computation blocks (STL) to capture long-range dependencies through window-based attention calculations. RSTB uses a sliding window mechanism to divide the input tensor into non-overlapping sub-matrices of size 4×4, which are flattened into labels to more effectively utilize the relevant information between regions in the flow field matrix.

[0103] 4. Upsampling (PixelShuffle): The PixelShuffle upsampling operator is used to reconstruct the low-resolution feature map into a high-resolution output to restore the temperature and pressure distribution on the blade surface and avoid the generation of checkerboard artifacts. The upsampling process converts the channel size into the size of height and width to ensure that the prediction result has a good reconstruction effect at high resolution.

[0104] 3. Model Training

[0105] The flow field prediction network model is trained so that the model can learn the mapping relationship from boundary conditions to the temperature field and pressure field on the blade surface.

[0106] Detailed description:

[0107] 1. Training data set and test data set division: The prepared normalized data set is divided into training set and test set. The training set is used for network parameter optimization, and the test set is used to verify the generalization ability of the model.

[0108] 2. Network training process: The supervised learning method is used to train the network, and the Adam optimizer is used to optimize the network parameters, with the objective function of minimizing the mean square error (MSE). The initial learning rate of the Adam optimizer is set to 0.001, and a dynamic attenuation strategy is used to ensure that the model converges stably during the training process and avoids falling into the local optimal solution.

[0109] 3. Model training results: During the training process, the prediction ability of the model is evaluated through loss functions such as mean square error and relative error, and the parameters are continuously optimized until the value of the loss function remains stable and at a low level, indicating that the model training has achieved the expected effect.

[0110] 4. Performance Verification and Evaluation

[0111] The performance of the model is evaluated through a variety of indicators to ensure that its prediction of the blade surface field has sufficient accuracy and stability.

[0112] Detailed description:

[0113] 1. Selection of evaluation indicators: Mean square error (MSE), relative error (RE), gradient error (GL), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are used as evaluation indicators of model performance.

[0114] 2. Model performance verification: Verify the performance of the model on the training set and test set. The values ​​of various indicators such as mean square error, relative error, gradient error, etc. should be very close on the training set and test set, and the relative error should be less than 15%. The average error is about 0.07%, indicating that the model has good generalization ability under different boundary conditions.

[0115] 3. Reconstruction of high gradient areas: The gradient error (GL) is used to evaluate the reconstruction quality of the model when dealing with high gradient areas on the turbine blade surface. When the GL is at a low level, it means that the model can well reconstruct the flow field distribution in the high gradient area and avoid the generation of predicted outliers.

[0116] 5. Flow field prediction application

[0117] The trained model is used to quickly predict the surface temperature and pressure fields of turbine blades to meet the needs of the turbine blade design stage.

[0118] Detailed description:

[0119] 1. Boundary condition input: The normalized boundary condition data is input into the trained model. The model predicts the temperature field and pressure field on the blade surface based on the learned mapping relationship.

[0120] 2. Prediction speed: The response speed of the model is less than 0.1 second. Compared with traditional CFD numerical simulation, its prediction time is greatly shortened, which can significantly improve the efficiency of design and optimization.

[0121] 3. Prediction result verification: The quality of the prediction results was evaluated based on the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). The PSNR of all samples was above 44dB, and the SSIM values ​​were above 0.997, indicating that the prediction results had high reconstruction quality under different conditions and could provide reliable data support for blade design.

[0122] In the specific implementation process, through the collection, preprocessing, feature extraction and model construction of turbine blade surface field data, and then to model training and verification, the efficient and rapid prediction of turbine blade surface temperature field and pressure field is finally achieved. This method effectively solves the bottleneck problem of traditional numerical simulation methods in computing resources and time cost, and provides more accurate and efficient prediction results under complex flow fields.

[0123] Compared with the existing technical route based on turbulence model correction and initialization numerical simulation, the multi-channel self-attention network architecture proposed in this implementation has stronger feature extraction capabilities and can significantly improve computational efficiency while maintaining high accuracy. In addition, through the residual Swin Transformer block and sliding window mechanism, the model's ability to handle high Reynolds numbers and complex flow fields is significantly improved. These features together ensure high-quality prediction of the turbine blade surface field.

[0124] Ultimately, the prediction results can be used in the preliminary design and optimization stages of turbine blades, effectively reducing the cost of obtaining high-quality turbine blade surface field data and improving design efficiency, providing important technical support for the new generation of gas turbines and aircraft engines.

[0125] Specific:

[0126] 1. Flow field prediction network architecture

[0127] The flow field prediction is a multi-path network with three branches, and the data size flowing in each branch is different. First, the boundary conditions are copied and expanded to 32×128, and then two downsampling layers are used, so the operation sizes of the three paths are 32×128, 16×64, and 8×32 respectively. The downsampling operation is performed using a regular layer with a stride of 2. Features of different scales are connected in stages to capture the physical field characteristics of the blade surface at different granularity levels, thereby improving the prediction accuracy. We designed the same architecture for the three paths, namely a convolutional layer (CNN), a residual Swin Transformer block (RSTB), and an upsampling operator using PixelShuffle, where the convolutional layer mainly plays the role of feature integration.

[0128] The main network architecture is as follows Figure 1 As shown in (a), RSTB can be regarded as a variant of the residual unit ResUnit, which includes 6 SwinT computing blocks (STL), 1 convolutional layer and 1 direct connection path, as shown in Figure 1 (b) STL contains two serially connected Transformer encoders, and its structure is as follows Figure 1 (c) STL uses window-based attention calculation to more effectively utilize the relevant information in the flow field matrix. Specifically, the input tensor is split into non-overlapping sub-matrices of size 4×4, which are flattened into labeled So we get the transformed input like Figure 2 (a) shown.

[0129] However, this method will lead to the solidification of features within the window and lack of relevant information within the sub-matrix, thus limiting its modeling ability. Therefore, the sliding window method is introduced. Figure 2 As shown in (b), the window is shifted by (4, 4) pixels from the original window, resulting in new sub-blocks that are usually uneven in size and larger in number. A circular shift strategy is used to reduce computational complexity. The remaining part outside the left boundary is first shifted to the right; then the top part (including the part just shifted from the left) is placed at the bottom. Subsequently, the pixels are reshaped into the same region configuration as a regular window. Once the attention calculation is completed, the reverse process is performed to restore the original layout.

[0130] Multilayer perceptron is an important basic component of neural network. It is a series of stacked fully connected layers, which is essentially a continuous linear operation. The parameters of each layer of neurons include weight matrix and the bias vector Where m is the number of neurons in the current layer, and n is the input data size. Let ζ(·) be the activation function, then the output of the neurons in the lth layer can be written as:

[0131]

[0132] Select the leaky ReLU function to satisfy the mathematical relationship described below:

[0133]

[0134] The convolution calculation unit is essentially a linear operation based on a set of small matrices called convolution kernels. The output of the l-th layer convolution unit can be written as the following filtering formula:

[0135]

[0136] where h l-1 and h l Represents the input and output of the convolution unit; w pqrk represents a convolution kernel of size P×Q×R, b k is the bias. ζ(·) is the activation function, which is used to improve the nonlinear modeling ability of the neural network.

[0137] PixelShuffle is used as an upsampling method to avoid checkerboard artifacts. The idea is to convert the channel dimension into height and width dimensions. Given a low-resolution image of C×H×W, a convolution operation is first performed to obtain r 2 The feature map is C×H×W, where r is the upsampling factor. PixelShuffle then rearranges the feature map into a high-resolution image of C×rH×rW.

[0138] 2. Training strategy

[0139] The Adam algorithm is used to train the neural network, optimize the model parameters, and the mean square error is used to calculate the corresponding residual:

[0140]

[0141] In order to evaluate the performance of the proposed model, the mean square error MSE, relative error RE and gradient error GL are used to evaluate the deviation between the predicted value and the reference value.

[0142]

[0143] Where H and W refer to the height and width of the surface field matrix, N φ=2, which specifically refers to temperature and pressure in this embodiment. GL reflects the performance of the model in dealing with high gradient areas and the probability of unexpected outliers. When GL is at a relatively low level, the model achieves good reconstruction quality for the flow field in all areas including high gradient areas, while minimizing the possibility of prediction errors or extreme outliers. On the contrary, when GL is high, the model tends to have obvious errors in high gradient areas and has a high probability of extreme outliers.

[0144] Table 1 shows the MSE, RE, and GL values ​​of the current model for various inputs on the training and test sets. The evaluation indicators of the model on the two datasets are very close, proving that the model has strong generalization. Overall, the relative error of the model prediction value is less than 15%, and the average relative error is at the order of 0.07%, indicating that the model achieves high-quality flow field prediction based on boundary conditions. As for GL, it should be comparable in magnitude to the mean absolute error (MAE). The average GL value of the model is less than 5×10 -3 , indicating that most regions have high reconstruction quality, almost eliminating the possibility of unusual errors or outliers.

[0145] Table 1 Statistical indicators of model performance

[0146]

[0147] Figure 3 and Figure 4 Shows different exit isentropic Mach numbers Ma 2s The cloud maps of the reconstructed flow field and the reference flow field and their relative errors under the conditions are obtained. In general, the model shows satisfactory reconstruction results, and there are almost no large-scale and wide-range high error areas in the relative error cloud maps. The maximum relative error of the pressure field is less than 15%, while the relative error of the temperature field is less than 2.5%. In the obvious error area, most of the relative error values ​​are limited to the range of 4% to 6%. As Ma2s continues to increase, the overall error of the model prediction value increases slightly, but still remains at a low level. It shows that the multi-channel neural network proposed in this embodiment can predict the turbine surface temperature field and pressure field under complex flow patterns with high accuracy.

[0148] In order to more comprehensively evaluate the model performance, the peak signal-to-noise ratio (PSNR) and structural similarity criterion (SSIM) are introduced as evaluation indicators. The definition of the peak signal-to-noise ratio hybrid structural similarity criterion is as follows:

[0149]

[0150] where μ cfd and σ cfd represents the mean and standard deviation of the reference data; μ Fand σ F The mean and standard deviation of the predicted data; σ cfd，F is the covariance of the two sets of data; C1 and C2 are constants to maintain numerical stability SSIM∈[0,1], and the higher the value, the better the reconstruction quality. We use the violin plot of PSNR and SSIM to illustrate the performance of the model on the dataset, which combines the distribution plot and the box plot. The distribution of PSNR and SSIM is as follows Figure 5 As shown in Figure 2, these two indicators show similar distributions in terms of training and test sets. The PSNR values ​​of all samples are above 44dB and the SSIM values ​​are above 0.997, indicating that the prediction quality is high.

[0151] In implementation:

[0152] like Figure 6 As shown in the figure, the network needs to be trained in a targeted manner, and sufficient test data should be prepared and data extraction and preprocessing should be performed. First, uniform sampling points are arranged on the surface of the turbine blade. The surface of the turbine blade is evenly cut into 128 sections along the span direction, and 256 sampling points are evenly arranged on the suction side and the pressure side along the blade arc. Starting from the trailing edge of the pressure side, across the leading edge line to the trailing edge of the suction side, the three-dimensional blade surface is unfolded into a matrixed two-dimensional plane. Finally, each complete blade surface field is constructed into an array of shape 2×128×512, which stores the pressure P and temperature data T on the blade surface. Then, according to the statistical characteristics of the data, the data is normalized based on the following formula:

[0153]

[0154] Where μ is the mean and σ is the standard deviation.

[0155] The network and its associated codes are written in Python and built on the open source deep learning framework PaddlePaddle. The network is trained on a computer platform equipped with specific hardware devices. It is usually recommended to use the Adam optimizer with default parameters.

[0156] After the model is completed, the performance of the model on the dataset is evaluated by mean square error and relative error.

[0157] When the model is trained, in the prediction stage, only the normalized boundary conditions need to be provided and the corresponding turbine blade surface temperature and pressure fields can be obtained by importing the model, with a response speed of less than 0.1s. Among them, the normalization index should adopt the statistical characteristics of the CFD data in the training set.

[0158] The technical solution provided by the present invention is further described in detail above through several specific implementation modes in order to highlight the advantages and benefits of the technical solution provided by the present invention. However, the several specific implementation modes described above are not intended to be used as limitations on the present invention. Any reasonable modification and improvement of the present invention, combination of implementation modes and equivalent substitution within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for establishing a fast prediction model for turbine blade surface field based on a multi-channel self-attention network, characterized in that: include: Steps for collecting turbine blade surface field data; Collect the steps of the preset multi-channel self-attention network architecture; A step of preprocessing the preset multi-channel self-attention network architecture; The step of extracting features and modeling the turbine blade surface field according to the processed multi-channel self-attention network architecture; According to the feature extraction and modeling results, a step of establishing a "mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field" is performed.

2. According to claim 1, a method for establishing a fast prediction model of turbine blade surface field based on a multi-channel self-attention network is characterized in that: The preprocessing specifically includes configuring a convolutional layer, a residual Swin Transformer block, and a PixelShuffle upsampling operator in each path of the multi-channel self-attention network architecture.

3. The method for establishing a fast prediction model of turbine blade surface field based on a multi-channel self-attention network according to claim 2 is characterized in that: The convolutional layer is used for feature integration, the residual Swin Transformer block is used to calculate the correlation between different regions in the enhanced flow field matrix, and the PixelShuffle upsampling operator is used to reconstruct the low-resolution feature map into a high-resolution output.

4. The method for establishing a fast prediction model of turbine blade surface field based on a multi-channel self-attention network according to claim 1 is characterized in that: Through supervised learning, a "mapping relationship from boundary conditions to the temperature field and pressure field on the surface of turbine blades" is established.

5. A device for establishing a rapid prediction model of turbine blade surface field based on a multi-channel self-attention network, characterized in that: include: A module for collecting surface field data of turbine blades; Collect modules of preset multi-channel self-attention network architecture; A module for preprocessing the preset multi-channel self-attention network architecture; A module for feature extraction and modeling of the turbine blade surface field according to the processed multi-channel self-attention network architecture; According to the feature extraction and modeling results, a module of "mapping relationship from boundary conditions to temperature field and pressure field on the surface of turbine blades" is established.

6. A fast prediction method for turbine blade surface field based on multi-channel self-attention network, characterized in that: include: The step of predicting the surface temperature field and pressure field of the blade to be measured is carried out through the mapping relationship from the boundary conditions to the surface temperature field and pressure field of the turbine blade as described in claim 1.

7. A fast prediction device for turbine blade surface field based on multi-channel self-attention network, characterized in that: include: A module for predicting the surface temperature field and pressure field of a blade to be measured by using the mapping relationship from boundary conditions to the surface temperature field and pressure field of a turbine blade as described in claim 1.

8. A computer storage medium for storing a computing program, characterized in that: When the computer reads the computer program, the computer executes the method of claim 1.

9. A computer, comprising a processor and a storage medium, characterized in that: When the processor reads the computer program stored in the storage medium, the computer executes the method of claim 1 .

10. A computer program product, being a computer program, characterized in that When the computer program is executed, the method of claim 1 is implemented.

Images