A fast prediction method for turbine blade surface field based on multi-channel self-attention network

Abstract

The application discloses a kind of based on multi-channel self-attention network turbine blade surface field fast prediction method, it is related to data prediction technical field.To solve the technical problems existing in prior art, the existing turbine blade surface field prediction technology is long in time and high in calculation cost, insufficient in accuracy, and there is significant temperature and pressure distribution unevenness, the technical scheme provided by the application is as follows: a kind of based on multi-channel self-attention network turbine blade surface field fast prediction model establishment method, comprising: collecting turbine blade surface field data;Collecting preset multi-channel self-attention network architecture;Pretreatment is carried out;According to the multi-channel self-attention network architecture after processing, the turbine blade surface field is carried out feature extraction and modeling;According to the feature extraction and modeling result, establish the mapping relationship from boundary condition to turbine blade surface temperature field and pressure field.It is suitable for being applied to the work of turbine blade surface temperature field and pressure field prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data prediction, and particularly relates to a fast prediction of a turbine blade surface field. BACKGROUND

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

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

[0004] In recent years, machine learning has gradually attracted attention in the field of fluid dynamics. Scholars try to build surrogate models to replace part of the numerical simulation to accelerate the design and optimization process. For example, through machine learning methods for rapid prediction of flow field, which can be divided into the following main routes:

[0005] 1. Machine learning method based on turbulence model correction: predict anisotropic turbulence characteristic quantities through machine learning or linear algebra methods, and correct existing turbulence models to accelerate numerical calculation. This method can improve the convergence of numerical simulation, but still needs to call a numerical solver in complex flow field prediction, with high computational cost.

[0006] 2. Flow field prediction for initializing numerical simulation: use machine learning to build a rough flow field estimate as the initial condition of numerical simulation to reduce the number of iterations and accelerate convergence. Although this method speeds up the calculation process, the prediction accuracy is limited and it is difficult to 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 a significant advantage in response speed, but existing research mostly focuses on low Reynolds number flow (Re < 500) or relatively simple two-dimensional and three-dimensional flow patterns, and its application is limited under complex three-dimensional turbulent flow field and high Reynolds number conditions, making it 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 calculation is time-consuming and costly, while existing machine learning methods often lack prediction accuracy when dealing with complex wave system structures (such as shock waves and expansion waves) and high Reynolds number conditions. The turbine blade surface in a complex flow field usually exhibits complex three-dimensional flow characteristics with significant temperature and pressure distribution non-uniformity, which makes 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 field and pressure field on the surface of turbine blades to meet the design requirements of the new generation of gas turbines is a major challenge in the current technical field. SUMMARY

[0010] To solve the technical problems of existing turbine blade surface field prediction technology, such as long time-consuming and high computational cost, insufficient accuracy, and significant temperature and pressure distribution non-uniformity, the technical solution provided by the present application aims to reduce computational resources and time cost, improve the prediction accuracy of complex turbine blade surface flow field, 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] a step of collecting turbine blade surface field data;

[0013] a step of collecting a preset multi-channel self-attention network architecture;

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

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

[0016] a step of establishing a mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field according to the feature extraction and modeling results.

[0017] Further, a preferred embodiment is provided, wherein the preprocessing is specifically configuring convolutional layers, residual Swin Transformer blocks and PixelShuffle up-sampling operators in each path of the multi-path self-attention network architecture.

[0018] Further, a preferred embodiment is provided, wherein the convolutional layers are used for feature integration, the residual Swin Transformer blocks are used for calculating the correlation between different regions in the enhanced flow field matrix, and the PixelShuffle up-sampling operators are used for reconstructing low-resolution feature maps into high-resolution outputs.

[0019] Further, a preferred embodiment is provided, wherein the mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field is established through supervised learning.

[0020] Further, a preferred embodiment is provided, wherein the turbine blade surface field fast prediction model based on the multi-path self-attention network is established by the device, which comprises:

[0021] a module for collecting turbine blade surface field data;

[0022] a module for collecting a preset multi-path self-attention network architecture;

[0023] a module for preprocessing the preset multi-path self-attention network architecture;

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

[0025] a module for establishing a mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field according to the feature extraction and modeling results.

[0026] Further, a preferred embodiment is provided, wherein the turbine blade surface field fast prediction method based on the multi-path self-attention network comprises the step of predicting the surface temperature field and pressure field of the to-be-tested blade according to the mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field.

[0027] Further, a preferred embodiment is provided, wherein the turbine blade surface field fast prediction device based on the multi-path self-attention network comprises a module for predicting the surface temperature field and pressure field of the to-be-tested blade according to the mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field.

[0028] Further, a preferred embodiment is provided, wherein the computer storage medium is used for storing a computer program, and when the computer reads the computer program, the computer executes the method.

[0029] Further, a preferred embodiment is provided, a computer comprising a processor and a storage medium, the computer executing the method when the processor reads a computer program stored in the storage medium.

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

[0031] Compared with the prior art, the technical solution provided by the present application has the advantages that:

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

[0033] The residual Swin Transformer block is used in the flow field feature extraction, which enhances the understanding ability of the model to the correlation between different regional flow fields by introducing window-based attention calculation and sliding window mechanism. Compared with traditional convolutional neural networks (CNN), Swin Transformer can more efficiently capture long-range dependencies, especially when predicting complex flow fields at 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 convolutional layers and PixelShuffle up-sampling operators realizes the efficient reconstruction of low-resolution feature maps, thereby obtaining high-resolution prediction results. PixelShuffle avoids the checkerboard artifact problem commonly seen in traditional up-sampling methods, making the predicted turbine blade surface temperature and pressure field smoother and more natural. Compared with other up-sampling 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 by supervised learning, and the Adam optimizer is used to optimize the parameters of the network, which significantly improves the convergence speed and prediction performance of the model. Compared with a simple fully connected network without optimization, the scheme can reach the global optimum faster in the training process and avoid the problem of falling into local optimum, so that the mean square error and relative error on different test sets remain at a low level. This training strategy effectively improves the generalization performance of the model, so that it can still give high-quality prediction results when dealing with unseen boundary conditions.

[0036] The model performance is systematically verified by various evaluation indicators such as mean square error, peak signal-to-noise ratio and structural similarity, which shows the reconstruction quality and prediction accuracy of the model for complex flow fields. The difference between these indicators in the training set and the test set is very small, indicating that the model performs well in generalization ability. Compared with traditional CFD-based prediction methods, this scheme not only greatly shortens the calculation time, but also shows similar or even better results in reconstruction quality, especially when dealing with high gradient areas, significantly reducing the probability of abnormal value occurrence.

[0037] It is suitable for application in the working of turbine blade surface temperature field and pressure field prediction. BRIEF DESCRIPTION OF DRAWINGS

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

[0039] Figure 2 It is a sliding window self-attention schematic diagram;

[0040] Figure 3 It is a comparison cloud diagram of the model predicted surface field and the reference surface field under low Mach number;

[0041] Wherein (a) is the case of isentropic Mach number 0.4, (b) is the case of isentropic Mach number 1.0;

[0042] Figure 4 It is a comparison cloud diagram of the model predicted surface field and the reference surface field under high Mach number;

[0043] Wherein (a) is the case of isentropic Mach number 1.4, (b) is the case of isentropic Mach number 1.6;

[0044] Figure 5 It is a comparison cloud diagram of the model predicted surface field and the reference surface field;

[0045] Figure 6 It is a schematic diagram of data processing method. DETAILED DESCRIPTION

[0046] In order to make the advantages and beneficial effects of the technical solutions provided by the present application clearer, the technical solutions provided by the present application will be described in further detail below in conjunction with the drawings. Specifically,

[0047] Embodiment I provides a method for establishing a fast prediction model of a turbine blade surface field based on a multi-path self-attention network, comprising:

[0048] a step of collecting turbine blade surface field data;

[0049] a step of collecting a preset multi-path self-attention network architecture;

[0050] a step of preprocessing the preset multi-path self-attention network architecture;

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

[0052] a step of establishing a mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field according to the feature extraction and modeling results.

[0053] Embodiment II is a further limitation of the method for establishing a fast prediction model of a turbine blade surface field based on a multi-path self-attention network provided by Embodiment I, and the preprocessing specifically includes configuring a convolutional layer, a residual Swin Transformer block and a PixelShuffle upsampling operator in each path of the multi-path self-attention network architecture.

[0054] Embodiment III is a further limitation of the method for establishing a fast prediction model of a turbine blade surface field based on a multi-path self-attention network provided by Embodiment II, wherein the convolutional layer is used for feature integration, the residual Swin Transformer block is used for calculating the correlation between different regions in the enhanced flow field matrix, and the PixelShuffle upsampling operator is used for reconstructing a low-resolution feature map into a high-resolution output.

[0055] Embodiment IV is a further limitation of the method for establishing a fast prediction model of a turbine blade surface field based on a multi-path self-attention network provided by Embodiment I, and the mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field is established through supervised learning.

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

[0057] a module for collecting turbine blade surface field data;

[0058] a module for collecting a 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] a module for establishing a mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field according to the feature extraction and modeling results.

[0062] Embodiment six, the embodiment provides a turbine blade surface field fast prediction method based on a multi-channel self-attention network, comprising: predicting the blade surface temperature field and pressure field to be measured through the mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field provided by the embodiment one.

[0063] This embodiment is intended to be applied to the preliminary design stage of gas turbines and aero-engines, and to replace numerical simulation methods to improve design efficiency and reduce costs. The overall scheme is based on machine learning, adopts a multi-channel self-attention network, and predicts the surface temperature field and pressure field of the turbine blade by inputting boundary conditions.

[0064] Data preparation: Collect turbine blade surface field data and preprocess to form training and test data sets.

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

[0066] Model training: use a specific training algorithm to optimize network parameters to ensure that the model can accurately predict the blade surface temperature field and pressure field.

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

[0068] Flow field prediction application: use the trained model for prediction, and only provide the 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: 128 cross sections are cut along the spanwise direction on the surface of the turbine blade, and 256 sampling points are uniformly arranged along the blade profile on the suction side and the pressure side. From the pressure side trailing edge, across the leading edge to the suction side trailing edge, the three-dimensional blade surface is unfolded into a matrixed two-dimensional plane.

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

[0072] Data preprocessing: Normalization is performed based on statistical characteristics to eliminate the scale effect of the data and improve the model training effect.

[0073] 2. Flow field prediction network construction

[0074] Network architecture overview: The network is a multi-path structure, including three branches, each branch processing data of different sizes to capture different scale features. 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, the RSTB is used to extract deep features, and the 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 computational complexity is reduced through a cyclic shift strategy, and the efficiency of flow field matrix information utilization is improved.

[0077] 3. Model training

[0078] Training algorithm selection: The Adam optimizer is used to train the model, and the hyperparameters in the network are optimized. 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] Data set division: The data set is divided into training set and test set, which are used for model parameter optimization and performance verification respectively. The mean square error, relative error and other indicators are used to evaluate the performance of the model in 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 on 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, which 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 in high gradient areas, minimizes the possibility of abnormal values, and can provide reliable data support for turbine blade design and optimization.

[0086] Embodiment seven, the embodiment provides a turbine blade surface field fast prediction device based on a multi-channel self-attention network, comprising: a module for predicting the surface temperature field and pressure field of the turbine blade to be measured through the mapping relationship from the boundary condition to the surface temperature field and pressure field of the turbine blade provided in embodiment one.

[0087] Embodiment eight, the 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 one.

[0088] Embodiment nine, the embodiment provides a computer comprising a processor and a storage medium, when the processor reads the computer program stored in the storage medium, the computer executes the method provided in embodiment one.

[0089] Embodiment ten, the embodiment provides a computer program product as a computer program, when the computer program is executed, the method provided in embodiment one is realized.

[0090] Embodiment eleven, in combination Figures 1-6 This embodiment is described in detail, and the specific embodiments provided above are further described in detail, specifically:

[0091] I. Data preparation

[0092] Collect and pretreat 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 spanwise direction, and 256 sampling points are evenly arranged along the camber line on each section, including the suction side and the pressure side. The specific operation is to start from the pressure side trailing edge, cross the leading edge line to the suction side trailing edge, so as to spread the three-dimensional blade surface into a matrixed two-dimensional plane.

[0095] 2. Data structuring: Each complete leaf surface field is structured as an array with shape 2x128x512 to store the pressure (P) and temperature (T) information of the leaf surface. These arrays will serve as input data for the model.

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

[0097] II. Flow field prediction network construction

[0098] A multi-path self-attention network is built to efficiently extract and predict the temperature and pressure fields on the turbine blade surface.

[0099] Detailed description:

[0100] 1. Multi-path feature extraction: The normalized data is input into the multi-path self-attention network. The network contains three paths, each with different data sizes (32x128, 16x64, and 8x32 respectively). First, the boundary conditions are expanded to 32x128 input size by copying, and then the input data is adjusted to the specific size of each path through the down-sampling layer.

[0101] 2. Down-sampling and convolution processing: Convolution operation with a step size of 2 is used to realize down-sampling, so as to compress the input data and extract features. The convolution layer is used to integrate features and enhance the ability to capture features at different granularities.

[0102] 3. Residual Swin Transformer Block (RSTB): After the convolution processing of the features in each path, they enter the Residual Swin Transformer Block (RSTB), which contains multiple Swin Transformer Calculation Blocks (STL) connected in series. The RSTB uses a sliding window mechanism to divide the input tensor into non-overlapping sub-matrices of size 4x4, which are flattened into tokens. This allows more effective use of the correlation information between different regions in the flow field matrix.

[0103] 4. Up-sampling (PixelShuffle): The PixelShuffle up-sampling operator is used to reconstruct the low-resolution feature map into a high-resolution output, so as to restore the temperature and pressure distribution on the blade surface and avoid the generation of chessboard artifacts. The up-sampling process converts the channel size into height and width size, ensuring that the prediction results have good reconstruction effect at high resolution.

[0104] III. Model Training

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

[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 parameter optimization of the network, and the test set is used for verifying the generalization ability of the model.

[0108] 2. Network training process: The network is trained by using supervised learning method, and the network parameters are optimized by using Adam optimizer, taking the mean square error (MSE) as the objective function. The initial learning rate of Adam optimizer is set to 0.001, and dynamic decay strategy is adopted to ensure the stable convergence of the model in the training process and avoid falling into local optimal solution.

[0109] 3. Model training result: In the training process, the prediction ability of the model is evaluated by the loss function such as mean square error and relative error, and the parameters are continuously optimized until the value of the loss function remains stable and low, indicating that the training of the model reaches the expected effect.

[0110] IV. Performance verification and evaluation

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

[0112] Detailed description:

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

[0114] 2. Model performance verification: The performance of the model on the training set and test set is verified, the values of mean square error, relative error, gradient error and other indicators on the training set and test set should be very close, 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. High gradient area reconstruction: The gradient error (GL) is used to evaluate the reconstruction quality of the model in processing the high gradient area of the turbine blade surface. When GL is at a low level, it indicates that the model can well reconstruct the flow field distribution in the high gradient area and avoid the generation of abnormal values in prediction.

[0116] V. Flow field prediction application

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

[0118] DETAILED DESCRIPTION

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

[0120] 2. Prediction speed: the response speed of the model is less than 0.1 seconds, which is significantly shorter than the traditional CFD numerical simulation, and can significantly improve the efficiency of design and optimization.

[0121] 3. Prediction result verification: the quality of the prediction result is evaluated according to the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). The PSNR of all samples is above 44dB, and the SSIM value is above 0.997, indicating that the prediction result has high reconstruction quality under different conditions, and can provide reliable data support for the design of the blade.

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

[0123] Compared with the existing technology route based on turbulence model correction and initialization numerical simulation, the multi-channel self-attention network architecture proposed in this embodiment has stronger feature extraction capability, which can greatly improve the calculation efficiency while maintaining high accuracy. In addition, through the residual Swin Transformer block and the sliding window mechanism, the processing capability of the model for high Reynolds number and complex flow field is significantly improved, which together guarantees the high-quality prediction of the turbine blade surface field.

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

[0125] In particular

[0126] 1. Flow field prediction network architecture

[0127] The flow field prediction is a multi-path network, containing three branches with different data size in each branch. First, the boundary condition is replicated and expanded to 32x128, then two down-sampling layers are adopted, so the operation size of the three paths are 32x128, 16x64 and 8x32 respectively. The down-sampling operation is performed by a regular layer with a stride of 2. Features of different scales are connected hierarchically to capture the physical field characteristics of the blade surface at different granularity levels, thereby improving the prediction accuracy. We design the same architecture for the three paths, i.e., a convolutional layer (CNN), a residual Swin Transformer block (RSTB) and an up-sampling operator using PixelShuffle, where the convolutional layer mainly undertakes the role of feature integration.

[0128] The network architecture is shown in FIG. 1 Figure 1 (a), where the RSTB can be regarded as a variant of the residual unit ResUnit, which includes 6 SwinT computation blocks (STL), 1 convolutional layer and 1 direct connection path, as shown in FIG. 1 Figure 1 (b). The STL contains two serially connected Transformer encoders, whose structure is shown in FIG. 1 Figure 1 (c). The STL uses window-based attention computation to more effectively utilize the relevant information in the flow field matrix. Specifically, the input tensor is divided into non-overlapping sub-matrices of size 4x4, which are flattened into the label to obtain the transformed input as shown in FIG. 1 Figure 2 (a).

[0129] However, this method causes the features within the window to be fixed, lacking relevant information within the sub-matrices, thereby limiting its modeling ability. Therefore, a sliding window method is introduced. As shown in FIG. 1 Figure 2 (b), the window is displaced by (4, 4) pixels from the original window, thereby generating new sub-blocks that are usually uneven in size and more in number. A cyclic 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 reorganized into the same regional configuration as the regular window. Once the attention computation is completed, the reverse process is performed to restore the original layout.

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

[0131]

[0132] The leaky ReLU function is selected, which satisfies the mathematical relationship as 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 of convolution units can be written as a filtering formula as follows:

[0135]

[0136] where h l-1 and h l represent the input and output of the convolution unit; w pqrk represents a convolution kernel with a size of P x Q x R, b k is the bias. ζ(·) is an activation function, which is used to improve the non-linear 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 size into height and width size. Given a low-resolution image of C x H x W, first perform a convolution operation to obtain a feature map of r 2 C x H x W, where r is the upsampling factor. Then PixelShuffle rearranges the feature map into a high-resolution image of C x rH x rW.

[0138] 2. Training strategy

[0139] The Adam algorithm is used to train the neural network, 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, the relative error RE and the 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, in the present embodiment. GL reflects the performance of the model in handling high gradient regions, as well as the probability of producing unexpected outliers. When GL is at a relatively low level, the model achieves good reconstruction quality for the flow field of all regions, including high gradient regions, while minimizing the likelihood of prediction errors or extreme outliers. Conversely, when GL is at a high level, the model tends to have significant errors in high gradient regions, and there is 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 metrics of the model on the two data sets are very close, demonstrating the strong generalizability of the model. Overall, the relative error of the model prediction value is less than 15%, and the average relative error is on the order of 0.07%, indicating that the model achieves high-quality flow field prediction from boundary conditions. As for GL, it should be on the order of the mean absolute error (MAE). The average GL value of the model is less than 5 x 10 -3 , indicating that most regions have high reconstruction quality, almost eliminating the possibility of abnormal errors or outliers.

[0145] Table 1 Statistical indicators of model performance

[0146]

[0147] Figure 3 and Figure 4 shows the cloud images of the reconstructed flow field and the reference flow field under different outlet isentropic Mach numbers Ma 2s and their relative errors. Overall, the model shows satisfactory reconstruction results, and the relative error cloud image almost does not have large-scale and large-range high error areas. The maximum relative error of the pressure field is less than 15%, and the relative error of the temperature field is less than 2.5%. In the obvious error region, most of the relative error values are limited to the range of 4% to 6%. As Ma2s continues to rise, the overall error of the model prediction value increases slightly, but still maintains a low level. This shows that the multi-pass neural network proposed in the present embodiment can predict the turbine surface temperature field and pressure field under complex flow patterns with high accuracy.

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

[0149]

[0150] where μ cfd and σ cfd represent the mean and standard deviation of the reference data; μ Fand σ F Mean and standard deviation of prediction data; σ cfd，F Covariance of two groups of data; C1 and C2 are constants to maintain numerical stability SSIM ∈ [0, 1], the higher the value represents the better 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 box plot. The distribution of PSNR and SSIM is shown in Figure 5 , the two indicators show similar distribution in training set and test set. The PSNR value of all samples is above 44dB, and the SSIM value is above 0.997, indicating high prediction quality.

[0151] In implementation:

[0152] As shown in Figure 6 , the network needs to be trained specifically, and sufficient experimental 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 turbine blade surface is uniformly cut into 128 sections along the spanwise direction, and 256 sampling points are uniformly arranged on the suction side and pressure side along the blade profile curve. Starting from the pressure side trailing edge, across the leading edge line to the suction side trailing edge, the three-dimensional blade surface is unfolded into a matrixed two-dimensional plane. Finally, each complete blade surface field is constructed into an array with a shape of 2x128x512, storing the pressure P and temperature data T of 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 the accompanying code are written in Python language and built based on the open source deep learning framework PaddlePaddle. The network is trained on a computer platform equipped with specific hardware devices, and the Adam optimizer is usually recommended 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 field and pressure field can be obtained by importing the model, with a response speed less than 0.1s. Among them, the normalization index should use the statistical characteristics of the CFD data in the training set.

[0158] The technical solutions of the present application are described in further detail through several specific embodiments above, in order to highlight the advantages and benefits of the technical solutions provided by the present application. However, the above several specific embodiments are not used as a limitation to the present application, and any reasonable modifications and improvements, combinations and equivalent replacements, etc. of the present application within the scope of the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for building a fast prediction model of turbine blade surface field based on a multi-pass self-attention network, characterized in that, The application relates to a method for establishing a mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field. The method comprises the following steps: a step of collecting turbine blade surface field data; Specifically: sampling point arrangement: the turbine blade surface is cut into 128 sections along the spanwise direction, and 256 sampling points are uniformly arranged along the blade profile curve on the suction side and the pressure side; a three-dimensional blade surface is unfolded into a matrixed two-dimensional plane from the pressure side trailing edge, across the leading edge to the suction side trailing edge; data structuring: each complete blade surface field is structured into an array with a shape of 2*128*512, and pressure and temperature data of the blade surface are stored; a step of establishing a preset multi-path self-attention network architecture; The multi-path self-attention network architecture is a multi-path structure, comprising three branches, each branch processing data of different sizes, and the boundary condition being expanded to 32*128, and passing through two down-sampling layers, the three paths being 32*128, 16*64 and 8*32 respectively; Each branch comprises a convolution layer, a residual Swin Transformer block and a up-sampling operator PixelShuffle, the convolution layer being responsible for feature integration, the residual Swin Transformer block being used for extracting deep features, and the PixelShuffle being used for high-resolution reconstruction; The multi-path self-attention network architecture introduces a sliding window method, increases the correlation between sub-matrices, and reduces the computational complexity through a cyclic shift strategy; a step of preprocessing the preset multi-path self-attention network architecture; The preprocessing specifically comprises: configuring a convolution layer, a residual Swin Transformer block and a PixelShuffle up-sampling operator in each path of the multi-path self-attention network architecture; a step of extracting features and modeling the turbine blade surface field according to the processed multi-path self-attention network architecture; a step of establishing a mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field according to the feature extraction and modeling results; 2. The method of claim 1, wherein the method is characterized by: a step of obtaining a trained fast prediction model.

3. The method of claim 1, wherein the method is characterized by: The convolution layer is used for feature integration, the residual Swin Transformer block is used for calculating the correlation between different regions in the enhanced flow field matrix, and the PixelShuffle up-sampling operator is used for reconstructing a low-resolution feature map into a high-resolution output.

4. A multi-pass self-attention network-based turbine blade surface field fast prediction model establishment device, characterized in that, The mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field is established through supervised learning. The application relates to a method for establishing a mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field. The method comprises the following steps: a module for collecting turbine blade surface field data; Specifically: sampling point arrangement: the turbine blade surface is cut into 128 sections along the spanwise direction, and 256 sampling points are uniformly arranged along the blade profile curve on the suction side and the pressure side; a three-dimensional blade surface is unfolded into a matrixed two-dimensional plane from the pressure side trailing edge, across the leading edge to the suction side trailing edge; data structuring: each complete blade surface field is structured into an array with a shape of 2*128*512, and pressure and temperature data of the blade surface are stored; a module for establishing a preset multi-path self-attention network architecture; The multi-path self-attention network architecture is a multi-path structure, including three branches, each branch processing data of different sizes, the boundary condition being expanded to 32x128, passing through two down-sampling layers, and the three paths being 32x128, 16x64 and 8x32 respectively; Each branch includes a convolutional layer, a residual Swin Transformer block and a up-sampling operator PixelShuffle, the convolutional layer being responsible for feature integration, the residual Swin Transformer block being used for extracting deep features, and the PixelShuffle being used for high-resolution reconstruction; The multi-path self-attention network architecture introduces a sliding window method, increases the correlation between sub-matrices, and reduces the computational complexity through a cyclic shift strategy; A module for preprocessing the preset multi-path self-attention network architecture; The preprocessing specifically includes configuring a convolutional layer, a residual Swin Transformer block and a PixelShuffle up-sampling operator in each path of the multi-path self-attention network architecture; A module for feature extraction and modeling of the turbine blade surface field according to the processed multi-path self-attention network architecture; A module for establishing a mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field according to the feature extraction and modeling results; A module for obtaining a trained fast prediction model.

5. A method for fast prediction of turbine blade surface field based on multi-pass self-attention network, characterized in that, Comprising: A step of predicting the surface temperature field and pressure field of a to-be-tested blade through the mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field of claim 1.

6. A device for fast prediction of turbine blade surface field based on multi-pass self-attention network, characterized in that, Comprising: A module for predicting the surface temperature field and pressure field of a to-be-tested blade through the mapping relationship from boundary conditions to turbine blade surface temperature field and pressure field of claim 1.

7. 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.

8. 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.

9. Computer program product as computer program, characterized in that When the computer program is executed, the method of claim 1 is implemented.

Images