Three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network

Abstract

The invention provides a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and a neural network. The method comprises the steps of obtaining a simulated ice shape two-dimensional slice image under a first target icing working condition; inputting the simulated ice-shaped two-dimensional slice image data into a pre-trained first target neural network, and predicting ice-shaped two-dimensional slice images of different spanwise positions of the wing under a first target icing condition, the first target neural network is obtained by training according to the simulated ice shape two-dimensional slice image and a test ice shape two-dimensional slice image obtained by an icing wind tunnel test; and performing three-dimensional reconstruction on the ice-shaped two-dimensional section images at different spanwise positions to obtain a predicted wing icing three-dimensional geometrical shape. According to the method, the neural network is used for extracting main features of the icing form, and rapid prediction and reconstruction of the three-dimensional icing form are achieved. Through multi-fidelity data fusion, prediction result precision can be improved, calculation complexity can be reduced, and icing prediction precision and real-time performance can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of wing icing prediction, and particularly relates to a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network. Background Art

[0002] Icing poses a serious threat to aircraft flight safety, especially in the cold and humid high-altitude flight environment. Wing icing can lead to a decrease in lift, an increase in drag, and even change the aerodynamic characteristics of the aircraft, seriously affecting flight performance and safety.

[0003] With the development of computational fluid dynamics (CFD), aircraft icing simulation has become an important tool for aircraft design and certification. However, there are still problems with the accuracy of CFD simulation at present. Especially when simulating icing on large-scale three-dimensional airfoils, it is unable to accurately reproduce the complex ice shapes (such as "scaly" or "lobster tail" types) observed in experiments. In addition, the accuracy of CFD is also limited by the ice density assumption, the limitations of roughness modeling, and the non-linear effects of complex heat transfer processes. The improvement of CFD simulation accuracy awaits further theoretical improvement and computational performance enhancement.

[0004] Although CFD has been widely used, experimental verification is still indispensable, especially in capturing the detailed characteristics of icing phenomena and in the later stage of aircraft design. Traditional icing prediction methods usually rely on experimental data or numerical simulation, but these methods often need to make compromises between data accuracy and acquisition cost, and it is difficult to achieve both, resulting in high computational costs or low accuracy of existing icing prediction methods. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a three-dimensional icing rapid prediction method and device based on multi-fidelity data fusion and neural network to meet the requirements of ensuring icing prediction accuracy and reducing the amount of calculation.

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

[0007] According to the first aspect, the present invention provides a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network, including: obtaining a two-dimensional slice image of the simulated ice shape under the first target icing condition; inputting the two-dimensional slice image of the simulated ice shape into a pre-trained first target neural network to predict the two-dimensional slice images of the ice shapes at different spanwise positions of the wing under the first target icing condition, where the first target neural network is trained according to the two-dimensional slice images of the simulated ice shape and the two-dimensional slice images of the experimental ice shape obtained from the icing wind tunnel test; performing three-dimensional reconstruction on the two-dimensional sectional images of the ice shapes at different spanwise positions to obtain the predicted three-dimensional geometric shape of the wing icing.

[0008] Optionally, the first target neural network is a U-net convolutional neural network. The U-net convolutional neural network includes an encoder and a decoder. The encoder includes multiple convolutional blocks and a max pooling layer. Each convolutional block includes convolution, batch normalization, ReLU activation, and a residual. The max pooling layer converts the multi-dimensional features of the two-dimensional ice shape cross-sectional data at the first resolution into a second resolution, which is lower than the first resolution but extracts the deep features of the data. The decoder gradually restores the image resolution.

[0009] Optionally, the training process of the U-net convolutional neural network is as follows: Determine the simulated ice shape two-dimensional slice image according to numerical simulation, and determine the experimental ice shape two-dimensional slice image according to the icing wind tunnel test; Use the simulated ice shape two-dimensional slice image as a sample and the experimental ice shape two-dimensional slice image as a sample learning label to form a training set and a test set; There is a corresponding relationship between the positions of the simulated ice shape and the experimental ice shape, and the spanwise coordinates are consistent; Input the training set into the U-net convolutional neural network for training, adjust the weights of MSE, SSIM, and edge loss in the preset loss function. During the training process, the weight of MSE remains unchanged, and gradually increase the weights of SSIM and edge loss until the network converges; Input the test set into the trained U-net convolutional neural network to obtain the prediction result. Among them, the test set contains ice shape cross-sectional data that is similar but not repeated to the training set; Perform error analysis on the prediction result according to the root mean square error formula to evaluate the performance of the trained U-net convolutional neural network. When the preset performance requirements are not met, adjust the division ratio of the training set and the test set, the number of U-net network layers, and the weights of each item in the loss function, and repeat the training process until the preset performance requirements are met.

[0010] Optionally, the preset loss function Loss of the U-net convolutional neural network is:

[0011] Loss = w mse ·MSE(y true , y pred ) + w ssim ·(1 - SSIM(y true , y pred )) + w edge ·MSE(Edge(y true ), Edge(y pred ));

[0012] Among them, y true is the experimental ice shape two-dimensional slice image; y pred is the image predicted by the U-net convolutional neural network. Edge(y) represents extracting the edge information of the image y through the Sobel operator. w mse , w ssim , w edgeThey are the weight coefficients of MSE, SSIM, and edge loss respectively.

[0013] Optionally, perform three-dimensional reconstruction on the two-dimensional cross-sectional images of ice shapes at different spanwise positions to obtain the predicted three-dimensional geometric shape of wing icing, including: filling the gaps between adjacent slices of the two-dimensional cross-sectional images of ice shapes at different spanwise positions by using an interpolation method to obtain ice shape data; organizing the ice shape data in a three-dimensional grid, dividing the three-dimensional grid into multiple voxels, each voxel having eight corner points, and each corner point having a numerical value; for any voxel, comparing the numerical values of the eight corner points with a preset isosurface value to determine the relationship between the corner points and the isosurface; selecting a triangle combination from a preset template to represent the isosurface according to the relationship between the corner points and the isosurface; constructing the three-dimensional geometric shape of wing icing according to the triangle combination.

[0014] Optionally, a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network further includes: obtaining the test two-dimensional cross-sectional images of ice shapes obtained according to the simulated ice shape two-dimensional slice images and icing wind tunnel tests under a second target icing condition; thawing the target network parameters in the first target neural network, and training the thawed first target neural network by using the test two-dimensional cross-sectional images of ice shapes obtained according to the simulated ice shape two-dimensional slice images and icing wind tunnel tests under the second target icing condition, and adjusting the thawed target network parameters to obtain a second target neural network.

[0015] According to a second aspect, a three-dimensional icing rapid prediction device based on multi-fidelity data fusion and neural network includes: a slice acquisition module, configured to acquire the simulated two-dimensional cross-sectional images of ice shapes under a first target icing condition; a prediction module, configured to input the simulated two-dimensional cross-sectional images of ice shapes into a pre-trained first target neural network to predict the two-dimensional cross-sectional images of ice shapes at different spanwise positions of the wing under the first target icing condition, where the first target neural network is trained according to the simulated two-dimensional cross-sectional images of ice shapes and the test two-dimensional cross-sectional images of ice shapes obtained from icing wind tunnel tests; a construction module, configured to perform three-dimensional reconstruction on the two-dimensional cross-sectional images of ice shapes at different spanwise positions to obtain the predicted three-dimensional geometric shape of wing icing.

[0016] According to a third aspect, an embodiment of the present invention provides an electronic device, where the device includes: a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor executes the steps of the three-dimensional icing rapid prediction method according to the first aspect or any implementation manner of the first aspect.

[0017] According to a fourth aspect, an embodiment of the present invention provides a computer storage medium, on which computer instructions are stored, and when the instructions are executed by a processor, the steps of the three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network according to the first aspect or any implementation manner of the first aspect are implemented.

[0018] The present invention provides a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network, develops a deep learning model that fuses experimental data and simulation data, uses a neural network to extract the main features of the icing morphology, and effectively realizes the rapid prediction and reconstruction of the three-dimensional icing morphology. Through the multi-fidelity data fusion technology, combined with the calibration of high and low precision data, the accuracy of the prediction result is significantly improved. Therefore, the proposed three-dimensional icing prediction method based on data fusion and neural network reduces the computational complexity and improves the accuracy and real-time performance of icing prediction, and has wide engineering application value.

[0019] Other advantages, objectives, and features of the present invention will be described in the following specification, and to some extent, will be obvious to those skilled in the art, or those skilled in the art can obtain teachings from the practice of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the following specification. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to make the objectives, technical solutions, and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0021] Figure 1 It is a specific example flowchart of a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network in the present invention;

[0022] Figure 2 It is a schematic diagram of the overall process of the three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network proposed by the present invention;

[0023] Figure 3 It is a schematic diagram of the structure of the U-net network in the present invention;

[0024] Figure 4 It is a modeling effect diagram of predicting the three-dimensional icing of the NACA0012 30° swept wing based on the U-net network in the present invention;

[0025] Figure 5 It is a three-dimensional icing prediction effect diagram based on the U-net network and transfer learning method for other icing conditions and airfoil conditions in the present invention;

[0026] Figure 6 It is a schematic block diagram of a specific example of an electronic device in an embodiment of the present invention. Detailed Implementation Modes

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, and it can also be the communication inside two components. It can be a wireless connection or a wired connection. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0029] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0030] An embodiment of the present invention provides a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network, as Figure 1 shown, including:

[0031] S101, obtaining a two-dimensional slice image of the simulated ice shape under the first target icing condition;

[0032] S102, inputting the two-dimensional slice image of the simulated ice shape into a pre-trained first target neural network to predict the two-dimensional slice images of the ice shape at different spanwise positions of the wing under the first target icing condition. The first target neural network is trained according to the two-dimensional slice images of the simulated ice shape and the two-dimensional slice images of the experimental ice shape obtained from the icing wind tunnel test;

[0033] S103, performing three-dimensional reconstruction on the two-dimensional cross-sectional images of the ice shape at different spanwise positions to obtain the predicted three-dimensional geometric shape of the wing icing.

[0034] Exemplarily, the first target icing condition can be frost ice, clear ice, or mixed ice, etc. The specific method for obtaining the two-dimensional slice image of the simulated ice shape under the first target icing condition can be numerical simulation using Fluent software and FENSAP-ICE software.

[0035] The first target neural network can be a convolutional neural network (U-net). The process of training the convolutional neural network according to the two-dimensional slice images of the simulated ice shape and the two-dimensional slice images of the experimental ice shape obtained from the icing wind tunnel test is as Figure 2As shown in the figure, first, numerical simulation is used to obtain two-dimensional slice images of low-precision simulated ice shapes, and icing wind tunnel tests are used to obtain two-dimensional slice images of high-precision test ice shapes. The two-dimensional slice images of high-precision test ice shapes are used as learning labels, and the two-dimensional slice images of low-precision simulated ice shapes are used as samples to construct a training set and a test set. Then, a first target neural network is built, and the training set is input into the first target neural network for training until the network converges; the trained first target neural network is verified using the test set, and the root mean square error (RMSE) is used to evaluate the network performance to ensure the generalization ability of the model on the test set and prevent overfitting, thereby obtaining a trained first target neural network.

[0036] Finally, the two-dimensional cross-sectional images of the ice shape are three-dimensionally reconstructed using the Marching Cubes algorithm to obtain the three-dimensional geometric shape of the wing icing. After obtaining the complete three-dimensional geometric shape of the wing icing, these three-dimensional geometric shapes of the wing icing can be used for subsequent flight performance analysis, especially for evaluating the impact of icing on the aerodynamic performance of the wing.

[0037] The following gives a specific implementation scenario: When it is necessary to predict the icing situation of the wing, icing wind tunnel tests can be carried out for the current icing conditions of the wing to obtain two-dimensional slice images of the test ice shape. For example, two-dimensional cross-sectional images of the test ice shape within the range of 0.8 m - 1 m on the wing are used as sample learning labels, and two-dimensional slice images of the simulated ice shape at the same position obtained by simulation are used as samples. After training the first target neural network with the two-dimensional slice images of the test ice shape and the two-dimensional slice images of the simulated ice shape, a trained first target neural network is obtained. At this time, the first target neural network has learned the mapping relationship between the simulated ice shape and the test ice shape. Therefore, the first target neural network can predict and restore the icing situation of the wing at corresponding positions in a wider wingspan direction under the same working conditions according to the two-dimensional slice images of the simulated ice shape.

[0038] The embodiment of the present invention provides a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural networks, develops a deep learning model that fuses experimental data and simulation data, extracts the main features of the icing morphology using neural networks, and effectively realizes the rapid prediction and reconstruction of the three-dimensional icing morphology. Through the multi-fidelity data fusion technology and the calibration of high- and low-precision data, the accuracy of the prediction results is significantly improved. Therefore, the proposed three-dimensional icing prediction method based on data fusion and neural networks reduces the computational complexity and improves the accuracy and real-time performance of icing prediction, and has wide engineering application value.

[0039] As an alternative implementation, the first target neural network is a U-net convolutional neural network. The U-net convolutional neural network includes an encoder and a decoder. The encoder includes multiple convolutional blocks and a max-pooling layer. Each convolutional block includes convolution, batch normalization, ReLU activation, and a residual. The max-pooling layer converts the multi-dimensional features of the two-dimensional ice shape cross-sectional data at the first resolution into a second resolution, which is lower than the first resolution but extracts the deep features of the data. The decoder gradually restores the image resolution.

[0040] Exemplarily, a U-net network is constructed according to hyperparameters (such as the number of network layers and the number of nodes in the bottleneck layer). The U-net network, as Figure 3 shown, is divided into two parts: an encoder and a decoder. The encoder contains multiple convolutional blocks, and each block contains convolution, batch normalization, ReLU activation, and residual connections. Then, the max-pooling layer converts the multi-dimensional features of the two-dimensional cross-section into a low-resolution representation and transmits it to the right. Then, the decoder gradually restores the image resolution. The unique skip connections in U-net combine low-level and high-level features to improve image details and accuracy.

[0041] As an alternative implementation, the training process of the U-net convolutional neural network is as follows:

[0042] Determine the simulated ice shape two-dimensional slice image according to numerical simulation, and determine the experimental ice shape two-dimensional slice image according to the icing wind tunnel test; use the simulated ice shape two-dimensional slice image as a sample and the experimental ice shape two-dimensional slice image as a sample learning label to form a training set and a test set; the positions of the simulated ice shape and the experimental ice shape have a corresponding relationship, and the spanwise coordinates are consistent; input the training set into the U-net convolutional neural network for training, adjust the weights of MSE, SSIM, and edge loss in the preset loss function. During the training process, the MSE weight remains unchanged, and the weights of SSIM and edge loss are gradually increased until the network converges; input the test set into the trained U-net convolutional neural network to obtain the prediction result. Among them, the test set contains ice shape cross-sectional data that is similar but not repeated to the training set; perform error analysis on the prediction result according to the root mean square error formula to evaluate the performance of the trained U-net convolutional neural network. When the preset performance requirements are not met, adjust the division ratio of the training set and the test set, the number of U-net network layers, and the weights of each item in the loss function, and repeat the training process until the preset performance requirements are met.

[0043] Exemplarily, the training set and the test set are obtained through Ansys Fensapice numerical simulation and icing wind tunnel test respectively. The following gives specific examples of obtaining the training set and the test set:

[0044] Example 1: Taking the three-dimensional icing of a NACA0012 30° swept wing as an example, the proposed three-dimensional icing rapid prediction method based on the U-net network was verified. Numerical simulation studies were carried out using Fluent software and FENSAP-ICE software. The FENSAP-ICE software was used to calculate the flow field, water droplet impingement characteristics, and ice layer growth process, and Fluent was used for grid update. The grid was generated using pointwise software, and the generated grid contained approximately 8,394,000 volume grids and 102,000 surface grids for discretizing the wing geometry. All grids used structured boundary layers and satisfied y + <1, which means that the viscous effects within the boundary layer were fully resolved. The simulated ice shapes obtained by the FENSAP-ICE software and the first 100 two-dimensional icing slice profiles of the two-dimensional test section obtained from the icing test were used as the training set, while the subsequent 100 two-dimensional icing slice profiles were regarded as the test set.

[0045] As an alternative implementation, the preset loss function Loss of the U-net convolutional neural network is:

[0046] Loss = w mse ·MSE(y true , y pred ) + w ssim ·(1 - SSIM(y true , y pred )) + w edge ·MSE(Edge(y true ), Edge(y pred ));

[0047] Among them, y true is the two-dimensional slice image of the experimental ice shape; y pred is the image predicted by the U-net convolutional neural network, Edge(y) represents extracting the edge information of image y through the Sobel operator, and w mse , w ssim , w edge are the weight coefficients of MSE, SSIM, and edge loss respectively. In the initial stage of training (for example, the first 50 epochs), the weight of MSE is relatively high (for example, w mse = 1.0, w ssim = 0.1, w edge = 0.1); as the training progresses, the weight of MSE remains unchanged, while the weights of SSIM and edge loss are gradually increased (for example, w ssim = 0.5, w edge = 0.3) to enhance the model's ability to recover the image structure and details.

[0048] The above loss function ensures the accurate recovery of ice shape edges and details, prevents overfitting, and effectively improves the robustness of the model. According to the set loss function, hyperparameters such as the number of network layers, the number of nodes in the bottleneck layer, and the activation function are jointly optimized, and the network is trained using the training set until the network converges. During the training process, the weights of each item in the loss function are gradually adjusted to enhance the network's learning ability for global structures, local details, and complex edge features.

[0049] The trained U-net network is verified using the test set, and the root mean square error (RMSE) is used to evaluate the network performance to ensure the generalization ability of the model on the test set and prevent overfitting. Specifically, the test set is input into the trained U-net network. The test set should have similar ice shape data as the training set but not be repeated to ensure the fairness and generalization ability of model evaluation. The test set cross-sectional data is predicted through the U-net network to obtain the predicted results output by the network. Calculate the root mean square error (RMSE) between the predicted results of the test set and the true values.

[0050]

[0051] Among them, X i represents the pixel value of the predicted image generated by the network, Y i represents the pixel value of the corresponding experimental ice shape image, and n is the total number of all pixels in the test set. The performance of the model is evaluated according to the calculated RMSE value. A lower RMSE value indicates that the predicted results of the model are more accurate and the generalization ability is stronger; a higher RMSE value indicates that the performance of the model on the test set is poor and there may be overfitting.

[0052] As an optional implementation manner, three-dimensional reconstruction is performed on the two-dimensional cross-sectional images of ice shapes at different spanwise positions to obtain the predicted three-dimensional geometric shape of wing icing, including:

[0053] The interpolation method is used to fill the gaps between adjacent slices of the two-dimensional cross-sections of ice shapes at different spanwise positions to obtain ice shape data; the ice shape data is organized in a three-dimensional grid, and the three-dimensional grid is divided into multiple voxels. Each voxel has eight corner points, and each corner point has a numerical value; for any voxel, the relationship between the corner points and the isosurface is determined by comparing the numerical values of the eight corner points with a preset isosurface value; according to the relationship between the corner points and the isosurface, a triangle combination is selected from a preset template to represent the isosurface; according to the triangle combination, the three-dimensional geometric shape of wing icing is constructed.

[0054] Exemplarily, first, obtain the two-dimensional ice shape cross-sectional data output by the U-net network, usually multiple slice data generated by the U-net network. These slices are the contours of the two-dimensional ice shape. Each slice represents the icing morphology of the wing surface at different positions. Use interpolation methods to fill the gaps between adjacent slices of the two-dimensional ice shape cross-sections at different spanwise positions, filling the gaps between adjacent slices. Methods such as linear interpolation or spline interpolation can be used to ensure smooth transitions between each two-dimensional slice and accurately reflect the actual three-dimensional ice shape distribution. The goal of interpolation is to generate uniformly distributed slice data so as to provide richer input for subsequent three-dimensional reconstruction.

[0055] Then, organize the ice shape data in a three-dimensional grid. This grid is divided into multiple voxels (small cubes), and each voxel can be regarded as a pixel point in 3D space. Voxels are the pixels of 3D space, which are quantified and point clouds with fixed sizes. Each unit has a fixed size and discrete coordinates. Each voxel has eight corner points, and each corner point has a numerical value, which represents the physical quantity at that corner point. For any voxel, compare the numerical values of the eight corner points (i.e., the ice shape data obtained by interpolation) with a preset isosurface value to judge its inner and outer surfaces. The core idea is to approximate the isosurface through linear interpolation. According to the relationship between the corner points and the isosurface, select a combination of triangles from a preset template to represent the isosurface. For each voxel, according to the relative position of the vertices and the isosurface, connect the intersection points of the isosurface and the voxel edges in a certain way to generate the isosurface, which is an approximate representation of the isosurface within the voxel. According to the triangle combination, construct the three-dimensional geometric shape of the complete wing icing. Finally, post-processing can also be performed on the three-dimensional reconstructed ice shape data to ensure the smoothness of the ice shape surface, remove possible noise or outliers, and further improve the accuracy and quality of the reconstruction results.

[0056] This embodiment gives the modeling effect diagram of the three-dimensional geometric shape of the wing icing obtained by three-dimensional reconstruction of the two-dimensional cross-sectional image of the ice shape of the NACA0012 30° swept wing three-dimensional icing prediction according to the above method, as Figure 4 shown. It can be seen that the prediction details are rich and the prediction effect is good.

[0057] The embodiment of the present invention provides a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network. Through the method of three-dimensional reconstruction, the state of the ice shape on the wing can be restored, which is beneficial to subsequent flight performance analysis, especially the evaluation of the impact of icing on the aerodynamic performance of the wing.

[0058] As an alternative implementation, a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network further includes:

[0059] Obtain the two-dimensional sliced image of the test ice shape obtained from the simulation ice shape two-dimensional sliced image and the icing wind tunnel test under the second target icing condition;

[0060] Thaw the target network parameters in the first target neural network, and use the two-dimensional sliced image of the test ice shape obtained from the simulation ice shape two-dimensional sliced image and the icing wind tunnel test under the second target icing condition to train the thawed first target neural network, and adjust the thawed target network parameters to obtain the second target neural network.

[0061] Exemplarily, first, obtain the above-mentioned U-net neural network model that has been trained and converged, and this model can already effectively predict the three-dimensional ice shape under the first target icing condition. This network includes two parts: an encoder and a decoder, as well as the network weights and parameters that have been trained. Then, determine the target tasks that need to apply transfer learning. These target tasks include predicting the three-dimensional ice shapes of different airfoils under other icing conditions (the second target icing condition). Therefore, prepare a small amount of two-dimensional sliced images of the test ice shape obtained from the simulation ice shape two-dimensional sliced image and the icing wind tunnel test under the second target icing condition, and use them as the input of a new training set to fine-tune the first target neural network (U-net network).

[0062] In this embodiment, transfer learning is performed on the first target neural network for three-dimensional icing prediction of the NACA0012 30° swept wing to obtain the second target neural network, and the second target neural network is used for three-dimensional icing prediction of the RG-15 airfoil. This embodiment provides Example 2 to determine the training set and the test set. Example 2 is specifically as follows:

[0063] Example 2: The second example for verifying the three-dimensional icing rapid prediction method, the three-dimensional icing of the RG-15 airfoil. Use the Fluent software and the FENSAP-ICE software for numerical simulation research. The FENSAP-ICE software is used to calculate the flow field, the water droplet impingement characteristics, and the ice layer growth process, and Fluent is used for grid update. The grids all use structured boundary layers and satisfy y + <1, which means that the viscous effects within the boundary layer are fully resolved. The first 30 two-dimensional icing sliced profiles of the simulation ice shape obtained by the FENSAP-ICE software and the two-dimensional test slices obtained from the icing test are used as the training set (for the training of the second target neural network, and the number of its training sets is greatly reduced), while the subsequent 100 two-dimensional icing sliced profiles are regarded as the test set.

[0064] In transfer learning, some of the underlying network parameters are unfrozen. The core of transfer learning is to utilize the knowledge in the pre-trained model and apply it to new but related tasks. At this time, only the lower convolutional layers of the U-net network (i.e., the underlying parameters) are unfrozen. These layers usually learn more general features (such as edges, textures, etc.) and are applicable to ice shape prediction under different icing conditions. By unfreezing these underlying parameters, the feature representation of the network in the source task can be retained and utilized, while allowing the network to adapt to the data features of the new task.

[0065] The model is fine-tuned using the second target icing condition and airfoil data. In transfer learning, fine-tuning means further training the unfrozen part of the network parameters using a new dataset to adapt to the new icing conditions and airfoil features. During the training process, other high-level network parameters are kept unchanged, and mainly the weights of the underlying network are trained to ensure that the model can quickly adapt to the features of different icing conditions.

[0066] Based on the input new icing conditions and airfoil data, the network makes predictions. Through the trained model, rapid predictions are made to generate new three-dimensional icing shapes. The prediction results will retain the efficient characteristics of the U-net network and can accurately reconstruct the three-dimensional ice shapes of different icing conditions and airfoils in a relatively short time.

[0067] As Figure 5 shown, it is the three-dimensional icing prediction effect diagram based on the U-net network and transfer learning method for other icing conditions and airfoil conditions. Among them, the dashed lines respectively represent the prediction effects of the U-net neural network model trained with the NACA0012 swept wing mixed ice as the research object on the NACA0012 swept wing mixed ice, RG-15 three-dimensional airfoil mixed ice, and RG-15 three-dimensional airfoil clear ice. The solid lines respectively represent the prediction models obtained by fine-tuning (transfer learning) the U-net neural network model trained with the NACA0012 swept wing mixed ice as the research object. For the new working conditions of the NACA0012 swept wing mixed ice, RG-15 three-dimensional airfoil mixed ice, and RG-15 three-dimensional airfoil clear ice, the prediction effects are in Figure 5 The ordinate represents the root mean square error (RMSE) between the prediction results of the test set and the true values, and the abscissa represents the size of the training samples. By comparing the errors between the network prediction results after transfer learning and the experimental data, it is ensured that transfer learning effectively improves the prediction accuracy under new working conditions.

[0068] The embodiment of the present invention provides a three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network. By adopting a transfer learning strategy, the generalization ability of the model under different icing conditions is further improved, making the three-dimensional ice shape prediction under different icing conditions more efficient and accurate.

[0069] An embodiment of the present invention provides a three-dimensional icing rapid prediction device based on multi-fidelity data fusion and neural network, including:

[0070] A slice acquisition module, configured to acquire a two-dimensional slice image of the simulated ice shape under a first target icing condition; for details, refer to the corresponding part of the above method embodiment and will not be elaborated here.

[0071] A prediction module, configured to input the two-dimensional slice image of the simulated ice shape into a pre-trained first target neural network to predict the two-dimensional slice images of the ice shape at different spanwise positions of the wing under the first target icing condition, where the first target neural network is trained based on the two-dimensional slice image of the simulated ice shape and the two-dimensional slice image of the experimental ice shape obtained from the icing wind tunnel test; for details, refer to the corresponding part of the above method embodiment and will not be elaborated here.

[0072] A construction module, configured to perform three-dimensional reconstruction on the two-dimensional cross-sectional images of the ice shape at different spanwise positions to obtain the predicted three-dimensional geometric shape of the wing icing. For details, refer to the corresponding part of the above method embodiment and will not be elaborated here.

[0073] An embodiment of the present application also provides an electronic device, as Figure 6 shown, a processor 501 and a memory 502, where the processor 501 and the memory 502 can be connected through a bus or other means.

[0074] The processor 501 can be a central processing unit (CPU). The processor 501 can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. chips, or a combination of the above types of chips.

[0075] The memory 502, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network in the embodiment of the present invention. The processor executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions, and modules stored in the memory.

[0076] The memory 502 may include a program storage area and a data storage area. Among them, the program storage area can store an operating system and application programs required for at least one function; the data storage area can store data created by the processor and the like. In addition, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory 502 may optionally include a memory remotely set relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0077] The one or more modules are stored in the memory 502 and, when executed by the processor 501, execute the three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network in the embodiments as Figure 1 shown.

[0078] Specific details of the above electronic device can be correspondingly referred to Figure 1 the relevant descriptions and effects in the corresponding embodiments shown for understanding, and will not be elaborated here.

[0079] This embodiment also provides a computer storage medium. The computer storage medium stores computer-executable instructions, and these computer-executable instructions can execute the three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network in any of the above method embodiments. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), a flash memory, a hard disk drive (abbreviation: HDD), or a solid-state drive (SSD), etc.; the storage medium can also include a combination of the above types of memories.

[0080] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network, characterized in that: include: Acquire a two-dimensional slice image of a simulated ice shape under a first target icing condition; Inputting the simulated ice shape two-dimensional slice image into a pre-trained first target neural network, predicting the ice shape two-dimensional slice image at different spanwise positions of the wing under the first target icing condition, wherein the first target neural network is trained based on the simulated ice shape two-dimensional slice image and the experimental ice shape two-dimensional slice image obtained from the icing wind tunnel test; The two-dimensional cross-sectional images of ice shapes at different spanwise positions are reconstructed in three dimensions to obtain the predicted three-dimensional geometric shape of wing icing.

2. The method for rapid three-dimensional icing prediction based on multi-fidelity data fusion and neural network according to claim 1, characterized in that: The first target neural network is the U-net convolutional neural network, which includes an encoder and a decoder. The encoder includes multiple convolutional blocks and a maximum pooling layer. Each convolutional block includes convolution, batch normalization, ReLU activation and residual. The maximum pooling layer converts the multi-dimensional features of the two-dimensional ice section data at the first resolution into the second resolution. The second resolution is lower than the first resolution, but the deep features of the data are extracted. The decoder gradually restores the image resolution.

3. The method for rapid three-dimensional icing prediction based on multi-fidelity data fusion and neural network according to claim 2, characterized in that: The training process of the U-net convolutional neural network is as follows: Determine a two-dimensional slice image of a simulated ice shape according to numerical simulation, and determine a two-dimensional slice image of a test ice shape according to an icing wind tunnel test; The simulated ice shape 2D slice image is used as a sample, and the experimental ice shape 2D slice image is used as a sample learning label to form a training set and a test set; the positions of the simulated ice shape and the experimental ice shape have a corresponding relationship, and the spanwise coordinates are consistent; The training set is input into the U-net convolutional neural network for training, and the weights of MSE, SSIM and edge loss in the preset loss function are adjusted. During the training process, the MSE weight remains unchanged, and the weights of SSIM and edge loss are gradually increased until the network converges; The test set is input into the trained U-net convolutional neural network to obtain the prediction results, where the test set contains ice cross-section data that is similar to but not repeated in the training set; The prediction results are analyzed according to the root mean square error formula to evaluate the performance of the trained U-net convolutional neural network. If the preset performance requirements are not met, the ratio of the training set and the test set, the number of U-net network layers, and the weights of the loss function are adjusted, and the training process is repeated until the preset performance requirements are met.

4. The method for rapid three-dimensional icing prediction based on multi-fidelity data fusion and neural network according to claim 3 is characterized in that: The preset loss function Loss of the U-net convolutional neural network is: Loss=w mse ·MSE(y true ,y pred )+w ssim ·(1-SSIM(y true ,y pred ))+w edge ·MSE(Edge(y true ),Edge(y pred )); Among them, y true is the two-dimensional slice image of the experimental ice shape; y pred is the image predicted by the U-net convolutional neural network, Edge(y) represents the edge information of image y extracted by the Sobel operator, and w mse 、w ssim 、w edge are the weight coefficients of MSE, SSIM and marginal loss respectively.

5. The method for rapid three-dimensional icing prediction based on multi-fidelity data fusion and neural network according to claim 1, characterized in that: The 2D slice images of ice at different spanwise positions are reconstructed in 3D to obtain the predicted 3D geometric shape of wing icing, including: The interpolation method is used to fill the gaps between adjacent slices of the two-dimensional ice shape sections at different spanwise positions to obtain ice shape data. Organizing ice shape data in a three-dimensional grid, the three-dimensional grid is divided into a plurality of voxels, each voxel has eight corner points, and each corner point has a numerical value; For any voxel, the relationship between the corner points and the isosurface is determined by comparing the values ​​of the eight corner points with the preset isosurface values; According to the relationship between the corner points and the isosurface, a triangle combination is selected from a preset template to represent the isosurface; Based on the combination of triangles, the three-dimensional geometric shape of wing icing is constructed.

6. The method for rapid three-dimensional icing prediction based on multi-fidelity data fusion and neural network according to claim 1, characterized in that: Also includes: Acquire a two-dimensional slice image of the experimental ice shape obtained from a simulated ice shape two-dimensional slice image and an icing wind tunnel test under a second target icing condition; The target network parameters in the first target neural network are thawed, and the thawed first target neural network is trained using the simulated ice shape two-dimensional slice image and the experimental ice shape two-dimensional slice image obtained by the icing wind tunnel test under the second target icing condition, and the thawed target network parameters are adjusted to obtain the second target neural network.

7. A three-dimensional icing rapid prediction device based on multi-fidelity data fusion and neural network, characterized in that: include: A slice acquisition module, used to acquire a two-dimensional slice image of a simulated ice shape under a first target icing condition; A prediction module is used to input the simulated ice shape two-dimensional slice image into a pre-trained first target neural network to predict the ice shape two-dimensional slice image at different spanwise positions of the wing under the first target icing condition, wherein the first target neural network is trained based on the simulated ice shape two-dimensional slice image and the experimental ice shape two-dimensional slice image obtained by the icing wind tunnel test; The construction module is used to perform three-dimensional reconstruction of the two-dimensional cross-sectional images of ice shapes at different spanwise positions to obtain the predicted three-dimensional geometric shape of wing icing.

8. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the steps of the three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network as described in any one of claims 1 to 6.

9. A computer storage medium having computer instructions stored thereon, characterized in that: When the instruction is executed by the processor, the steps of the three-dimensional icing rapid prediction method based on multi-fidelity data fusion and neural network described in any one of claims 1-6 are implemented.

Images