A real-time simulation method for aircraft engines based on convolutional neural networks

Abstract

The present application provides a method for real-time simulation of an aero-engine based on a convolutional neural network, which belongs to the field of aero-engine simulation technology. Specifically, the method includes obtaining original grid data and the distribution of various physical quantities, constructing a training input data set and a training label data set after imaging the simulation data, building a neural network model, and designing a loss function to train the neural network model using the training input data set and the training label data set; imaging the boundary conditions of the real-time simulation data of the engine operation, inputting the imaged real-time data into the trained neural network model, and the trained neural network model outputting an imaged prediction result. By accurately identifying the boundaries of the calculation domain, constructing a mask, and retaining the valid area data, the simulation prediction accuracy and real-time performance based on the convolutional neural network are greatly improved.

Description

Technical Field

[0001] The present application relates to the field of aero-engine simulation, and in particular to a real-time aero-engine simulation method based on convolutional neural networks. Background Art

[0002] During operation, the performance of aircraft engines is affected by a variety of factors, including the operating environment, operating conditions, and design parameters. While traditional numerical simulation methods offer high accuracy, they are often computationally intensive and time-consuming, making it difficult to achieve real-time performance. In recent years, deep learning methods, particularly convolutional neural networks, have significantly improved computational efficiency in image processing and feature extraction. Visualizing numerical simulation results and using convolutional neural networks for rapid predictions are key areas that require urgent breakthroughs. However, existing techniques for visualizing simulation data suffer from significant interpolation errors and invalid data interference, impacting network training convergence and prediction accuracy. Summary of the Invention

[0003] In view of this, the present application provides a real-time simulation method for aircraft engines based on convolutional neural networks, which solves the problems in the existing technology, realizes fast and accurate prediction of the aircraft engine status, and meets the needs of real-time simulation.

[0004] The present application provides a convolutional neural network-based real-time simulation method for aircraft engines, which employs the following technical solutions:

[0005] A convolutional neural network-based real-time simulation method for an aerospace engine comprises the following steps:

[0006] Perform simulation calculations on a variety of simulation entities to obtain original grid data and distribution of various physical quantities;

[0007] Each boundary condition is processed into an image to obtain input images of multiple channels as a training input data set;

[0008] The physical quantities to be predicted are processed into images to form multi-channel label data as the training label data set, where the number of channels is consistent with the number of physical quantities to be predicted;

[0009] Normalize the image data to the [0,1] interval and save it;

[0010] Build a neural network model;

[0011] Design loss function;

[0012] Train the neural network model using the training input dataset and the training label dataset;

[0013] The boundary conditions of the real-time simulation data of the engine operation are processed graphically, and the graphical real-time data is input into the trained neural network model, and the trained neural network model outputs the graphical prediction results;

[0014] The specific steps of image processing include:

[0015] Select a predefined fixed resolution as the target image size and calculate the pixel value by interpolation to obtain the interpolated result ;

[0016] Construct the masked area and obtain the mask matrix ;

[0017] Multiply the interpolation result by the mask matrix pixel by pixel to get the final pixel value .

[0018] Optionally, CFD numerical simulation software is used to perform simulation calculations on the simulation entity to obtain original grid data and distribution of various physical quantities.

[0019] Optionally, interpolate the pixel values ​​to obtain the interpolated result The specific steps include:

[0020] Map the discrete grid point coordinates (x, y) and physical quantity values ​​output by the CFD numerical simulation software to the image coordinate system, where each grid point corresponds to a number of pixel positions;

[0021] Calculate the weight coefficients of the four nearest grid points in the CFD grid space for the center coordinates of each pixel in the image, and use bilinear or cubic interpolation methods to calculate the physical value at each pixel.

[0022] Optionally, construct a mask matrix The specific steps include:

[0023] From the CFD mesh output, all nodes marked as contours or flow field edges are filtered to construct a preliminary boundary node set. Using the node-cell topology relationship, the candidate boundary nodes are traversed adjacently. The boundary points are arranged into a closed sequence in the order of connection to fully describe the boundary contours inside and outside the computational domain.

[0024] Map physical coordinates (X, Y) to image pixel coordinates ( , ), the formula is:

[0025] ;

[0026] ;

[0027] in, is the number of horizontal pixels of the target image, is the number of vertical pixels of the target image;

[0028] If multiple physical nodes fall on the same pixel, the sub-pixel average or the closest center point is taken;

[0029] If the adjacent mapped pixels ( , )and( , ) If the distance is greater than 1, the Bresenham straight line algorithm is used to interpolate the middle pixels, and the discrete pixel sequence is fitted with cubic spline or linear interpolation, and resampled in a way with a step size of less than 1 pixel to eliminate aliasing and obtain a smooth closed contour;

[0030] Project the boundary points in the CFD grid model and project all the grid points belonging to the boundary into the image coordinate system to obtain one or more closed curves. In the image coordinate system, use the scan line algorithm or polygon filling algorithm to fill the internal area of ​​the closed curve and mark it as the inside of the calculation domain. The external area is marked as an invalid area to obtain the mask matrix. .

[0031] Optionally, get the mask matrix In the process, the sub-pixel boundary interpolation method is used to remove edge artifacts by weight scaling for pixels adjacent to the boundary and with a mask value of 0 but a distance calculation domain smaller than the pixel scale;

[0032] Gaussian blur is performed on pixels at the mask boundary to smooth the mask edge.

[0033] Optionally, when designing the loss function, the error is calculated by weighting pixels only within the computational domain according to the following formula: :

[0034] ;

[0035] in, is the network prediction value, is the true label value, is the binary mask generated during the masking process, A small constant to prevent the denominator from being zero.

[0036] In summary, this application has the following beneficial technical effects:

[0037] This application focuses on the visualization of simulation data. By accurately identifying the boundaries of the calculation domain, constructing masking masks, and retaining valid area data, the generated image can more realistically reflect the internal flow field distribution of the simulation, thereby greatly improving the simulation prediction accuracy and real-time performance based on convolutional neural networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0039] Figure 1 The flowchart of the real-time simulation method of aircraft engines based on convolutional neural networks. DETAILED DESCRIPTION

[0040] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0041] The following describes the embodiments of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the features in the following embodiments and embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative work are within the scope of protection of this application.

[0042] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this application, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0043] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. The illustrations only show components related to the present application and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0044] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described may be practiced without these specific details.

[0045] An embodiment of the present application provides a real-time simulation method for an aerospace engine based on a convolutional neural network.

[0046] like Figure 1 As shown, a real-time simulation method for an aircraft engine based on a convolutional neural network includes the following steps:

[0047] Perform simulation calculations on a variety of different simulation entities to obtain original grid data and the distribution of various physical quantities.

[0048] Each boundary condition is processed into an image to obtain input images of multiple channels as a training input data set.

[0049] The physical quantities to be predicted are processed into images to form multi-channel label data as a training label data set, and the number of channels is consistent with the number of physical quantities to be predicted.

[0050] Normalize the image data to the [0,1] interval and save it.

[0051] Build a neural network model.

[0052] Design a loss function.

[0053] The neural network model is trained using the training input dataset and the training label dataset.

[0054] The boundary conditions of the real-time simulation of engine operation are processed in a graphical manner, and the graphical real-time data is input into the trained neural network model, which then outputs the graphical prediction results.

[0055] This application provides a specific process of a real-time simulation method for an aircraft engine based on a convolutional neural network:

[0056] Step 1: Use CFD numerical simulation software to simulate the simulation entity and obtain the original mesh data and the distribution of various physical quantities. Taking the compressor blade as an example, the original mesh data and the distribution of various physical quantities obtained include pressure, temperature, Mach number, etc. The simulation should cover multi-dimensional working conditions such as different blade geometric parameters, flow rate, speed, etc. to ensure the diversity and integrity of the database samples. Blade geometric parameters include curvature, thickness, etc. A total of 500 sets of sample data are generated, and each set of samples contains different inlet and outlet boundary conditions and different blade geometric parameters. Each simulation outputs a two-dimensional irregular grid and the corresponding physical quantity scalar field, such as Mach number M, static pressure p, temperature T, etc.

[0057] Step 2: Process each boundary condition into images to obtain multiple channel input images as training input data sets. Boundary conditions such as inlet pressure P in , inlet temperature T in , outlet pressure P out .

[0058] Step 3: The physical quantities to be predicted are processed into images to form multi-channel labeled data as a training label dataset. The number of channels matches the number of physical quantities to be predicted. The predicted physical quantities include the flow field Mach number distribution, static pressure distribution, and temperature distribution.

[0059] The specific methods of image processing in steps 2 and 3 are as follows:

[0060] Step 21 : Select a predefined fixed resolution as the target image size; in one embodiment, the resolution of the target image is 256×256 pixels.

[0061] Step 22: Define the origin of the image coordinate system to correspond to the lower left corner of the physical coordinate system, and linearly map the vertical and horizontal directions to the maximum range of the grid. Map the discrete grid point coordinates (x, y) and physical values ​​output by the CFD numerical simulation software to the image coordinate system. Each grid point corresponds to a number of pixel positions, and the pixel values ​​are calculated by interpolation to obtain the difference result. .

[0062] The specific method of calculating pixel values ​​by interpolation is:

[0063] Calculate the weight coefficients of the four nearest grid points in the CFD grid space for the center coordinates of each pixel in the image, and use bilinear or cubic bicubic interpolation methods to calculate the physical value at the pixel The formula for the interpolation algorithm is ;in, is the physical quantity of the adjacent i-th grid point, and the weight Calculated based on the distance between the pixel center and the grid point. To avoid artifacts caused by numerical oscillations, the CFD data is pre-processed by Gaussian filtering or curve fitting before interpolation.

[0064] For each CFD grid node, its real physical coordinates are first mapped to the image coordinate system to obtain the floating point pixel coordinates ( , ); corresponding to its physical value .

[0065] For each pixel center , find the nearest 4 nodes in the CFD grid data, interpolate according to the bilinear interpolation formula, and get the pixel center The physical value of :

[0066] ;

[0067] in, 、 To surround The upper and lower boundary coordinates of 、 、 、 is the grid node value of each corresponding node.

[0068] Step 23: From the CFD mesh output, filter all nodes marked as contours or flow field edges, construct a preliminary boundary node set, and use the node-cell topology relationship to perform adjacency traversal on the candidate boundary nodes. Arrange the boundary points into a closed sequence in the order of connection to fully describe the boundary contours inside and outside the computational domain.

[0069] Map physical coordinates (X, Y) to image pixel coordinates ( , ), the formula is:

[0070] ;

[0071] ;

[0072] in, is the number of horizontal pixels of the target image, is the number of vertical pixels of the target image;

[0073] If multiple physical nodes fall on the same pixel, take the sub-pixel average or the closest center point to improve accuracy.

[0074] If the adjacent mapped pixels ( , )and( , ) distance is greater than 1, the Bresenham algorithm is used to interpolate intermediate pixels to ensure a continuous closed curve at the pixel level. The discrete pixel sequence is resampled using cubic spline or linear interpolation with a step size of less than 1 pixel to eliminate aliasing and obtain a smooth closed contour, facilitating subsequent mask generation.

[0075] Step 24: Project the boundary points in the CFD mesh model. Project all mesh points belonging to the boundary into the image coordinate system to obtain one or more closed curves. In the image coordinate system, use the scanline algorithm Scanline or polygon filling algorithm such as fillPoly in OpenCV to fill the internal area of ​​the closed curve and mark it as the inside of the calculation domain and define the mask value as 1. Mark the external area as an invalid area and define the mask value as 0 to obtain the mask matrix. , multiply the interpolation result by the mask matrix pixel by pixel to get the final pixel value , suppressing spurious interpolation data outside the computational domain:

[0076] .

[0077] Get the mask matrix During the process, sub-pixel boundary interpolation is used to remove edge artifacts by weight scaling for pixels adjacent to the boundary whose mask value is 0 but whose distance calculation domain is smaller than the pixel scale. Specifically, sub-pixel processing is performed on mask boundary pixels: if the distance between the center of a pixel and the polygon boundary is less than half a pixel, the mask value of that pixel is set to α∈(0,1) and linearly decayed. For example, pixels with a distance of 0 have a mask value of 1, pixels with a distance of 0.5 have a mask value of 0.5, and pixels with a distance of ≥1 have a mask value of 0.

[0078] Gaussian blur processing is performed on the pixels at the mask boundary to smooth the mask edge so that the subsequent convolutional neural network can automatically learn the boundary transition, wherein the Gaussian blur processing uses a Gaussian kernel with a radius of 1 to 2 pixels, for example.

[0079] Step 4: normalize the image data to the interval [0, 1] and save it in an efficient storage format, such as a NumPy array.

[0080] Step 5: Build the neural network model: A convolutional neural network based on the U-Net architecture is used, with the input size consistent with the image size. The encoder consists of N convolutional blocks, each of which includes two 3×3 convolutional layers, batch normalization (BatchNorm) and Reluctant Unit (ReLU) activations, and a 2×2 max pooling layer for downsampling, doubling the number of channels with each downsampling. The decoder has a symmetrical deconvolutional ConvTranspose structure similar to the encoder. Each layer is fused with the corresponding encoder output using skip connections to fully utilize multi-scale features. Finally, the output layer uses a 1×1 convolution to reduce the number of channels to the number of channels of the predicted physical quantity, followed by a sigmoid or linear activation. ReLU, short for Rectified Linear Unit, is a commonly used activation function in artificial neural networks.

[0081] Step 6, design loss function: This application proposes an improved multi-channel relative error loss function, which calculates the error L by pixel weighting only within the calculation domain, thereby avoiding the influence of invalid pixels in the masked area on network training:

[0082]

[0083] in, is the network prediction value, is the true label value, The binary mask generated in step 3, A small constant to prevent the denominator from being zero.

[0084] In step 7, the neural network model is trained using the training input dataset and the training label dataset. Specifically, the dataset is split into training, validation, and test sets with a ratio of 70%:20%:10%. AdamW is used as the optimizer, with an initial learning rate of 0.001 and 100 training iterations. The loss is evaluated on the validation set every five epochs, and the learning rate or early stopping is dynamically adjusted. The evaluation metrics used are the mean relative error and peak error within the computational domain. If necessary, the structural similarity metric (SSIM) is used to measure the spatial consistency of the graphical simulation results.

[0085] Step 8: The actual operating parameters of the aircraft engine simulation data P in 、T in 、P out Following the same process as above, visualization and masking are performed, inputting the trained model to quickly output a graphical prediction of each physical quantity within the computational domain. Finally, the graphical prediction results can be mapped back to the physical coordinate system through inverse interpolation to obtain a point-by-point distribution of physical quantities, enabling engineers to perform visual analysis or provide real-time decision support. Specifically, the predicted image is denormalized to restore the actual range of the physical quantity. To display it on a CAD / CFD grid, an inverse interpolation algorithm is used: the image pixels are mapped to the CFD grid, and the image prediction values ​​are converted back to each grid point through bilinear or distance-weighted interpolation. Finally, a visualization interface displays the engine's internal flow field cloud map, contour line distribution, and more.

[0086] This application systematically designs the interpolation algorithm and image resolution selection, so that the simulation results can be seamlessly converted into visual information, which is convenient for direct processing by the convolutional network and significantly improves the network convergence speed and prediction accuracy. Through mask generation and sub-pixel boundary processing based on boundary projection, the interpolation noise outside the calculation domain is effectively removed, ensuring that the network training only performs error calculations within the effective area, further improving the simulation accuracy. The method of this application is not only suitable for real-time simulation of two-dimensional blades, but can also be extended to complex flow fields such as three-dimensional blades and entire engines. It is only necessary to adjust the boundary recognition algorithm in the mask generation step accordingly. With the help of image data and lightweight U-Net network structure, millisecond-level predictions can be achieved on conventional GPUs or edge computing devices to meet the real-time monitoring and fault warning needs of aircraft engines. The generated image prediction results can be directly used in visualization interfaces such as thermodynamic cloud maps and flow field distribution maps without the need for additional post-processing, so that engineers can understand the working status of the engine at a glance.

[0087] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A real-time simulation method for an aircraft engine based on a convolutional neural network, characterized in that: The steps include: Perform simulation calculations on a variety of different simulation entities to obtain original grid data and distribution of various physical quantities; Each boundary condition is processed into an image to obtain input images of multiple channels as a training input data set; The physical quantities to be predicted are processed into images to form multi-channel label data as the training label data set, where the number of channels is consistent with the number of physical quantities to be predicted; Normalize the image data to the [0,1] interval and save it; Build a neural network model; Design loss function; Train the neural network model using the training input dataset and the training label dataset; The boundary conditions of the real-time simulation data of the engine operation are processed graphically, and the graphical real-time data is input into the trained neural network model, and the trained neural network model outputs the graphical prediction results; The specific steps of image processing include: Select a predefined fixed resolution as the target image size and calculate the pixel value by interpolation to obtain the interpolated result ; Construct the masked area and obtain the mask matrix ; Multiply the interpolation result by the mask matrix pixel by pixel to get the final pixel value ; Use CFD numerical simulation software to simulate the simulation entity and obtain the original grid data and the distribution of various physical quantities; The interpolated result is obtained by interpolating the pixel value The specific steps include: Map the discrete grid point coordinates (x, y) and physical quantity values ​​output by the CFD numerical simulation software to the image coordinate system, where each grid point corresponds to a number of pixel positions; Calculate the weight coefficients of the four nearest grid points in the CFD grid space for the center coordinates of each pixel in the image, and use bilinear or cubic interpolation methods to calculate the physical value at each pixel; Constructing the mask matrix The specific steps include: From the CFD mesh output, all nodes marked as contours or flow field edges are filtered to construct a preliminary boundary node set. Using the node-cell topology relationship, the candidate boundary nodes are traversed adjacently. The boundary points are arranged into a closed sequence in the order of connection to fully describe the boundary contours inside and outside the computational domain. Map physical coordinates (X, Y) to image pixel coordinates ( , ), the formula is: ； ； in, is the number of horizontal pixels of the target image, is the number of vertical pixels of the target image; If multiple physical nodes fall on the same pixel, the sub-pixel average or the closest center point is taken; If the adjacent mapped pixels ( , )and( , ) If the distance is greater than 1, the Bresenham straight line algorithm is used to interpolate the middle pixels, and the discrete pixel sequence is fitted with cubic spline or linear interpolation, and resampled in a way with a step size of less than 1 pixel to eliminate aliasing and obtain a smooth closed contour; Project the boundary points in the CFD grid model and project all the grid points belonging to the boundary into the image coordinate system to obtain one or more closed curves. In the image coordinate system, use the scan line algorithm or polygon filling algorithm to fill the internal area of ​​the closed curve and mark it as the inside of the calculation domain. The external area is marked as an invalid area to obtain the mask matrix. .

2. The method for real-time simulation of an aircraft engine based on a convolutional neural network according to claim 1, characterized in that: Get the mask matrix In the process, the sub-pixel boundary interpolation method is used to remove edge artifacts by weight scaling for pixels adjacent to the boundary and with a mask value of 0 but a distance calculation domain smaller than the pixel scale; Gaussian blur is performed on pixels at the mask boundary to smooth the mask edge.

3. The method for real-time simulation of an aircraft engine based on a convolutional neural network according to claim 1, characterized in that: When designing the loss function, the error is calculated by weighting pixels only within the calculation domain according to the following formula: : ； in, is the network prediction value, is the true label value, is the binary mask generated during the masking process, A small constant to prevent the denominator from being zero.

Images