Electromagnetic simulation multi-scale region segmentation method based on deep learning

Abstract

The invention discloses an electromagnetic simulation multi-scale region segmentation method based on deep learning, and the method comprises the steps: constructing and training a deep learning model which is used for recognizing a key region and a non-key region in an electromagnetic simulation region; in the training process, a data set is labeled, and the model performance is optimized by using a cross validation method; simulating the key area by adopting a full-wave solution method to obtain field distribution of the key area; simulating the non-critical area by adopting a high-frequency approximation method, and simulating an electromagnetic field generated after an incident wave is reflected by the surface of an object; dynamic boundary conditions are generated at the boundary of the key area and the non-key area, and continuity and consistency of field values are ensured. According to the method, in a large-scale electromagnetic simulation scene, through combination of deep learning and a mixed solution strategy, optimal configuration of computing resources is achieved, limited computing force is focused on a high-precision demand area, and simulation efficiency and overall performance are remarkably improved.

Description

Technical Field

[0001] The patent of this invention relates to the field of electromagnetic simulation, and in particular to a multi-scale region segmentation method for electromagnetic simulation based on deep learning, which is used to optimize the computational efficiency and accuracy in the electromagnetic simulation process. Background Art

[0002] Large-scale electromagnetic simulation involves complex electromagnetic wave propagation, scattering and radiation problems, and is widely used in antenna design, radar cross section (RCS) analysis, electromagnetic compatibility (EMC) testing, wireless communication coverage and other fields. As the scale of simulation increases, traditional small-scale simulation methods have encountered bottlenecks in computing time and resource requirements, so various advanced technologies have emerged: full-wave solution method, high-frequency approximation method, hybrid method, parallel computing and high-performance computing. Although the existing electromagnetic simulation technology has made significant progress, there are still several major problems when dealing with large-scale electromagnetic simulation:

[0003] 1) Excessive computational complexity and resource requirements: In large-scale problems, the computational complexity of full-wave solutions increases dramatically with frequency, geometry size, and number of grids. Especially in high-frequency, large-scale simulations, the sparse matrices generated by full-wave methods are very large, and solving these matrices requires a lot of memory and long computation time.

[0004] 2) Trade-off between accuracy and efficiency: Although high-frequency approximation methods can improve efficiency in large-scale, far-field problems, they are not accurate enough in complex geometries or near-field regions because they ignore some details in the electromagnetic field and cannot capture key electromagnetic effects. Although hybrid methods can balance accuracy and efficiency, they have high requirements for region division and boundary condition processing. Manually setting these region divisions and boundary conditions is not only time-consuming, but also prone to introduce errors.

[0005] 3) Complex region division and boundary condition processing: In the hybrid method, different numerical methods are required for different regions, and the manual division of regions of interest and regions of no interest may not be precise enough, resulting in inaccurate simulation results or waste of computing resources.

[0006] 4) Insufficient simulation automation and intelligence: In the current simulation process, many steps (such as area division, grid generation, boundary condition setting, etc.) require manual intervention and the degree of automation is low. In large-scale simulation processes, these manual operations are time-consuming and error-prone. Summary of the invention

[0007] Purpose of the invention: In view of the shortcomings of the above-mentioned technology, the purpose of the present invention is to provide a multi-scale region segmentation method for electromagnetic simulation based on deep learning, so as to solve the problems in the prior art of high computational complexity of large-scale electromagnetic simulation, difficulty in balancing accuracy and efficiency, complex region division and boundary condition processing, and insufficient simulation automation and intelligence.

[0008] Technical solution: The electromagnetic simulation multi-scale region segmentation method based on deep learning described in the present invention specifically includes:

[0009] (1) Build and train a deep learning model to identify critical and non-critical areas in electromagnetic simulation areas. The training process uses an annotated dataset and a cross-validation method to optimize model performance.

[0010] (2) The full-wave solution method is used to simulate the key area to obtain the field distribution in the key area;

[0011] (3) The high-frequency approximation method is used to simulate the non-critical area to simulate the electromagnetic field generated by the incident wave reflected by the surface of the object;

[0012] (4) Generate dynamic boundary conditions at the boundaries between critical areas and non-critical areas to ensure the continuity and consistency of field values.

[0013] Furthermore, the implementation process of identifying the key area and the non-key area in the electromagnetic simulation area in step (1) is as follows:

[0014] The deep learning model is a convolutional neural network, which extracts features from the image data of the top view of the simulation area and the distribution data of the electromagnetic field. The feature extraction layer includes an edge detection module and a gradient analysis module. The edge detection module extracts the contour information of the area, and its convolution kernel is initialized as the Sobel operator:

[0015]

[0016] Calculate the gradient magnitude after convolution:

[0017]

[0018] The gradient analysis module is used to capture the changing pattern of electromagnetic field intensity by extracting gradient features through the convolution kernel of Gaussian Laplacian operator:

[0019]

[0020] Downsample the feature map in the pooling layer to reduce the data dimension:

[0021] f pool (x)=max{x1,x2,...,x n}

[0022] The extracted features are input into the fully connected network in the fully connected layer, and the probabilities of key areas and non-key areas are output in the classification layer, thereby realizing the boundary segmentation of key areas and non-key areas.

[0023] Furthermore, the training process in step (1) uses a labeled data set and a cross-validation method to optimize the model performance. The implementation process is as follows:

[0024] The cross entropy loss function is used to measure the gap between the predicted value and the true label:

[0025]

[0026] Among them, y i is the actual label, is the model output; divide the data into k subsets, use k-1 of them for training and 1 for validation each time; repeat the training cycle k times and calculate the average validation performance.

[0027] Furthermore, the simulation of the key area using the full-wave solution method in step (2) is implemented using the finite element method or the finite difference time domain method.

[0028] Furthermore, the implementation process of step (2) is as follows:

[0029] The finite element method is used to further divide the key area into a finite number of small sub-areas. In each small area, the solution of the global equation is approximated by solving the local equation; the Maxwell equation is transformed into a weak form, namely:

[0030]

[0031] The entire calculation area is divided into small finite elements by meshing. Assuming that the electric and magnetic fields are polynomial functions on each element, the field distribution of the entire area is finally obtained by solving the problem of each small element one by one and assembling them into global equations.

[0032] Furthermore, in step (3), the high-frequency approximation method is used to simulate the non-critical area and the electromagnetic field solution is achieved using physical optics or ray tracing.

[0033] Furthermore, the implementation process of step (3) is as follows:

[0034] For electromagnetic waves on a surface, it is assumed that the propagation directions of the reflected and incident waves are only related to the surface normal and the angle of incidence; according to the basic formula of physical optics:

[0035]

[0036] Calculations are performed on non-critical areas to simulate the electromagnetic field generated when the incident wave is reflected by the surface of the object.

[0037] Furthermore, the implementation process of step (4) is as follows:

[0038] According to the key area and the non-key area, the field value of the boundary area is weighted averaged using weighted polynomial interpolation; assuming that at the boundary, the field value of the key area is f1(x), and the field value of the non-key area is f2(x), the field value f(x) at the boundary is represented by a weighted polynomial, which is in the form of:

[0039] f(x)=w(x)f1(x)+(1-w(x))f2(x)

[0040] Among them, w(x) is the weight function, and the weight function is designed using quadratic interpolation:

[0041]

[0042] Among them, x1 and x2 are the positions on both sides of the boundary, and the weight function will achieve continuous transition at the boundary as the position changes.

[0043] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: 1. Intelligent region division: With the help of deep learning technology, the transition from manual division to automatic generation is realized, and a reasonable and multi-scale region segmentation scheme can be automatically generated according to the actual distribution characteristics of the electromagnetic field; for the identified fine calculation area, a high-precision full-wave solution method is used for detailed simulation; and for the simple calculation area, an efficient high-frequency approximation method is applied to reduce the calculation complexity; 2. Accurate boundary condition generation: Utilizing the predictive ability of deep learning, the changing law of field values ​​between regions is accurately captured, and then accurate and reasonable boundary conditions are automatically generated, which effectively reduces the error in the boundary processing process and ensures seamless connection between different calculation areas; 3. Efficient computing resource allocation: In large-scale electromagnetic simulation scenarios, through the combination of deep learning and hybrid solution strategies, the optimal configuration of computing resources is realized, and the limited computing power is focused on the high-precision requirement area, which significantly improves the simulation efficiency and overall performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the present invention;

[0045] Figure 2 Schematic diagram of electromagnetic simulation area segmentation. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0047] like Figure 1 As shown, the present invention proposes a multi-scale region segmentation method for electromagnetic simulation based on deep learning, and the specific implementation process is as follows:

[0048] Step 1: Build and train a deep learning model to identify critical and non-critical areas in the electromagnetic simulation area. The training process uses a labeled dataset and a cross-validation method to optimize model performance.

[0049] The deep learning model is a convolutional neural network, which extracts features from the image data of the top view of the simulation area and the distribution data of the electromagnetic field. The feature extraction layer includes an edge detection module and a gradient analysis module. The edge detection module is used to extract the contour information of the area, and its convolution kernel is initialized as the Sobel operator:

[0050]

[0051] Calculate the gradient magnitude after convolution:

[0052]

[0053] The gradient analysis module is used to capture the changing pattern of electromagnetic field intensity by extracting gradient features through the convolution kernel of Gaussian Laplacian operator:

[0054]

[0055] Downsample the feature map in the pooling layer to reduce the data dimension:

[0056] f pool (x)=max{x1,x2,...,x n}

[0057] The extracted features are input into the fully connected network at the fully connected layer, and the probabilities of the key area (1) and the non-key area (0) are output at the classification layer. In the process of training and optimizing the model, the cross entropy loss function is used to measure the gap between the predicted value and the true label:

[0058]

[0059] Among them, y i is the actual label, is the model output. Divide the data into k subsets, use k-1 of them for training and 1 for validation each time. Repeat the training cycle k times and calculate the average validation performance.

[0060] Step 2: Simulate the critical areas using full-wave solutions, including the finite element method (FEM) and finite-difference time-domain method (FDTD).

[0061] The full-wave solution method can fully consider the propagation, scattering and interaction of electromagnetic waves in space, and has high accuracy. The key area divided by the finite element method step 1 is further divided into a finite number of small sub-areas, and the solution of the global equation is approximated by solving the local equation in each small area. Convert the Maxwell equation into a weak form, that is:

[0062]

[0063] By using meshing, the entire calculation area is divided into small finite elements. Assuming that the electric field and magnetic field are polynomial functions in each element, the problem of each small element is solved one by one and assembled into a global equation, and finally the field distribution of the entire area is obtained.

[0064] Step 3: Simulate non-critical areas using high-frequency approximations, including physical optics (PO) and ray tracing (RT).

[0065] In the high-frequency limit, the reflection and refraction of electromagnetic fields can be described by simple geometric optics laws. For electromagnetic waves on a surface, it is assumed that the propagation directions of the reflected and incident waves are only related to the surface normal and the angle of incidence. According to the basic formula of physical optics:

[0066]

[0067] Calculation is performed based on the non-critical area divided in step 1 to simulate the electromagnetic field generated after the incident wave is reflected by the surface of the object.

[0068] Step 4: Generate dynamic boundary conditions at the boundaries between critical areas and non-critical areas to ensure the continuity and consistency of field values.

[0069] According to the boundary area segmented by the convolutional neural network in step 1, the field value of the boundary area is weighted averaged using weighted polynomial interpolation. Assuming that at the boundary, the field value of the key area is f1(x), and the field value of the non-key area is f2(x), the field value f(x) at the boundary is represented by a weighted polynomial, which is in the form of:

[0070] f(x)=w(x)f1(x)+(1-w(x))f2(x)

[0071] Among them, w(x) is the weight function, and the weight function is designed using quadratic interpolation:

[0072]

[0073] Among them, x1 and x2 are the positions on both sides of the boundary, and the weight function will achieve continuous transition at the boundary as the position changes.

[0074] In this embodiment, the electromagnetic simulation area is determined. For example, in the simulation of the impact of a wind farm on a radar, it is necessary to determine the geometry of the wind turbine modeling in the wind farm, the material properties, and the operating frequency of the radar. After the electromagnetic simulation area is determined, the simulation area is manually labeled to mark the key area and the non-key area. Figure 2 As shown in the figure, there are many wind turbines distributed in the simulation area. The areas with dense distribution of wind turbines are the key areas in the simulation, and the areas with sparse distribution of wind turbines or no wind turbines are the non-key areas. The full-wave method is used to simulate the electromagnetic field in the key areas, and the high-frequency approximation method is used to simplify the simulation in the non-key areas, so as to generate high-quality training data for this area and provide input and labels for the deep learning model.

[0075] The UNet segmentation model in the convolutional neural network is selected to process the spatial data of the electromagnetic field distribution. The input data is the electromagnetic field distribution data and the image data of the simulation area. The output data is the regional partition mask (1 represents the full-wave solution area, and 0 represents the high-frequency approximation area). The cross entropy loss is used to measure the gap between the predicted value and the true label, and the model parameters are optimized based on this value.

[0076] In the full-wave solution area, the finite element method (FEM) and the finite-difference time domain method (FDTD) are used to solve the electromagnetic field with high precision, and in the high-frequency approximation area, physical optics (PO) and ray tracing (RT) are used to quickly solve the electromagnetic field. Field value coupling is performed at the junction of the full-wave solution area and the high-frequency approximation area to ensure that the field solutions in different solution areas can be seamlessly connected to ensure physical accuracy.

[0077] Compare the field value distributions in the full-wave solution area and the high-frequency approximation area, check the continuity of the field values ​​at the junction, and ensure that the boundary conditions are correctly applied. Compare the simulation results with experimental data or theoretical solutions (such as analytical solutions, approximate solutions) to ensure simulation accuracy. Identify the source of simulation errors, analyze boundary conditions, grid density, and time step issues, and optimize and adjust the errors. If the boundary conditions or regional divisions are not accurate enough, readjust the regional divisions or boundary condition settings through the feedback mechanism of the deep learning model to further improve the accuracy of the simulation results.

[0078] The above content has been described in detail the method of the present invention, but it is obvious that the specific implementation of the present invention is not limited thereto. For those skilled in the art, various obvious changes made to it without departing from the spirit of the present invention and the scope of the claims are within the protection scope of the present invention.

Claims

1. A multi-scale region segmentation method for electromagnetic simulation based on deep learning, characterized in that: The following steps are involved: (1) Build and train a deep learning model to identify critical and non-critical areas in the electromagnetic simulation area; The training process uses a labeled dataset and a cross-validation method to optimize model performance; (2) The full-wave solution method is used to simulate the key area to obtain the field distribution in the key area; (3) The high-frequency approximation method is used to simulate the non-critical area to simulate the electromagnetic field generated by the incident wave reflected by the surface of the object; (4) Generate dynamic boundary conditions at the boundaries between critical areas and non-critical areas to ensure the continuity and consistency of field values.

2. According to the deep learning-based electromagnetic simulation multi-scale region segmentation method of claim 1, it is characterized in that: The implementation process of identifying the key area and the non-key area in the electromagnetic simulation area in step (1) is as follows: The deep learning model is a convolutional neural network, which extracts features from the image data of the top view of the simulation area and the distribution data of the electromagnetic field; The feature extraction layer includes an edge detection module and a gradient analysis module; the edge detection module extracts the contour information of the region, and its convolution kernel is initialized as the Sobel operator: Calculate the gradient magnitude after convolution: The gradient analysis module is used to capture the changing pattern of electromagnetic field intensity by extracting gradient features through the convolution kernel of Gaussian Laplacian operator: Downsample the feature map in the pooling layer to reduce the data dimension: f pool (x)=max{x1,x2,...,x n } The extracted features are input into the fully connected network in the fully connected layer, and the probabilities of key areas and non-key areas are output in the classification layer, thereby realizing the boundary segmentation of key areas and non-key areas.

3. The method for multi-scale region segmentation of electromagnetic simulation based on deep learning according to claim 1, characterized in that: The training process in step (1) uses a labeled data set and a cross-validation method to optimize the model performance. The implementation process is as follows: The cross entropy loss function is used to measure the gap between the predicted value and the true label: Among them, y i is the actual label, is the model output; divide the data into k subsets, use k-1 of them for training and 1 for validation each time; repeat the training cycle k times and calculate the average validation performance.

4. The method for multi-scale region segmentation of electromagnetic simulation based on deep learning according to claim 1, characterized in that: The simulation of the key area using the full-wave solution method in step (2) is implemented using the finite element method or the time-domain finite difference method.

5. The method for multi-scale region segmentation of electromagnetic simulation based on deep learning according to claim 1, characterized in that: The implementation process of step (2) is as follows: The finite element method is used to further divide the key area into a finite number of small sub-areas. In each small area, the solution of the global equation is approximated by solving the local equation; the Maxwell equation is transformed into a weak form, namely: The entire calculation area is divided into small finite elements by meshing. Assuming that the electric and magnetic fields are polynomial functions on each element, the field distribution of the entire area is finally obtained by solving the problem of each small element one by one and assembling them into global equations.

6. The method for multi-scale region segmentation of electromagnetic simulation based on deep learning according to claim 1, characterized in that: In step (3), the high-frequency approximation method is used to simulate the non-critical areas and the electromagnetic field is solved using physical optics or ray tracing.

7. The method for multi-scale region segmentation of electromagnetic simulation based on deep learning according to claim 1, characterized in that: The implementation process of step (3) is as follows: For electromagnetic waves on a surface, it is assumed that the propagation directions of the reflected and incident waves are only related to the surface normal and the angle of incidence; according to the basic formula of physical optics: Calculations are performed on non-critical areas to simulate the electromagnetic field generated when the incident wave is reflected by the surface of the object.

8. The method for multi-scale region segmentation of electromagnetic simulation based on deep learning according to claim 1, characterized in that: The implementation process of step (4) is as follows: According to the key area and the non-key area, the field value of the boundary area is weighted averaged using weighted polynomial interpolation; assuming that at the boundary, the field value of the key area is f1(x), and the field value of the non-key area is f2(x), the field value f(x) at the boundary is represented by a weighted polynomial, which is in the form of: f(x)=w(x)f1(x)+(1-w(x))f2(x) Among them, w(x) is the weight function, and the weight function is designed using quadratic interpolation: Among them, x1 and x2 are the positions on both sides of the boundary, and the weight function will achieve continuous transition at the boundary as the position changes.

Images