CN116151135A - Electromagnetic simulation method and system for electric large-size target - Google Patents
Electromagnetic simulation method and system for electric large-size target Download PDFInfo
- Publication number
- CN116151135A CN116151135A CN202310436666.0A CN202310436666A CN116151135A CN 116151135 A CN116151135 A CN 116151135A CN 202310436666 A CN202310436666 A CN 202310436666A CN 116151135 A CN116151135 A CN 116151135A
- Authority
- CN
- China
- Prior art keywords
- electric field
- electric
- axis direction
- electromagnetic simulation
- field component
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000004088 simulation Methods 0.000 title claims abstract description 67
- 238000000034 method Methods 0.000 title claims abstract description 51
- 230000005684 electric field Effects 0.000 claims abstract description 73
- 238000003062 neural network model Methods 0.000 claims abstract description 32
- 238000013213 extrapolation Methods 0.000 claims abstract description 9
- 230000009466 transformation Effects 0.000 claims abstract description 7
- 239000013598 vector Substances 0.000 claims description 53
- 238000005070 sampling Methods 0.000 claims description 36
- 230000006870 function Effects 0.000 claims description 21
- 239000011159 matrix material Substances 0.000 claims description 15
- 230000007246 mechanism Effects 0.000 claims description 7
- 230000008569 process Effects 0.000 claims description 7
- 210000002569 neuron Anatomy 0.000 claims description 6
- 238000009827 uniform distribution Methods 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 3
- 238000010276 construction Methods 0.000 claims description 3
- 239000000284 extract Substances 0.000 claims description 2
- 238000004364 calculation method Methods 0.000 description 21
- 238000013528 artificial neural network Methods 0.000 description 12
- 238000004422 calculation algorithm Methods 0.000 description 9
- 238000013135 deep learning Methods 0.000 description 8
- 230000006872 improvement Effects 0.000 description 8
- 238000012549 training Methods 0.000 description 8
- 230000005672 electromagnetic field Effects 0.000 description 7
- 210000004027 cell Anatomy 0.000 description 4
- 238000012795 verification Methods 0.000 description 3
- 230000006399 behavior Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 230000035699 permeability Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000012952 Resampling Methods 0.000 description 1
- 238000010521 absorption reaction Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 239000004020 conductor Substances 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 230000001131 transforming effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/20—Finite element generation, e.g. wire-frame surface description, tesselation
- G06T17/205—Re-meshing
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E60/00—Enabling technologies; Technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Geometry (AREA)
- Computer Graphics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to an electromagnetic simulation method and system for an electric large-size target, belongs to the technical field of electromagnetic simulation, and solves the problems of long time consumption and low precision of electromagnetic simulation for the electric large-size target in the prior art. Comprising the following steps: three-dimensional modeling and grid division are carried out on the electric large-size target, and a simulation space is obtained; based on the time step, transmitting the three-dimensional coordinates of grid points to be simulated in the simulation space and corresponding moments into a trained neural network model, and predicting electric field and magnetic field components of the grid points at the corresponding moments; taking the electric field and magnetic field components at each moment as scattered near fields in a period of time; based on the near-far field extrapolation principle, obtaining a far-field scattered electric field according to an electric field component in the scattered near field; based on Fourier transformation, radar scattering cross section data are calculated according to the far-zone scattering electric field and the simulated selected incident wave. And the electromagnetic simulation of a rapid and accurate electric large-size target is realized.
Description
Technical Field
The invention relates to the technical field of electromagnetic simulation, in particular to an electromagnetic simulation method and system for an electric large-size target.
Background
In the research of scattering characteristics of an electric large-size target, the electric large-size target needs to establish a full electromagnetic field simulation model by using a method of calculating electromagnetism so as to simulate a scattering mechanism of a real target. With the deeper research of the electromagnetic field, how to meet the demands of large-scale calculation and calculation efficiency and accuracy in the electromagnetic simulation system is a problem that is getting more and more attention as the calculation scale of the electromagnetic simulation system is further enlarged.
The traditional FDTD (Finite Time-Domain) algorithm is mainly used for solving a Maxwell equation set, an explicit frog-and-fly (leap-frog) format is adopted on Time steps and space points, electromagnetic field E, H components are scattered in space and Time in an alternating sampling mode, a Maxwell Wei Xuandu equation containing Time variables is converted into a set of differential equations in a discrete mode, and the algorithm simulates the change of electromagnetic waves in space and Time in the Time Domain through the discrete Maxwell equation set. Since fine meshing requires a long time for iterative solution, a large amount of computing resources are required to support electromagnetic simulation computation in the face of large-scale electrically large-sized targets. When larger calculation regions are used to ensure accuracy, the computational complexity increases and may lead to numerical instability.
The GPU is adopted to improve the calculation efficiency of the FDTD, and the problem of large calculation scale in electromagnetic simulation is solved to a certain extent. However, when the complexity of the problem rises to a certain level, it becomes very difficult to solve the problem by completely relying on the conventional calculation, and a large amount of computer resources are required to be consumed. Many conventional fields begin to try to use deep learning to assist calculation, however, in the existing method of using a deep learning network model, a time step and a space step are still determined according to a stability condition, iterating is performed in the time step, and an applied method or a differential method is performed, each step has an error in the iterating process, the error is accumulated continuously, so that memory and calculation time are relatively large, and calculation accuracy is reduced.
Disclosure of Invention
In view of the above analysis, the embodiment of the invention aims to provide an electromagnetic simulation method and an electromagnetic simulation system for an electric large-size target, which are used for solving the problems of long time consumption and low precision of electromagnetic simulation of the existing electric large-size target.
In one aspect, an embodiment of the present invention provides an electromagnetic simulation method for an electrically large-sized target, including the steps of:
three-dimensional modeling and grid division are carried out on the electric large-size target, and a simulation space is obtained;
based on the time step, transmitting the three-dimensional coordinates of grid points to be simulated in the simulation space and corresponding moments into a trained neural network model, and predicting electric field and magnetic field components of the grid points at the corresponding moments; taking the electric field and magnetic field components at each moment as scattered near fields in a period of time;
based on the near-far field extrapolation principle, obtaining a far-field scattered electric field according to an electric field component in the scattered near field; based on Fourier transformation, radar scattering cross section data are calculated according to the far-zone scattering electric field and the simulated selected incident wave.
Based on the further improvement of the method, the structure of the neural network model sequentially comprises: the device comprises a first full-connection module, a transducer module and a second full-connection module; the model input is a matrix of n×4, the model output is a matrix of n×3, and N is the number of samples input.
Based on further improvement of the method, a mean square error is adopted for a loss function of the neural network model, wherein the error is an electric field and magnetic field component error constructed based on a two-dimensional FDTD iterative equation, and the error comprises an electric field component loss in the x-axis direction, an electric field component loss in the y-axis direction and a magnetic field component loss in the z-axis direction.
Based on a further improvement of the above method, the loss function of the neural network model is obtained by the following formula:
wherein ,
representing the loss of the electric field component in the x-axis direction, +.>
Representing the loss of the electric field component in the y-axis direction, +.>
Indicating the loss of the magnetic field component in the z-axis direction,Mrepresents the number of samples in each batch, +.>
Representing dielectric permittivity,/-, and>
indicating permeability->
、/>
and />
Respectively represent input samples +.>
The corresponding three outputs: an electric field component in the x-axis direction, an electric field component in the y-axis direction, and a magnetic field component in the z-axis direction.
Based on further improvement of the method, the first full-connection module maps Cheng Gaowei vectors of four-dimensional space-time vectors of each sample in the model input, processes the high-dimensional vectors through a nonlinear activation function to obtain new feature vectors, and transmits the new feature vectors into the transducer module; the four-dimensional space-time vector includes an x-axis coordinate value, a y-axis coordinate value, a z-axis coordinate value, and a time value.
Based on a further improvement of the above method, the transducer module comprises: the position coding is carried out by using a sine and cosine function according to the parity of different positions of the output vector of the first full connection module, and the vector added with the position coding is input into the coder; the encoder outputs the characteristic vector according to the self-attention mechanism and transmits the characteristic vector into the decoder; the decoder extracts information from the feature vectors using the set of attention vectors.
Based on a further improvement of the method, the output layer in the second fully-connected module is set to 3 neurons, so that each sample outputs 3 values, which respectively represent an electric field component in the x-axis direction, an electric field component in the y-axis direction and a magnetic field component in the z-axis direction.
Based on further improvement of the method, the trained neural network model is obtained through the following steps:
constructing a three-dimensional space according to the region of the electric large-size target, respectively obtaining sampling points and time values through uniform distribution, wherein the three-dimensional coordinate value of each sampling point and the corresponding time value are taken as a sample, and obtaining a sample set;
and inputting the sample set into the neural network model for unsupervised learning according to batches, obtaining output in forward propagation, calculating a loss function in backward propagation, optimizing model parameters until iteration is finished or preset precision is reached, and obtaining the trained neural network model.
Based on the further improvement of the method, a three-dimensional space is constructed according to the region of the electric large-size target, and sampling points and time values are respectively obtained through uniform distribution, and the method comprises the following steps:
according to the region of the electric large-size target, constructing a sampling space which does not exceed the region;
based on preset sampling density and sampling point number, respectively acquiring coordinate values on an x axis, a y axis and a z axis in a sampling space according to the generated uniform random number, and combining the coordinate values to obtain sampling points;
and taking one electromagnetic simulation period as a time interval, and generating uniformly distributed numerical values as time values in the time interval.
In another aspect, an embodiment of the present invention provides an electromagnetic simulation system for an electrically large-sized target, including:
the simulation model construction module is used for carrying out three-dimensional modeling and grid division on the electric large-size target to obtain a simulation space;
the simulation data acquisition module is used for transmitting the three-dimensional coordinates of grid points to be simulated in the simulation space and corresponding moments into the trained neural network model based on the time step, and predicting electric field and magnetic field components of the grid points at the corresponding moments; taking the electric field and magnetic field components at each moment as scattered near fields in a period of time;
the RCS data acquisition module is used for acquiring a far-field scattered electric field according to an electric field component in a scattered near field based on a near-far field extrapolation principle; based on Fourier transformation, radar scattering cross section data are calculated according to the far-zone scattering electric field and the simulated selected incident wave.
Compared with the prior art, the invention has at least one of the following beneficial effects: by combining deep learning with traditional numerical iteration, the whole learning process does not select grid step length for iteration, but avoids step length problem and truncation error of grid division through the capability of approximating any continuous function of a neural network, and the accuracy is ensured without adopting a larger calculation area when large-size simulation is performed, so that the calculation complexity is reduced; the solution of the differential equation is represented by a self-attention model in deep learning, and the solution of the numerical value is calculated by integrating all information through a neural network, so that the field quantities of the magnetic field and the electric field at the current moment in different spaces are output, and compared with the traditional FDTD algorithm, the calculation operation efficiency of field quantities one by one is higher, and the method is easy to expand to a distributed platform. The method is used for simulating the electric large-size target, so that the influence of errors of FDTD differential iteration on simulation results can be greatly reduced, the simulation efficiency is greatly improved, and the accuracy of calculation results of radar scattering cross sections is improved.
In the invention, the technical schemes can be mutually combined to realize more preferable combination schemes. Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention may be realized and attained by the structure particularly pointed out in the written description and drawings.
Drawings
The drawings are only for purposes of illustrating particular embodiments and are not to be construed as limiting the invention, like reference numerals being used to designate like parts throughout the drawings;
fig. 1 is a flowchart of an electromagnetic simulation method for an electrically large-sized object in embodiment 1 of the present invention.
Detailed Description
Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form a part hereof, and together with the description serve to explain the principles of the invention, and are not intended to limit the scope of the invention.
Example 1
In one embodiment of the present invention, an electromagnetic simulation method for an electrically large-sized object is disclosed, as shown in fig. 1, comprising the steps of:
and S11, performing three-dimensional modeling and grid division on the electric large-size target to obtain a simulation space.
It should be noted that, the electrical large-size target is the target to be simulated, and a three-dimensional modeling tool, such as CAD software, is used to build a three-dimensional geometric model of the target to be simulated; and selecting different grid densities according to the complexity to be simulated and the precision requirement of the simulation result, and carrying out grid division on the three-dimensional geometric model. In addition, the method also comprises the steps of setting material parameters of an object to be simulated, such as electromagnetic parameters of dielectric constant, magnetic permeability and the like, and setting boundary conditions, such as a free space boundary, an electric conductor boundary, an absorption boundary and the like.
The simulation space is the complete area or partial area of the electrically large target.
S12, based on the time step, transmitting the three-dimensional coordinates of grid points to be simulated in the simulation space and corresponding moments into a trained neural network model, and predicting electric field and magnetic field components of the grid points at the corresponding moments; the electric field and magnetic field components at each moment are taken as scattered near fields over a period of time.
In the prior art, an FDTD algorithm is adopted to solve a numerical solution of an electromagnetic field differential equation by a differential iteration method, but in the embodiment, a neural network model is trained by an unsupervised learning method, and a mapping relationship between space-time vectors and electric field and magnetic field components is established, so that the FDTD algorithm is replaced, and the electric field and magnetic field components at the current moment are obtained directly according to the input space coordinates and moment.
Specifically, the structure of the neural network model sequentially includes: the device comprises a first full-connection module, a transducer module and a second full-connection module; the model input is a matrix of n×4, the model output is a matrix of n×3, N is the number of samples input, namely: each sample is a four-dimensional space-time vector, and comprises an x-axis coordinate value, a y-axis coordinate value, a z-axis coordinate value and a time value; the output of each sample is a three-dimensional vector comprising: an electric field component in the x-axis direction, an electric field component in the y-axis direction, and a magnetic field component in the z-axis direction.
From the network structure, the first full-connection module maps Cheng Gaowei vectors of four-dimensional vectors of each sample in the model input, processes high-dimensional vectors through a nonlinear activation function to obtain new feature vectors, and transmits the new feature vectors into the transducer module.
Preferably, the first fully-connected module adopts a feedforward neural network FNN, wherein the number of hidden layers and the number of dimensions are set according to the number of samples, so as to obtain better expression capability and generalization capability. Illustratively, when the number of samples is less than 10000, the learning rate is set to 0.001, and the number of FNN hidden layer neurons is set to 50; setting the learning rate to be 0.0005 and the number of FNN hidden layer neurons to be 150 when the number of samples is [10000,50000 ]; when the number of samples is greater than 50000, the learning rate is set to 0.0001, and the number of FNN hidden layer neurons is 200.
The transducer module includes: position encoder, encoder and decoder.
Specifically, the position coding uses sine or cosine functions to perform position coding according to the parity of different positions of the output vector of the first full connection module, so that the model can sense the relative distance of different positions in the input sequence, and feature extraction and reasoning can be performed better.
In situations where computational power is limited, using an attention mechanism to dynamically allocate resources handles more important information, for increasing a position-coded vector sequence
After being mapped to three attention vectors through three weight matrixes, the three weight matrixes are input into a self-attention mechanism to encode the attention vectors, and the formula is as follows:
wherein ,
is the input vector +.>
Dimension of->
，/>
，/>
Query weight matrix, key weight matrix and value, respectivelyA weight matrix; each input vector is mapped linearly to a corresponding attention vector according to three weight matrices: query vector, key vector and value vector are respectively stacked to form matrix, and then the query vector matrix is calculatedQAnd key vector matrixKWhen the inner product of each row of vectors is equal to the inner product, the inner product is divided by +.>
Square root of (2); />
Representing a key vector matrixKIs a transposed matrix of (a).
To extract more interactive information in the sequence information, multi-head self-attention is added, and the multi-head self-attention is formed by combining a plurality of self-attention and is expressed by the following formula:
wherein ,
for outputting the projection matrix>
Is the firstmSelf-attention, self-care>
,/>
and />
Is the firstmThree of the self-attentiveness are provided.
The encoder in the transducer module is used for capturing the dependency relationship of the input sequence and sequentially comprises a multi-head self-attention mechanism and a feedforward neural network, wherein the self-attention mechanism and the feedforward neural network are connected through residual errors and normalization; the decoder includes a multi-headed self-attention module, a cross-attention module, and a feed-forward neural network. And taking the output of the multi-head self-attention module of the decoder as a query vector, dynamically calculating weights in the cross attention module according to the query vector and the key vector and the value vector output by the encoder, extracting information, transmitting the information into the feedforward neural network, and outputting all information obtained by the synthesis of the feedforward neural network to the second full-connection module.
The output layer in the second fully-connected module is set to 3 neurons, so that each sample outputs 3 values, which respectively represent an electric field component in the x-axis direction, an electric field component in the y-axis direction and a magnetic field component in the z-axis direction.
Further, the neural network model's loss function employs a mean square error (Mean Square Error, MSE), where the error is an electric and magnetic field component error constructed based on a two-dimensional FDTD iterative equation, including the loss of the electric field component in the x-axis direction
Loss of electric field component in y-axis direction +.>
And loss of magnetic field component in the z-axis direction +.>
. That is, the present embodiment uses the iterative formula of the conventional two-dimensional FDTD as a priori knowledge to provide the optimization direction of the neural network model.
Specifically, the loss function of the neural network model is expressed by the following formula:
wherein ,Mindicating the number of samples in each batch,
representing dielectric permittivity,/-, and>
representing permeability, space-time vector of input sample +.>
The corresponding output result is the electric field component in the x-axis direction +.>
Electric field component in the y-axis direction +.>
And magnetic field component in the z-axis direction +.>
Similarly, the->
、/>
And
is a space-time vector +.>
Is provided.
From the loss function, the method is obtained through a neural network model
After corresponding output, constructing ++according to the preset step length of each dimension>
、/>
、/>
、/>
And
the space-time vector afferent neural network model obtains respective outputs, and then substitutes the corresponding outputs into a loss function formula to calculate +.>
Loss function of sample.
After the neural network model is built, training is carried out through the following steps of:
(1) and constructing a three-dimensional space according to the electrically large-size target area, respectively obtaining sampling points and time values through uniform distribution, wherein the three-dimensional coordinate value of each sampling point and the corresponding time value are taken as a sample, and obtaining a sample set.
Specifically, the sampling is performed by:
according to the region of the electric large-size target, constructing a sampling space which does not exceed the region;
based on preset sampling density and sampling point number, respectively acquiring coordinate values on an x axis, a y axis and a z axis in a sampling space according to the generated uniform random number, and combining the coordinate values to obtain sampling points;
and taking one electromagnetic simulation period as a time interval, and generating uniformly distributed numerical values as time values in the time interval.
It should be noted that the higher the sampling density, the greater the number of sampling points and the more accurate the sampling result. The sampling density may be balanced against the simulation requirements and computational resources. The sampling density may be determined by the number of sampling points, which should generally be able to cover the sampling space.
According to the sampling density, equally dividing three axial dimensions (x axis, y axis and z axis) of a sampling space into a plurality of cells, generating a group of random numbers uniformly distributed in the [0,1] sections, and mapping the random numbers to the cells of each axial dimension respectively through the following formula to obtain corresponding coordinate values:
wherein ,
representation ofrThe coordinate values of the axial dimension,rrepresents x, y or z, < >>
and />
Is thatrUpper and lower bounds of the cell-to-cell in the axis dimension,randis a random number.
Illustratively, the sampling density is 100, the uniform random number generated is 0.3, and the sampling range in the x-axis dimension is [0,1]The length of each cell is 0.01, and the uniform random number is mapped to the 31 st cell of the x-axis dimension, and the coordinate value of the x-axis dimension
。
For the time of day value, a set of random numbers uniformly distributed over a time interval, illustratively [0,20] seconds, is directly obtained.
(2) And inputting the sample set into the neural network model for unsupervised learning according to batches, obtaining output in forward propagation, calculating a loss function in backward propagation, optimizing model parameters until iteration is finished or preset precision is reached, and obtaining the trained neural network model.
It should be noted that, the sample set may be proportionally divided into a training set and a verification set, for example, 5:1, where each round of learning is performed through the training set, and each round of learning effect is verified through the verification set. Each round of learning employs resampling to generate a sample set.
During training, in forward propagation, 3 output values are mapped according to the final output layer, the current loss gradient is calculated through a backward propagation algorithm and a random gradient descent algorithm, and current model parameters are updated until the maximum iteration number is reached or the current model meets the error precision of a verification set, and training is stopped.
Preferably, three-dimensional coordinates of grid points, and real electric field and magnetic field components at each moment are extracted from the historical simulation data to serve as a test set, and the performance of the trained neural network model on the test set is verified.
And the training process is migrated to the multi-GPU distributed model to provide uniform gradient calculation and updating, so that the model training efficiency is improved by more effectively utilizing the GPU performance.
Compared with the prior art, the embodiment combines deep learning and traditional numerical iteration, and the step length problem and truncation error of grid division are avoided by the approximate continuous function capability of the neural network. The nonlinear characteristics of deep learning can be more generalized to complex electromagnetic field behaviors, and uncertainty behaviors in the electromagnetic field can be better simulated through the uncertain combination of the super parameters of the neural network.
After training the neural network model, predicting electric field and magnetic field components of the grid points at corresponding moments according to the three-dimensional coordinates and the corresponding moments of the input grid points to be simulated; after a plurality of moments are transmitted according to time step length, electric field and magnetic field components at each moment are used as scattering near fields in a period of time.
S13, obtaining a far-field scattered electric field according to an electric field component in a scattered near-field based on a near-far-field extrapolation principle; based on Fourier transformation, radar scattering cross section data are calculated according to the far-zone scattering electric field and the simulated selected incident wave.
The far-field scattered field is required for calculation of the radar cross section, and thus the value of the far-field scattered field is extrapolated using the scattered near-field value obtained in step S12.
Specifically, electric field components at all times in the z-axis direction are obtained according to the electric field components at all times in the x-axis direction and the y-axis direction;
obtaining the scattering electric field of the far-field at each moment based on the near-far-field extrapolation principle according to the electric field components at each moment in the z-axis direction
；
Fourier transform is carried out through the following formula to obtain a frequency domain field corresponding to the far-region scattered electric field
：
Fourier transforming the incident wave to obtain its frequency domain field
；
Radar cross section data RCS is calculated by the following formula:
wherein ,Rrepresenting the distance of the grid point from the radar,lg(. Cndot.) represents taking the logarithm.
It should be noted that, compared with the result obtained by solving the traditional FDTD algorithm, the field value of the electromagnetic field obtained by using the self-focusing neural network model is improved in precision, so that the field value of the remote field obtained by extrapolation is more accurate, and the precision of the calculation result of the RCS is improved.
Compared with the prior art, the electromagnetic simulation method for the electric large-size target combines deep learning with traditional numerical iteration, the whole learning process does not select grid step length for iteration, but avoids step length problems and truncation errors of grid division through the capability of a neural network for approximating any continuous function, and accuracy is ensured without adopting a larger calculation area during electric large-size simulation, so that the calculation complexity is reduced; the solution of the differential equation is represented by a self-attention model in deep learning, and the solution of the numerical value is calculated by integrating all information through a neural network, so that the field quantities of the magnetic field and the electric field at the current moment in different spaces are output, and compared with the traditional FDTD algorithm, the calculation operation efficiency of field quantities one by one is higher, and the method is easy to expand to a distributed platform. The method is used for simulating the electric large-size target, so that the influence of errors of FDTD differential iteration on simulation results can be greatly reduced, the simulation efficiency is greatly improved, and the accuracy of calculation results of radar scattering cross sections is improved.
Example 2
In another embodiment of the present invention, an electromagnetic simulation system of an electrically large-sized object is disclosed, thereby realizing an electromagnetic simulation method of an electrically large-sized object in embodiment 1. The specific implementation of each module is described with reference to the corresponding description in embodiment 1. The system comprises:
the simulation model construction module is used for carrying out three-dimensional modeling and grid division on the electric large-size target to obtain a simulation space;
the simulation data acquisition module is used for transmitting the three-dimensional coordinates of grid points to be simulated in the simulation space and corresponding moments into the trained neural network model based on the time step, and predicting electric field and magnetic field components of the grid points at the corresponding moments; taking the electric field and magnetic field components at each moment as scattered near fields in a period of time;
the RCS data acquisition module is used for acquiring a far-field scattered electric field according to an electric field component in a scattered near field based on a near-far field extrapolation principle; based on Fourier transformation, radar scattering cross section data are calculated according to the far-zone scattering electric field and the simulated selected incident wave.
Since the relevant points of the electromagnetic simulation method for the large-size object in this embodiment can be referred to each other, the description is repeated here, and thus the description is omitted here. The principle of the system embodiment is the same as that of the method embodiment, so the system embodiment also has the corresponding technical effects of the method embodiment.
Those skilled in the art will appreciate that all or part of the flow of the methods of the embodiments described above may be accomplished by way of a computer program to instruct associated hardware, where the program may be stored on a computer readable storage medium. Wherein the computer readable storage medium is a magnetic disk, an optical disk, a read-only memory or a random access memory, etc.
The present invention is not limited to the above-mentioned embodiments, and any changes or substitutions that can be easily understood by those skilled in the art within the technical scope of the present invention are intended to be included in the scope of the present invention.
Claims (10)
1. An electromagnetic simulation method of an electrically large-size target is characterized by comprising the following steps:
three-dimensional modeling and grid division are carried out on the electric large-size target, and a simulation space is obtained;
based on the time step, transmitting the three-dimensional coordinates of grid points to be simulated in the simulation space and corresponding moments into a trained neural network model, and predicting electric field and magnetic field components of the grid points at the corresponding moments; taking the electric field and magnetic field components at each moment as scattered near fields in a period of time;
based on the near-far field extrapolation principle, obtaining a far-field scattered electric field according to an electric field component in the scattered near field; based on Fourier transformation, radar scattering cross section data are calculated according to the far-zone scattering electric field and the simulated selected incident wave.
2. The electromagnetic simulation method of an electrically large-sized object according to claim 1, wherein the structure of the neural network model sequentially includes: the device comprises a first full-connection module, a transducer module and a second full-connection module; the model input is a matrix of n×4, the model output is a matrix of n×3, and N is the number of samples input.
3. The electromagnetic simulation method of an electrically large-sized object according to claim 2, wherein the loss function of the neural network model adopts a mean square error, wherein the error is an electric field and magnetic field component error constructed based on a two-dimensional FDTD iterative equation, and includes an electric field component loss in an x-axis direction, an electric field component loss in a y-axis direction, and a magnetic field component loss in a z-axis direction.
4. The electromagnetic simulation method of an electrically large-sized object according to claim 3, wherein the loss function of the neural network model is obtained by the following formula:
wherein ,
representing the loss of the electric field component in the x-axis direction, +.>
Indicating the loss of the electric field component in the y-axis direction,
indicating the loss of the magnetic field component in the z-axis direction,Mrepresents the number of samples in each batch, +.>
Representing dielectric permittivity,/-, and>
indicating permeability->
、/>
and />
Respectively represent input samples +.>
The corresponding three outputs: an electric field component in the x-axis direction, an electric field component in the y-axis direction, and a magnetic field component in the z-axis direction.
5. The electromagnetic simulation method of the electric large-size target according to claim 2, wherein the first full-connection module maps Cheng Gaowei vectors of four-dimensional space-time vectors of each sample in the model input, processes high-dimensional vectors through a nonlinear activation function to obtain new feature vectors, and transmits the new feature vectors into the transducer module; the four-dimensional space-time vector includes an x-axis coordinate value, a y-axis coordinate value, a z-axis coordinate value, and a time value.
6. The electromagnetic simulation method of an electrically large-sized object according to claim 2, wherein the transducer module comprises: the position coding is carried out by using a sine and cosine function according to the parity of different positions of the output vector of the first full connection module, and the vector added with the position coding is input into the encoder; the encoder outputs the feature vector according to the self-attention mechanism and transmits the feature vector into the decoder; the decoder extracts information from the feature vectors using the set of attention vectors.
7. The electromagnetic simulation method of an electrically large-sized object according to claim 2, wherein the output layer in the second fully-connected module is set to 3 neurons, so that each sample outputs 3 values representing an electric field component in the x-axis direction, an electric field component in the y-axis direction, and a magnetic field component in the z-axis direction, respectively.
8. The electromagnetic simulation method of an electrically large-sized target according to claim 2, wherein the trained neural network model is obtained by:
constructing a three-dimensional space according to the region of the electric large-size target, respectively obtaining sampling points and time values through uniform distribution, wherein the three-dimensional coordinate value of each sampling point and the corresponding time value are taken as a sample, and obtaining a sample set;
and inputting the sample set into the neural network model for unsupervised learning according to batches, obtaining output in forward propagation, calculating a loss function in backward propagation, optimizing model parameters until iteration is finished or preset precision is reached, and obtaining the trained neural network model.
9. The electromagnetic simulation method of an electrically large-sized object according to claim 8, wherein the constructing a three-dimensional space according to the region of the electrically large-sized object, respectively obtaining the sampling points and the time values by uniform distribution, comprises:
constructing a sampling space not exceeding the area according to the area of the electric large-size target;
based on preset sampling density and sampling point number, respectively acquiring coordinate values on an x axis, a y axis and a z axis in a sampling space according to the generated uniform random number, and combining the coordinate values to obtain sampling points;
and taking one electromagnetic simulation period as a time interval, and generating uniformly distributed numerical values as time values in the time interval.
10. An electromagnetic simulation system of an electrically large-sized object, comprising:
the simulation model construction module is used for carrying out three-dimensional modeling and grid division on the electric large-size target to obtain a simulation space;
the simulation data acquisition module is used for transmitting the three-dimensional coordinates of grid points to be simulated in the simulation space and corresponding moments into the trained neural network model based on the time step, and predicting electric field and magnetic field components of the grid points at the corresponding moments; taking the electric field and magnetic field components at each moment as scattered near fields in a period of time;
the RCS data acquisition module is used for acquiring a far-field scattered electric field according to an electric field component in a scattered near field based on a near-far field extrapolation principle; based on Fourier transformation, radar scattering cross section data are calculated according to the far-zone scattering electric field and the simulated selected incident wave.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310436666.0A CN116151135B (en) | 2023-04-23 | 2023-04-23 | Electromagnetic simulation method and system for electric large-size target |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310436666.0A CN116151135B (en) | 2023-04-23 | 2023-04-23 | Electromagnetic simulation method and system for electric large-size target |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116151135A true CN116151135A (en) | 2023-05-23 |
| CN116151135B CN116151135B (en) | 2023-06-27 |
Family
ID=86354728
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310436666.0A Active CN116151135B (en) | 2023-04-23 | 2023-04-23 | Electromagnetic simulation method and system for electric large-size target |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116151135B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117725805A (en) * | 2024-02-08 | 2024-03-19 | 合肥工业大学 | Magnetic field rapid calculation method of optimized depth operator network |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020119107A1 (en) * | 2018-12-13 | 2020-06-18 | 北京遥测技术研究所 | Integrated research and development platform for terahertz focal plane imaging system |
| CN111581886A (en) * | 2020-05-12 | 2020-08-25 | 上海无线电设备研究所 | Electromagnetic field rapid simulation solving method based on convolutional neural network parallel acceleration |
| CN112114295A (en) * | 2020-07-29 | 2020-12-22 | 中国人民解放军海军航空大学 | Target identification method and system for full-polarization radar |
| CN113721212A (en) * | 2021-07-16 | 2021-11-30 | 西安电子科技大学 | Radar scattering cross section confidence evaluation method based on neural network |
| CN113722951A (en) * | 2021-08-17 | 2021-11-30 | 西安中电科西电科大雷达技术协同创新研究院有限公司 | Scatterer three-dimensional finite element grid optimization method based on neural network |
| CN114611387A (en) * | 2022-03-01 | 2022-06-10 | 中国人民解放军国防科技大学 | Method and device for improving measurement precision of electromagnetic characteristics of equipment |
| US11567171B1 (en) * | 2017-10-09 | 2023-01-31 | Ansys, Inc. | Systems and methods for radar signature modeling using a rotating range profile reduced-order model |
-
2023
- 2023-04-23 CN CN202310436666.0A patent/CN116151135B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11567171B1 (en) * | 2017-10-09 | 2023-01-31 | Ansys, Inc. | Systems and methods for radar signature modeling using a rotating range profile reduced-order model |
| WO2020119107A1 (en) * | 2018-12-13 | 2020-06-18 | 北京遥测技术研究所 | Integrated research and development platform for terahertz focal plane imaging system |
| CN111581886A (en) * | 2020-05-12 | 2020-08-25 | 上海无线电设备研究所 | Electromagnetic field rapid simulation solving method based on convolutional neural network parallel acceleration |
| CN112114295A (en) * | 2020-07-29 | 2020-12-22 | 中国人民解放军海军航空大学 | Target identification method and system for full-polarization radar |
| CN113721212A (en) * | 2021-07-16 | 2021-11-30 | 西安电子科技大学 | Radar scattering cross section confidence evaluation method based on neural network |
| CN113722951A (en) * | 2021-08-17 | 2021-11-30 | 西安中电科西电科大雷达技术协同创新研究院有限公司 | Scatterer three-dimensional finite element grid optimization method based on neural network |
| CN114611387A (en) * | 2022-03-01 | 2022-06-10 | 中国人民解放军国防科技大学 | Method and device for improving measurement precision of electromagnetic characteristics of equipment |
Non-Patent Citations (1)
| Title |
|---|
| 黄思倩: "电大尺寸目标的FDTD并行计算", 《中国优秀硕士学位论文全文数据库 基础科学辑》, no. 04, pages 17 - 28 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117725805A (en) * | 2024-02-08 | 2024-03-19 | 合肥工业大学 | Magnetic field rapid calculation method of optimized depth operator network |
| CN117725805B (en) * | 2024-02-08 | 2024-04-30 | 合肥工业大学 | Magnetic field rapid calculation method of optimized depth operator network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116151135B (en) | 2023-06-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Biancolini | Fast radial basis functions for engineering applications | |
| CN111812732B (en) | Geoelectromagnetic nonlinear inversion method based on convolutional neural network | |
| CN116151135B (en) | Electromagnetic simulation method and system for electric large-size target | |
| Sharpe et al. | Topology design with conditional generative adversarial networks | |
| CN113158492B (en) | Full-implicit double-time-step calculation method of time-varying electromagnetic field | |
| CN116451584B (en) | Thermal stress prediction method and system based on neural network | |
| Oishi et al. | A surface-to-surface contact search method enhanced by deep learning | |
| CN113158527A (en) | Method for calculating frequency domain electromagnetic field based on implicit FVFD | |
| Capek et al. | Characteristic mode decomposition using the scattering dyadic in arbitrary full-wave solvers | |
| Nobari et al. | Range-constrained generative adversarial network: Design synthesis under constraints using conditional generative adversarial networks | |
| CN116757052A (en) | Particle space-time propulsion algorithm under quasi-circular symmetry star simulator magnetic field configuration | |
| CN111144013A (en) | High-precision medium object scattering simulation method | |
| CN114692529B (en) | CFD high-dimensional response uncertainty quantification method and device, and computer equipment | |
| CN115659619A (en) | Geometric topological optimization and additive manufacturing based integrated method | |
| CN113313834B (en) | Point cloud generation method based on geometric interpretability | |
| CN115482893A (en) | Electromagnetic metamaterial design method based on deep learning and structural variables | |
| Kaplan et al. | Fast iterative integral equation solver for acoustic scattering by inhomogeneous objects using the butterfly approximation | |
| Wu et al. | Linear geometric algebra rotor estimator for efficient mesh deformation | |
| Zhao et al. | Spatio-temporal Model Combining VMD and AM for Wind Speed Prediction | |
| Namburu et al. | Scalable electromagnetic simulation environment | |
| Duan et al. | Ship hull surface reconstruction from scattered points cloud using an RBF neural network mapping technology | |
| WO2022260179A1 (en) | Training device, training method, program, and inference device | |
| CN117892598B (en) | Enhanced inverse finite element shape sensing reconstruction system for offshore wind turbine tower | |
| CN113076583B (en) | Cantilever beam structure design method based on self-attention mechanism neural network | |
| Shen et al. | 3D modeling and visualization of river systems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |