Lens antenna multi-objective optimization method based on prior knowledge neural network

Abstract

The application discloses a lens antenna multi-target optimization method based on prior knowledge neural network, takes INN as the main body of the algorithm, adopts a reverse neural network to perform reverse design on the lens antenna, and the input of the reverse neural network is electromagnetic response, and the output is a structure parameter, so that the inversion of the antenna structure parameter of multiple performance indexes is realized. Meanwhile, multiple sub-FNNs are introduced to provide prior knowledge corresponding to multiple performance indexes, respectively, the input of the forward neural network is the structure parameter, and the output is the electromagnetic response. Compared with the existing lens antenna design method, the design method has the following advantages: 1) the antenna design efficiency is high, and the designed antenna can realize good multi-target characteristics; 2) the proposed KBANN model relieves the problem of large neural network data requirement and provides a new solution. The application can realize the rapid optimization of the multi-target lens antenna, and relieves the problem of large neural network data requirement.

Description

Technical Field

[0001] This invention relates to the field of lens antenna technology, and specifically to a multi-objective optimization method for lens antennas based on prior knowledge neural network (KBANN). Background Technology

[0002] The increasingly complex electromagnetic environment and application demands necessitate continuous improvements in the performance of systems such as radar, guidance, communication, biomedicine, electronic warfare, and radio astronomy, which in turn places increasingly stringent requirements on antenna performance. Antenna design must consider satisfying multiple specific performance indicators, such as impedance bandwidth, aperture efficiency, gain, gain bandwidth, beamforming, and polarization characteristics / bandwidth (circular polarization). Therefore, modern antenna design faces significant challenges. On the other hand, lens antennas offer advantages such as rich morphological variations and excellent electromagnetic characteristics. Furthermore, the development of 3D printing technology has made it possible to develop complex all-dielectric lenses, making all-dielectric lens antennas an excellent candidate for meeting the needs of future antenna systems. Therefore, researching multi-objective optimization techniques for lens antennas is crucial.

[0003] Currently, most antenna designs in domestic and international literature are based on traditional trial-and-error methods. This requires continuous simulation (parameter scanning) by changing the antenna's structural parameters to obtain its electrical performance, resulting in inconsistent design outcomes. Therefore, this method is extremely time-consuming, labor-intensive, and inefficient. With the further enhancement of computing power and the development of machine learning, machine learning methods (including neural networks, support vector regression, Gaussian process regression, etc.) are gradually being applied to the electromagnetic field to accelerate design, especially neural network-based methods, which have achieved good results. For example, some literature uses neural network methods to analyze the S-axis of antennas... 11 Some studies have focused on predicting the transmission amplitude and phase of frequency-selective surface elements to enable the design of multimode resonant antennas. Others have applied neural networks to array synthesis, enabling rapid retrieval of array information from radiation patterns. However, research on neural network-based lens antenna design, particularly multi-target lens antenna design, is still quite lacking. Summary of the Invention

[0004] This invention provides a multi-objective optimization method for lens antennas based on prior knowledge neural networks, which can achieve rapid optimization of multi-objective lens antennas while alleviating the problem of large data requirements of neural networks.

[0005] To solve the above problems, the present invention is achieved through the following technical solution:

[0006] The multi-objective optimization method for lens antennas based on prior knowledge neural networks includes the following steps:

[0007] Step 1: For a given lens antenna structure, collect N simulation data points (x...) using full-wave simulation. i y i ); where x i Let y be an n-dimensional simulation structure parameter vector. i for A 3D simulated electromagnetic response vector;

[0008] Step 2: Normalize the N simulation data (x) according to the maximum-minimum value normalization principle. i y i After normalization, N normalized simulation data points are obtained. in Let be an n-dimensional normalized simulation structure parameter vector. The normalized simulated electromagnetic response vector of dimension ;

[0009] Step 3: Construct a prior knowledge-based neural network consisting of m forward neural networks and 1 inverse neural network; the input of the forward neural network is the structural parameters, and the output is the electromagnetic response; the input of the inverse neural network is the electromagnetic response, and the output is the structural parameters.

[0010] Step 4: Utilize N normalized simulation data Each of the m forward neural networks based on prior knowledge is trained to obtain m trained forward neural networks; during the training of the forward neural network corresponding to the k-th electromagnetic response attribute, the normalized simulation structure parameter vector is... The normalized simulated electromagnetic response vector is used as input to this positive neural network. z-corresponding to the electromagnetic response k The dimension is used as the output of the forward neural network, and the training error and test error are made to meet the specified requirements, thereby obtaining the corresponding trained forward neural network;

[0011] Step 5: First, randomly generate N0 n-dimensional random structure parameter vectors x′. j Then, the random structure parameter vector x′ j The inputs are fed into m pre-trained forward neural networks, resulting in m z-values. k A dimensional random electromagnetic response vector; then m z-vectors k The electromagnetic response vectors of the two dimensions are concatenated into The electromagnetic response vector y′ of dimension j This yields N0 random data points (x′). j y′ j );

[0012] Step 6: Use N0 random data points (x′)j y′ j The inverse neural network in the prior knowledge-based neural network is trained to obtain a trained inverse neural network; during the training process of the inverse neural network, the random structure parameter vector y′ is... j The random structure parameter vector x′ serves as the input to this inverse neural network. j The training error and test error are used as the output of the inverse neural network, and the training error and test error are made to meet the specified requirements, thus obtaining the trained inverse neural network;

[0013] Step 7: Determine the requirements that the lens antenna to be designed must meet. The n-dimensional electromagnetic response vector y is input into the trained inverse neural network to obtain the n-dimensional target structure parameter vector x required for the design of the lens antenna;

[0014] In the above, i = 1, 2, ..., N, where N represents the number of simulation data sets; j = 1, 2, ..., N0, where N0 is the number of random data sets; N0 > N; n is the number of structural parameter attributes of the lens antenna to be designed; z k Let be the number of discrete points for the k-th electromagnetic response attribute, k = 1, 2, ..., m, where m is the number of electromagnetic response attributes that the lens antenna needs to satisfy.

[0015] In step 2 above, the structural parameter vector x i The formula for normalizing the label value of the p-th structural parameter attribute in the data is:

[0016]

[0017] For the electromagnetic response vector y i The formula for normalizing the q-th label value of the k-th electromagnetic response attribute is:

[0018]

[0019] In the formula, This represents the normalized label value of the p-th structural parameter attribute of the i-th simulation data. minx represents the label value of the p-th structural parameter attribute of the i-th simulation data. p Maxx represents the minimum label value of the p-th structural parameter attribute in N simulation data. p This represents the maximum label value of the p-th structural parameter attribute in N simulation data, where p = 1, 2, ..., n, and n represents the number of structural parameter attributes. This represents the q-th normalized label value of the k-th electromagnetic response attribute of the i-th simulation data. Miny represents the q-th label value of the k-th electromagnetic response attribute of the i-th simulation data. k Maxy represents the minimum label value of the k-th electromagnetic response attribute in N simulation data. k Let q represent the maximum label value of the k-th electromagnetic response attribute in N simulation data, where q = 1, 2, ..., z k , z k The number of discrete points represents the k-th electromagnetic response attribute, k = 1, 2, ..., m, where m is the number of electromagnetic response attributes of the lens antenna that need to be satisfied; i = 1, 2, ..., N, where N represents the number of simulation data.

[0020] In step 2 above, the normalized simulation data... Before normalization, the label values ​​of electromagnetic responses that change beyond a set threshold need to be simplified by step smoothing using a step smoothing method.

[0021] Compared with existing technologies, this invention proposes a multi-objective optimization method for lens antennas based on a prior knowledge neural network (KBANN). This invention uses an inverse neural network (INN) as the core of the algorithm, employing an inverse neural network to reverse-engineer the lens antenna. The input of the inverse neural network is the electromagnetic response, and the output is structural parameters, enabling the inversion of antenna structural parameters for multiple performance indicators. Considering the requirements of multi-objective design, and the difficulty in achieving good training results with a small amount of data while requiring a large amount of time with a large amount of data, this invention introduces multiple sub-FNNs to provide prior knowledge corresponding to multiple performance indicators, ensuring the lens antenna meets multiple performance requirements. The input of the forward neural network is the structural parameters, and the output is the electromagnetic response. Firstly, the forward neural network is easier to train, achieving better training results with less data compared to the inverse neural network; secondly, the forward network is usually set as a single-objective network, with a simple structure, allowing for better training compared to multi-objective inverse networks. Compared with existing lens antenna design methods, the design method proposed in this invention has the following advantages: 1) High antenna design efficiency, and the designed antenna can achieve good multi-target characteristics; 2) The proposed KBANN model provides a new solution to the problem of large data requirements in neural networks. This invention can achieve rapid optimization of multi-target lens antennas while alleviating the problem of large data requirements in neural networks. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the KBANN model.

[0023] Figure 2 The diagram shows a lens antenna model. (a) is a side view and (b) is a top view.

[0024] Figure 3 This is a concrete KBANN model.

[0025] Figure 4 The S-type lens antenna designed for this embodiment 11 curve.

[0026] Figure 5 Gain-axis ratio curve of the lens antenna designed for this implementation.

[0027] Figure 6 The radiation efficiency curve of the lens antenna designed for this implementation.

[0028] Figure 7 The radiation patterns of the lens antenna designed for this embodiment at 65 GHz are shown in (a) xoz plane and (b) yoz plane.

[0029] Figure 8 The radiation pattern of the lens antenna designed in this embodiment at 75 GHz is shown in (a) xoz plane pattern and (b) yoz plane pattern.

[0030] Figure 9 The radiation patterns of the lens antenna designed for this embodiment at 85 GHz are shown in (a) xoz plane and (b) yoz plane. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific examples.

[0032] Neural networks applied in electromagnetics include forward neural networks (FNNs) and inverse neural networks (INNs). Compared to FNNs, INNs, once trained, can directly obtain output parameters from the input target, making them more efficient for lens antenna design. However, the application of INNs in antenna design faces a significant challenge: data collection is extremely time-consuming, especially for complex mapping relationships where the data volume requirement is even greater. Solving this problem can further accelerate lens antenna design and provide technical support for the application of neural networks in the field of antenna electromagnetics. This invention proposes a multi-objective optimization method for lens antennas that combines inverse and forward neural networks to form a neural network based on prior knowledge, such as... Figure 1 As shown.

[0033] According to design requirements, the lens antenna in this embodiment mainly consists of a lens and a feed antenna. It operates at a frequency of 60-90 GHz and has two functions: firstly, to convert linearly polarized waves into circularly polarized waves; and secondly, to improve the gain of the feed antenna. Based on these requirements, the following is provided: Figure 2 The lens structure shown mainly consists of three parts: the bottom is a cylindrical sleeve and a small square sleeve, primarily used to reinforce the lens and fix it to the horn; the middle part is a cylinder composed of dielectric gratings (ε1 = 2.9, tanδ = 0.01), with a thickness of w and a gap of g between the gratings; this middle part, as the most important part of the lens, is mainly used to achieve linear-to-circular polarization conversion and also improve antenna gain; the top is an approximately conical shape, mainly used to achieve high gain and also to reinforce the lens. The feed antenna is a horn antenna, model LB-12-15-AwithWR12.

[0034] Based on a given lens antenna structure, the present invention proposes a multi-objective optimization method for lens antennas based on prior knowledge neural networks, which includes the following steps:

[0035] Step 1: For a given lens antenna structure, collect N simulation data points (x...) using full-wave simulation. i y i Where i = 1, 2, ..., N, and N represents the number of simulation data; x i Let y be an n-dimensional vector of simulated structural parameters, where n is the number of structural parameter attributes of the lens antenna to be designed; i for The simulated electromagnetic response vector is constructed by discretizing each electromagnetic response curve, since each electromagnetic response attribute is represented by an electromagnetic response curve. Therefore, zi is used to construct the simulated electromagnetic response vector. k Let be the number of discrete points for the k-th electromagnetic response attribute, k = 1, 2, ..., m, where m is the number of electromagnetic response attributes that the lens antenna needs to satisfy.

[0036] In this embodiment, the full-wave simulation method used is a CST-MATLAB co-simulation method. That is, after determining the upper and lower intervals of the structural parameters, MATLAB software randomly generates combinations of parameter values, which are then substituted into CST software for structural modeling and simulation to obtain the corresponding electromagnetic response values. Electromagnetic response values ​​are then extracted according to a sampling principle. The number of collected simulation data points, N, is 300. The structural parameter attributes, as shown in Figure 2, include g, w, hlow, R, and h, i.e., n=5. The electromagnetic response attributes (i.e., optimization objectives) include axial ratio, gain, and S. 11 That is, m = 3.

[0037] Step 2: Normalize the simulation data (x) according to the maximum-minimum value normalization principle. i y i Normalization is performed to obtain normalized simulation data. in Let be an n-dimensional normalized simulation structure parameter vector. for The normalized simulated electromagnetic response vector of dimension 1.

[0038] The purpose of normalization is twofold: firstly, to distribute variables that originally belong to different intervals into a unified interval; secondly, for a single variable, since a variable is usually multidimensional, the normalization operation can change its distribution interval from a large interval to a small interval, which is beneficial for the subsequent learning of the neural network.

[0039] For the structural parameter vector x i The formula for normalizing the label value of the p-th structural parameter attribute in the data is:

[0040]

[0041] In the formula, This represents the normalized label value of the p-th structural parameter attribute of the i-th simulation data. minx represents the label value of the p-th structural parameter attribute of the i-th simulation data. p Maxx represents the minimum label value of the p-th structural parameter attribute in N simulation data. p Let i represent the maximum label value of the p-th structural parameter attribute in N simulation data, where p = 1, 2, ..., n, and n represents the number of structural parameter attributes; i = 1, 2, ..., N, and N represents the number of simulation data.

[0042] For the electromagnetic response vector y i The formula for normalizing the q-th label value of the k-th electromagnetic response attribute is:

[0043]

[0044] In the formula, This represents the q-th normalized label value of the k-th electromagnetic response attribute of the i-th simulation data. Miny represents the q-th label value of the k-th electromagnetic response attribute of the i-th simulation data. k Maxy represents the minimum label value of the k-th electromagnetic response attribute in N simulation data. k Let q represent the maximum label value of the k-th electromagnetic response attribute in N simulation data, where q = 1, 2, ..., z k , z kThe number of discrete points represents the k-th electromagnetic response attribute, k = 1, 2, ..., m, where m is the number of electromagnetic response attributes of the lens antenna that need to be satisfied; i = 1, 2, ..., N, where N represents the number of simulation data.

[0045] Considering the simulation data (x) i y i The label values ​​of some electromagnetic responses in the simulation data (x) change drastically, which is detrimental to the learning of neural networks. i y i Before normalization, the label values ​​of these drastically changing (i.e., changing beyond a set threshold) electromagnetic responses need to be simplified using a step-smoothing method. This addresses the issue of simplification for some targets (such as radiation patterns, |S...). 11 |) Describe the difficult problem.

[0046] Step 3: Construct a prior knowledge-based neural network consisting of three forward neural networks and one inverse neural network. The input to the forward neural network is the structural parameters, and the output is the electromagnetic response. The input to the inverse neural network is the electromagnetic response, and the output is the structural parameters. For example... Figure 3 As shown.

[0047] The input to an inverse neural network is an electromagnetic response, and the output is structural parameters. In this embodiment, the inverse neural network includes an input layer, an output layer, and three hidden layers. The input of an INN can be written as... in, Represents the axial ratio in the electromagnetic response. Represents the gain in the electromagnetic response. |S represents the electromagnetic response 11 |,z1=z2=z3=31,that is It has 93 dimensions. The output of an INN can be written as a vector. n is the number of structural parameters. Once the network model is determined, the mapping relationship from input to output can be expressed by the following formula:

[0048]

[0049] in, This represents the output value of the input layer neurons. This represents the output value of the hidden layer neurons. l represents the output value of the neuron in the output layer. l represents the l-th layer, C l This represents the number of neurons in the l-th layer. It represents the connection weight between the c-th neuron and the d-th neuron in an adjacent layer. and The input and output parameters determine the mapping relationship between them. Training a neural network involves adjusting these two parameters to make the mapping relationship as accurate as possible. l (x) represents the activation function of the l-th layer.

[0050] The output layer activation function of an INN is a variation of tanh(x):

[0051]

[0052] The above formula limits the output range to [0,1], which aligns with the normalized data. After repeated trials during training, the activation function for the hidden layer was finally determined to be the ReLU(x) function:

[0053] f(x)=relu(x)=max(0,x) (3)

[0054] To guide the training of a neural network, a loss function needs to be defined (which measures the difference between the predicted and actual values). The loss function is defined as the MSE function:

[0055]

[0056] in It is the true label of the i-th sample. Let w be the predicted value of the i-th sample, and N be the number of samples. Training a neural network essentially involves adjusting the weights (w) and biases (b) to reduce the loss value, because reducing the loss value means reducing the gap between the predicted value and the true value.

[0057] The input to a forward neural network is structural parameters, and the output is an electromagnetic response. In this embodiment, |S 11 |-FNN includes an input layer, an output layer, and two hidden layers; AR-FNN includes an input layer, an output layer, and one hidden layer; Gain-FNN includes an input layer, an output layer, and one hidden layer.

[0058] The activation function of the output layer is the same as in formula (2), and the activation function of the hidden layer is the same as in formula (3). The inputs of the three sub-FNNs are all the same set of structural parameters. The outputs of the three sub-FNNs are Representing axis ratio, gain, and |S| respectively. 11 The loss function is the same as in formula (4). The purpose of training the FNN is to reduce the gap between the predicted electromagnetic response and the actual electromagnetic response.

[0059] Step 4: Utilize N normalized simulation data Three forward neural networks based on prior knowledge were trained separately to obtain three trained forward neural networks. During the training process of the forward neural network corresponding to each electromagnetic response attribute: the normalized simulation structure parameter vector was... The normalized simulated electromagnetic response vector is used as the input to this positive neural network. z-corresponding to the electromagnetic response k The dimension is used as the output of this forward neural network (when training the forward neural network corresponding to the axis ratio, it will be...). As output; when training the forward neural network corresponding to the gain, As output; during training |S 11 When the corresponding forward neural network is |, As output); and make the training error and test error meet the specified requirements, thereby obtaining the corresponding trained positive neural network.

[0060] Step 5: First, randomly generate N0 n-dimensional random structure parameter vectors x′. j Then, the random structure parameter vector x′ j The inputs are fed into m pre-trained forward neural networks, resulting in m z-values. k A dimensional random electromagnetic response vector; then m z-vectors k The electromagnetic response vectors of the two dimensions are concatenated into The electromagnetic response vector y′ of dimension j This yields N0 random data points (x′). j y′ j Where j = 1, 2, ..., N0, and N0 is the set number of random data.

[0061] The number of random data points N0 is chosen based on neural network debugging experience and is usually much larger than N. Specifically, MATLAB is used to randomly generate N0 sets of structural parameter combinations, which are input into the FNN to obtain N0 sets of output electromagnetic responses. These N0 sets of structural parameter combinations and N0 sets of output electromagnetic responses are then combined to form N0 random data points (x′). j y′ j In this embodiment, the number of random data items N0 is 50,000.

[0062] Step 6: Use N0 random data points (x′) j y′ j The inverse neural network in the prior knowledge-based neural network is trained to obtain a trained inverse neural network; during the training process of the inverse neural network, the random structure parameter vector y′ is... j The random structure parameter vector x′ serves as the input to this inverse neural network. jThe output of this inverse neural network is used to make the training error and test error meet the specified requirements, thus obtaining the trained inverse neural network.

[0063] Step 7: Determine the requirements that the lens antenna to be designed must meet. The 1D electromagnetic response vector y is input into the trained inverse neural network to obtain the n-dimensional target structural parameter vector x required for the design of the lens antenna.

[0064] The design goals and input values ​​of the INN in this embodiment are shown in Table 1. Since the training data for the INN is normalized, the input target electromagnetic response of the INN also needs to be normalized. The specific input target electromagnetic response values ​​are set by comprehensively considering the design goals, the electromagnetic response of the INN, and the characteristics of the lens antenna itself. It is also worth mentioning that for the INN, a set of input electromagnetic response parameters results in a unique output. However, for a specific design goal, there are usually many sets of input electromagnetic response parameters that meet the requirements. For the designer, only one set of output parameters is needed. For the input electromagnetic responses in Table 1, the obtained output structural parameters are P. target =[31.52,13.43,0.731,1.24,43.6].

[0065] Table 1 Comparison of Design Objectives, Input Values ​​of INN, and Simulated Values ​​of CST

[0066]

[0067] The structural parameters obtained using the method of this invention will be used for structural modeling and simulation in CST to verify the effectiveness of this invention.

[0068] The lens was processed using an Objet500 Connex3 3D ​​printer, with RGD837 VeroPureWhite as the printer support material, and the processed lens was tested.

[0069] Figure 4 The S-type lens antenna designed for this implementation 11 |Curve. Simulation and Testing|S 11 The values ​​are all below -16dB, indicating that the lens antenna has good matching in the 60-90GHz range.

[0070] Figure 5The gain-to-axis ratio curve of the lens antenna designed in this embodiment is shown. The horn lens antenna improves the gain by about 10dB compared to the feed antenna, and the gain curve of the lens antenna is relatively flat throughout the entire operating bandwidth (60-90GHz), achieving a 40% 3-dB gain bandwidth. The results show that the simulated axial ratio is below 2dB in the 60-90GHz range, consistent with the design target. For the test results, the low-frequency band test results are good, while the axial ratio in the 84-90GHz range is 3-4.5dB. The discrepancy between the simulated and tested results in the high-frequency range may be due to the instability of the electromagnetic properties of the dielectric material.

[0071] Figure 6 The radiation efficiency curve of the lens antenna designed for this embodiment is shown. The radiation efficiency curve shows that the antenna's radiation efficiency is greater than 67%, and reaches a maximum of 83%. This indicates that the energy of the lens antenna pair is effectively radiated.

[0072] Figure 7-9 The radiation patterns of the lens antenna designed for this embodiment at different frequencies show that the antenna is a left-hand circularly polarized antenna, and the sidelobe electrical average is below -20dB.

[0073] It should be noted that although the embodiments described above are illustrative, they are not intended to limit the invention. Therefore, the invention is not limited to the specific embodiments described above. Any other embodiments obtained by those skilled in the art under the guidance of this invention without departing from its principles are considered to be within the protection scope of this invention.

Claims

1. A multi-objective optimization method for lens antennas based on prior knowledge neural networks, characterized by: The steps include the following: Step 1: For a given lens antenna structure, collect data using full-wave simulation. simulation data ;in for 3D simulation structure parameter vector, for A 3D simulated electromagnetic response vector; Step 2: Normalize according to the maximum-minimum value principle. simulation data After normalization, we get Normalized simulation data ;in for A normalized simulation structure parameter vector of dimension 1. for The normalized simulated electromagnetic response vector of dimension ; Step 3, construct from A neural network based on prior knowledge, consisting of one forward neural network and one inverse neural network; the input of the forward neural network is the structural parameters, and the output is the electromagnetic response; the input of the inverse neural network is the electromagnetic response, and the output is the structural parameters. The mapping relationship from input to output of an inverse neural network can be represented by the following formula: ， in, This represents the output value of the input layer neurons. This represents the output value of the hidden layer neurons. f represents the output value of the output layer neuron. l (x) represents the activation function of the l-th layer. It represents the connection weight between the c-th neuron and the d-th neuron in adjacent layers. It is the bias of the d-th neuron. Represents the number of layers, C l This represents the number of neurons in the l-th layer; Step 4, utilize Normalized simulation data In neural networks based on prior knowledge Train a positive neural network to obtain The first trained positive neural network; in the second... During the training process of the forward neural network corresponding to each electromagnetic response property, the normalized simulation structure parameter vector is... The normalized simulated electromagnetic response vector is used as input to this positive neural network. Corresponding electromagnetic response The dimension is used as the output of the forward neural network, and the training error and test error are made to meet the specified requirements, thereby obtaining the corresponding trained forward neural network; Step 5: First generate randomly indivual dimensional random structure parameter vector Then the random structure parameter vector Enter them separately In each of the trained forward neural networks, the following are obtained: indivual A dimensional random electromagnetic response vector; then... indivual The electromagnetic response vectors of the two dimensions are concatenated into dimensional electromagnetic response vector ; from this we can obtain Random data ; Step 6, utilize Random data The inverse neural network in the prior knowledge-based neural network is trained to obtain a trained inverse neural network; during the training process of the inverse neural network, the random structure parameter vector is... The random structure parameter vector serves as the input to this inverse neural network. As the output of the inverse neural network, and by ensuring that the training error and test error meet the specified requirements, a trained inverse neural network is obtained. Step 7: Determine the requirements that the lens antenna to be designed must meet. dimensional electromagnetic response vector The input is fed into a trained inverse neural network to obtain the design parameters for the lens antenna to be designed. 1D target structure parameter vector ; The above , Indicates the amount of simulation data set; , The set number of random data points; ; The number of structural parameter attributes for the lens antenna to be designed; For the first The number of discrete points for each electromagnetic response property. , This represents the number of electromagnetic response properties that the lens antenna needs to satisfy.

2. The multi-objective optimization method for lens antennas based on prior knowledge neural networks according to claim 1, characterized in that, In step 2, For the structural parameter vector x i The first in The formula for normalizing the label values ​​of each structural parameter attribute is: electromagnetic response vector The first in The first electromagnetic response property The formula for normalizing the label values ​​is: In the formula, Indicates the first The first simulation data Normalized label values ​​of each structural parameter attribute. Indicates the first The first simulation data The label value of each structural parameter attribute. express The first simulation data The minimum label value of each structural parameter attribute. express The first simulation data The maximum label value of each structural parameter attribute , Indicates the number of structural parameter attributes; Indicates the first The simulation data The first electromagnetic response property A normalized label value, Indicates the first The first simulation data The first electromagnetic response property Each tag value, express The first simulation data The minimum label value of an electromagnetic response attribute. express The first simulation data The maximum label value of an electromagnetic response attribute. , Indicates the first The number of discrete points for each electromagnetic response property. , The number of electromagnetic response properties that the lens antenna needs to satisfy; , This indicates the amount of simulation data.

3. The multi-objective optimization method for lens antennas based on prior knowledge neural networks according to claim 1, characterized in that, In step 2, the normalized simulation data... Before normalization, the label values ​​of electromagnetic responses that change beyond a set threshold need to be simplified by step smoothing using a step smoothing method.

Images