Generative antenna optimization design method based on generative adversarial network
By using generative adversarial networks for optimization design, the problems of insufficient efficiency and accuracy in antenna optimization design are solved, enabling the efficient generation of multiple antenna models that meet the requirements and addressing the issue of long dataset acquisition time.
Patent Information
- Application Number
- CN202410714501.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-04
- Publication Date
- 2026-02-27
- Estimated Expiration
- 2044-06-04
AI Technical Summary
Existing technologies for antenna optimization design are inefficient and inaccurate, and traditional methods are time-consuming and make it difficult to obtain multiple antenna models that meet the requirements at once.
A generative adversarial network (GAN) is used to construct an antenna optimization design method. This method involves establishing an initial antenna model, uniform sampling, obtaining performance parameters through simulation software, constructing a training set, training the GAN, and iteratively generating multiple antenna models that meet the requirements.
It improves the efficiency and accuracy of antenna optimization, and can generate multiple antenna models that meet the requirements at one time after multiple iterations, solving the problem of long dataset acquisition time and avoiding model collapse.
Smart Images
Figure CN118627148B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The embodiment of the present disclosure relates to the technical field of antenna design, in particular to a generative antenna optimization design method based on a generative adversarial network. BACKGROUND
[0002] At present, due to the complex communication environment and the increasingly high requirements of people on the performance of the antenna, the antenna usually has a complex topology and structure, and has a large number of size design parameters, which leads to that the electromagnetic problem in the real world cannot be analytically calculated. The traditional method is to fine-tune the geometric or material parameters of the antenna structure by using simulation software to improve the performance. However, for a complex antenna with many geometric parameters, this process will be very time-consuming.
[0003] In the antenna optimization design, machine learning and neural network methods can be used to reduce the time cost. However, a network that can accurately learn the mapping relationship between the antenna structure and the performance parameters often needs a large amount of data set to support its training, and obtaining a large amount of data set from the simulation software also takes a lot of time. Moreover, the general optimization method currently used often stops optimizing when the antenna structure that meets the requirements is obtained, and cannot obtain multiple antenna models that meet the requirements at one time, which limits the design efficiency of the antenna.
[0004] Therefore, there is an urgent need for a high-efficiency and high-precision generative antenna optimization design method based on a generative adversarial network. SUMMARY
[0005] Therefore, the embodiment of the present disclosure provides a generative antenna optimization design method based on a generative adversarial network, which at least partially solves the problem of poor optimization efficiency and precision in the prior art.
[0006] The embodiment of the present disclosure provides a generative antenna optimization design method based on a generative adversarial network, which comprises:
[0007] Step 1, establishing an initial antenna model and uniformly sampling the structure parameters thereof, and using simulation software to obtain the performance parameters corresponding to the structure parameters, and combining the structure parameters and the performance parameters to form an initial data set;
[0008] Step 2, dividing the initial data set into two types of samples with good and poor performance by taking the median of the preset antenna performance indicators as a dividing point, and using the good type of samples after normalization as a training set;
[0009] Step 3, constructing a generative adversarial network and training by using the training set;
[0010] Step 4, a new structure parameter is generated by using the trained generative adversarial network, a new performance parameter is obtained by using simulation software, a second data set is formed by combining the new structure parameter and the new performance parameter, and the second data set is divided into two categories of samples with good and poor performance according to a preset antenna performance index median value, the optimal class samples of the last time are combined with the optimal class samples of the current time, and the new training set is obtained by dividing again according to the preset antenna performance index median value to train the new generative adversarial network;
[0011] Step 5, a plurality of new antenna models are generated by using the new generative adversarial network and simulated, and it is judged whether there is a required antenna model, if yes, the structure parameter and the performance parameter corresponding to the antenna model are output, if not, step 4 is repeated until the requirement is met, and the trained generative adversarial network is obtained.
[0012] Step 6, random noise is input into the trained generative adversarial network, and a normalized value of the antenna structure parameter is obtained.
[0013] According to a specific implementation mode of the embodiment of the present disclosure, the performance parameter includes at least one of return loss, reflection coefficient, gain, antenna pattern and radiation efficiency of the antenna.
[0014] According to a specific implementation mode of the embodiment of the present disclosure, the step 2 specifically includes:
[0015] The performance parameter in the initial data set is transformed, the antenna performance index corresponding to each structure parameter is calculated according to the transformation result, the initial data set is divided into two categories of samples with good and poor performance according to the median value, and the structure parameter in the optimal class sample is normalized and used as the training set.
[0016] According to a specific implementation mode of the embodiment of the present disclosure, the normalized expression is
[0017]
[0018] Wherein, x represents the antenna structure parameter to be normalized, x min represents the minimum value in the design range of the antenna structure parameter, and x max represents the maximum value in the design range of the antenna structure parameter.
[0019] The generative antenna optimization design scheme based on the generative adversarial network in the embodiment of the disclosure includes: step 1, an initial antenna model is established, and uniform sampling is performed on the structure parameters thereof, and a simulation software is used to obtain performance parameters corresponding to the structure parameters, and the structure parameters and the performance parameters are combined to form an initial data set; step 2, the initial data set is divided into two types of samples with performance advantages and disadvantages with a median of preset antenna performance indicators as a dividing point, and the superior class samples are normalized and used as a training set; step 3, a generative adversarial network is constructed and trained by using the training set; step 4, a trained generative adversarial network is used to generate new structure parameters, and a simulation software is used to obtain new performance parameters, a second data set is formed by combining the new structure parameters and the new performance parameters, and the second data set is divided into two types of samples with performance advantages and disadvantages with the median of the preset antenna performance indicators as a dividing point, the superior class samples of the last time are combined with the superior class samples of the current time, and the new training set is obtained by dividing again according to the median of the preset antenna performance indicators, so as to train a new generative adversarial network; step 5, a plurality of new antenna models are generated by using the new generative adversarial network and simulated, and it is judged whether there is an antenna model meeting the requirements, if yes, the structure parameters and the performance parameters corresponding to the antenna model are output, if not, step 4 is repeated until the trained generative adversarial network is obtained; and step 6, random noise is input into the trained generative adversarial network to obtain a normalized value of the antenna structure parameter.
[0020] The beneficial effects of the embodiment of the disclosure are that: by using the scheme of the disclosure, a small amount of antenna samples can be used to optimize the antenna, the problem of too long data set acquisition time in the neural network optimization antenna problem can be solved, and the network can be used to effectively avoid the mode collapse problem. After multiple iterations, the generator can generate multiple antennas meeting the requirements at one time, and the optimization efficiency and accuracy are improved. BRIEF DESCRIPTION OF DRAWINGS
[0021] In order to more clearly illustrate the technical scheme of the embodiment of the disclosure, the drawings needed in the embodiment will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the disclosure, and other drawings can be obtained by those skilled in the art without creating labor.
[0022] Figure 1 A flowchart of a generative antenna optimization design method based on a generative adversarial network provided by the embodiment of the disclosure is shown in the figure.
[0023] Figure 2 A structure diagram of an ultra-wideband antenna provided by the embodiment of the disclosure is shown in the figure.
[0024] Figure 3 A working principle diagram of a double-discriminator neural network provided by the embodiment of the disclosure is shown in the figure.
[0025] Figure 4 A curve of change of the median of the antenna performance parameter in the data set with the increase of the number of iterations is provided for the embodiment of the present disclosure.
[0026] Figure 5 An S11 curve of the finally optimized antenna is provided for the embodiment of the present disclosure. DETAILED DESCRIPTION
[0027] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.
[0028] The embodiments of the present disclosure are described below by way of specific examples. Those of ordinary skill in the art can easily understand other advantages and effects of the present disclosure from the content disclosed in the specification. It is obvious that the described embodiments are only a part of the embodiments of the present disclosure, but not all the embodiments. The present disclosure can also be implemented or applied by other different specific embodiments, and the details in the specification can be modified or changed based on different views and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present disclosure.
[0029] It should be noted that various aspects of the embodiments described below are within the scope of the appended claims. It should be apparent that the aspects described herein can be embodied in a wide variety of forms and that any specific structure and / or function described herein is merely illustrative. Based on the teachings provided herein one skilled in the art should appreciate that an aspect described herein can be implemented independently of any other aspects and that an aspect described herein can be implemented both as any number of software and / or hardware structures and as any number of processes and / or operations. For example, an aspect can be implemented as a software program running on hardware that can process information such as, for example, a general purpose computer store in memory and execute under the control of one or more software programs.
[0030] It should also be noted that the figures provided in the following embodiments are only to illustrate the basic concept of the present disclosure in a schematic manner, and only the components related to the present disclosure are shown in the figures, not drawn according to the number, shape and size of the components in actual implementation, and the shape, number and proportion of each component in actual implementation can be arbitrarily changed, and the layout pattern of the components can also be more complex.
[0031] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, one skilled in the art will understand that the described aspects can be practiced without these specific details.
[0032] The embodiment of the present disclosure provides a generative antenna optimization design method based on a generative adversarial network, which can be applied to an antenna design process in a communication scenario.
[0033] Referring to Figure 1 , a flowchart of a generative antenna optimization design method based on a generative adversarial network is provided in the embodiment of the present disclosure. As shown in the figure, Figure 1 , the method mainly includes the following steps:
[0034] Step 1, an initial antenna model is established, and uniform sampling is performed on the structure parameters thereof, and a simulation software is used to obtain performance parameters corresponding to the structure parameters, and the structure parameters and the performance parameters are combined to form an initial data set;
[0035] Optionally, the performance parameters include at least one of return loss, reflection coefficient, gain, antenna pattern and radiation efficiency of the antenna.
[0036] In a specific implementation, an initial antenna model is established, 100-point uniform sampling is performed in a design range of antenna structure parameters to be optimized, and a simulation software is used to obtain performance parameters of 100 antennas, so as to construct an initial data set containing 100 antenna structure-performance data pairs. In this example, as shown in the figure, Figure 2 , the antenna is built on an L0xW0 FR4 epoxy substrate, the relative dielectric constant (εr) is 4.4, the tangent loss is 0.02, the height (H) is 1.6 mm. The defects of the ground plane edge and the square radiation surface sheet are realized in the form of a spline curve. The antenna parameter vector to be optimized is Y = [y, r, W, L] T , wherein y = [y1, y2, …, yn] T , r = [r1, r2, …, rm] T , W = [W1, W2] T , L = [L1, L2] T , n = 11, and m = 8. The value range is shown in Table 1. Specifically, in this example, the parameter to be optimized is the return loss (S11) of the antenna, and it is expected to achieve super wideband performance in the frequency range of 3-10 GHz, that is, the antenna satisfies S11 <-10 dB in the frequency range of 3-10 GHz. The simulation frequency range is 1-10 GHz, and the sampling points are 100.
[0037] Parameter name y r [CD AT W1] [CD AT W2] <![CDATA[L1]]> <L2> Range (4,14) (3.5,5) (17,23) (2,4) (26,34) (10,14)
[0038] Table 1
[0039] Step 2, the initial data set is divided into two types of samples with good and poor performance with a preset median of antenna performance indicators as a dividing point, and the good type samples are normalized and used as a training set;
[0040] On the basis of the above embodiment, the step 2 specifically comprises:
[0041] Transform the performance parameters in the initial data set, calculate the antenna performance index corresponding to each structure parameter according to the transformation result, divide the initial data set into two categories of samples with good and poor performance with the median, and normalize the structure parameters in the superior sample to obtain a training set.
[0042] Further, the normalized expression is
[0043]
[0044] Wherein, x represents the antenna structure parameter to be normalized, x min represents the minimum value of the antenna structure parameter in the design range, x max represents the maximum value of the antenna structure parameter in the design range.
[0045] In specific implementation, the initial data set is divided into two categories of samples with good and poor performance with the median of the self-defined antenna performance index as the dividing point, the antenna samples in the category of "1" are normalized to obtain a training set of the network. Let the value of S11 of 100 frequency points obtained by simulation be ci (i = 1, …, 100), and the S11 value is transformed as follows:
[0046]
[0047] In this example, the self-defined antenna performance index is:
[0048] SCORE = ∑C i
[0049] After calculating the performance scores of 100 antennas, the data set is divided into the category of "1" with higher scores and the category of "0" with the rest. In the process of network training, the data input into the network only contains the normalized antenna structure data, and the antenna performance parameters obtained by simulation are only used when the data set is divided. Therefore, the structure parameters of the "1" category sample data are normalized:
[0050]
[0051] The antenna structure data set subjected to the normalization processing is used as the training sample of the network.
[0052] Step 3, constructing a generative adversarial network and training by using the training set;
[0053] In implementation, a generative adversarial network model is constructed and trained. In this example, a double-discriminator generative adversarial network is used to increase the diversity of the small-sample GAN network generated data, and its working principle is as shown in Figure 3 The generator head is composed of two fully connected layers and a LeakyRelu function to extract data features, then enters a batch normalization layer to normalize the features, and then inputs a fully connected layer and a batch normalization layer, and uses a sigmoid function to limit the output result to the range of [0, 1]. The discriminator is composed of four fully connected layers, the activation functions of the first three linear layers are ReLU functions, and the activation function of the last layer is a Softplus function, two Dropout layers are added in the middle to improve the generalization performance of the discriminator. The output is a positive real number, not a probability between [0, 1]. The three-player game process of the generator and the discriminator is as follows:
[0054]
[0055] Step 4, use the trained generative adversarial network to generate new structure parameters, and use the simulation software to obtain new performance parameters, combine the new structure parameters and the new performance parameters to form a second data set, and divide the second data set into two classes of samples with good and poor performance according to the median of the preset antenna performance index. The good class samples of the last time are combined with the good class samples of the current time, and are divided again according to the median of the preset antenna performance index to obtain a new training set to train a new generative adversarial network.
[0056] In implementation, 100 new antenna structures are generated using the trained generative adversarial network, simulated using simulation software, and divided into new "1" class and new "0" class according to the method of step 2. The "1" class samples of the last time are combined with the "1" class samples of the current time, and are divided again according to the median to obtain a new training set to train a new generative adversarial network model.
[0057] Step 5, use the new generative adversarial network to generate multiple new antenna models and simulate to determine whether there is a required antenna model, if so, output the structure parameters and performance parameters corresponding to the antenna model, if not, repeat step 4 until the requirements are met, and obtain a trained generative adversarial network.
[0058] In implementation, 100 new antenna structures are generated using the new generative adversarial network model and simulated, and it is determined whether there is a required antenna, if not, repeat step 4 until the requirements are met.
[0059] As Figure 4The figure shows the change of the median of the antenna performance indicator SCORE in the training set during the 4 iteration processes, where 0 corresponds to the median of the antenna data set obtained by the first uniform sampling. As can be seen, the performance of the antenna generated by the GAN network gradually improves, and in the third GAN network generated data, an antenna structure with super wideband performance requirements of 3-10GHz has been generated, and the S11 curve of the antenna is as shown in the figure. Figure 5 The optimized antenna structure parameters are y = [10.95, 9.38, 5.75, 6.75, 4.93, 6.29, 13.19, 9.24, 9.47, 8.66, 10.45] T r = [3.84, 4.63, 3.9, 4.52, 4.5, 4.49, 4.65, 4.11] T W = [17.29, 2.25] T L = [28.77, 13.57] T , unit: mm.
[0060] Step 6: input random noise into the trained generative adversarial network to obtain the normalized value of the antenna structure parameter.
[0061] In actual implementation, in the practical application process of the trained generative adversarial network, the input of the trained generative adversarial network is random noise, and the output is the normalized value of the antenna structure parameter, and multiple antenna structure parameters meeting the requirements can be obtained at a time.
[0062] The generative antenna optimization design method based on the generative adversarial network provided in the embodiment can solve the problem of too long data set acquisition time in neural network optimization antenna problem by using a small amount of samples to train the network, and can effectively avoid the mode collapse problem by using the network. After multiple iterations, the generator can generate multiple antennas meeting the requirements at a time, improving the optimization efficiency and accuracy.
[0063] It should be understood that parts of the present disclosure can be realized by hardware, software, firmware or a combination thereof.
[0064] The above is only a specific implementation of the present disclosure, but the protection scope of the present disclosure is not limited thereto, and any changes or replacements within the technical range disclosed in the present disclosure can be easily thought of by those skilled in the art, which should be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.
Claims
1. A generative antenna optimization design method based on a generative adversarial network, characterized in that, The method comprises the following steps: Step 1, an initial antenna model is established, and the structural parameters thereof are uniformly sampled, and the performance parameters corresponding to the structural parameters are obtained by using simulation software, and the initial data set is formed by combining the structural parameters and the performance parameters; Step 2, the initial data set is divided into two types of samples with good and poor performance by taking the median of the preset antenna performance index as the dividing point, and the good type samples are normalized and used as the training set; Step 3, a generative adversarial network is constructed and trained by using the training set; Step 4, the trained generative adversarial network is used to generate new structural parameters, and the new performance parameters are obtained by using the simulation software, the second data set is formed by combining the new structural parameters and the new performance parameters, the second data set is divided into two types of samples with good and poor performance by taking the median of the preset antenna performance index as the dividing point, the good type samples of the last time are combined with the good type samples of the current time, and the new training set is obtained by dividing again according to the median of the preset antenna performance index, so as to train the new generative adversarial network; Step 5, a plurality of new antenna models are generated by using the new generative adversarial network and simulated, whether there is a required antenna model is judged, if yes, the structural parameters and the performance parameters corresponding to the antenna model are output, if not, step 4 is repeated until the required training generative adversarial network is obtained; Step 6, random noise is input into the trained generative adversarial network to obtain the normalized value of the antenna structural parameter.
2. The method of claim 1, wherein The performance parameters include at least one of return loss, reflection coefficient, gain, antenna pattern and radiation efficiency of the antenna.
3. The method of claim 1, wherein The step 2 specifically comprises: The performance parameters in the initial data set are transformed, the antenna performance index corresponding to each structural parameter is calculated according to the transformation result, the initial data set is divided into two types of samples with good and poor performance by taking the median as the dividing point, and the structural parameters in the good type samples are normalized and used as the training set.
4. The method of claim 2, wherein The normalized expression is where x represents the antenna structure parameter to be normalized, x min represents the minimum value within the design range of the antenna structure parameter, x max represents the maximum value within the design range of the antenna structure parameter.