An antenna optimization method based on multi-branch conditional generative adversarial network
By using multi-branch conditional generative adversarial networks, monotonicity analysis, and spectral clustering algorithms to filter low-quality data and training a multi-branch random forest model, the problem of high data acquisition costs in antenna optimization by machine learning models is solved, achieving efficient and accurate antenna optimization.
Patent Information
- Application Number
- CN202411560425.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-04
- Publication Date
- 2025-12-09
- Estimated Expiration
- 2044-11-04
AI Technical Summary
Existing technologies using machine learning models to optimize antennas are costly and time-consuming in terms of data acquisition, especially when dealing with complex antenna models where computational costs can be even higher.
A multi-branch conditional generative adversarial network is adopted. By collecting an initial dataset and mapping it to a latent representation using an encoder, a multi-branch conditional generative adversarial network is established. By combining monotonicity analysis and spectral clustering algorithms to filter low-quality data, a multi-branch random forest model is trained as a surrogate model to optimize the antenna.
It reduces data acquisition costs, shortens antenna optimization time, and establishes an accurate and efficient surrogate model that can quickly generate high-fidelity data to optimize antenna design.
Smart Images
Figure CN119578213B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of antenna optimization design, and particularly relates to an antenna optimization method based on a multi-branch conditional generative adversarial network. BACKGROUND
[0002] Traditional antenna design methods usually need to use full-wave electromagnetic simulators to optimize antenna size parameters through a large number of repeated parameter scans. Although this method can provide accurate electromagnetic responses, this process is not only computationally intensive but also time-consuming.
[0003] Machine learning methods can capture nonlinear relationships in data sets and accurately predict new data. In recent years, machine learning-based antenna optimization methods have been widely used. By using computationally inexpensive machine learning models to replace expensive electromagnetic simulations, the computational cost can be greatly reduced and the optimization process can be accelerated. However, in general, a large amount of training data is required to establish an accurate machine learning model.
[0004] Most current research focuses on how to improve the algorithm to reduce the use of antenna simulation data. However, this method still essentially uses simulators to increase new samples, and when dealing with complex antenna models, the cost can be even more expensive. SUMMARY
[0005] The purpose of the present application is to provide an antenna optimization method based on a multi-branch conditional generative adversarial network, which aims to solve the problem of high cost and time-consuming data acquisition when using machine learning models to optimize antennas.
[0006] To achieve the above purpose, the present application provides an antenna optimization method based on a multi-branch conditional generative adversarial network, comprising the following steps:
[0007] Step 1: Collect an initial data set of the antenna;
[0008] Step 2: According to S 21 The optimal frequency of the parameter divides the data set into multiple groups;
[0009] Step 3: Use an encoder to map the data to a latent representation;
[0010] Step 4: Establish a multi-branch conditional generative adversarial network and use the latent representation to generate the size parameters of the antenna and the corresponding response parameters;
[0011] Step 5: Use monotonicity analysis and spectral clustering algorithms to filter out low-quality data generated;
[0012] Step 6: Based on the remaining high-fidelity data, a multi-branch random forest model is trained as a surrogate model to optimize the antenna.
[0013] Optionally, the execution process of step 1 is specifically using a simulator to obtain a small amount of simulation data set of the antenna as an initial data set, which contains the size parameters of the antenna and the corresponding S 11 and S 21 response data.
[0014] Optionally, in the grouping process of step 2, the S 21 data in the initial data set is divided into the same group if the optimal frequency point is the same. Because the size parameters in the same group correspond to a group of S 11 and S 21 response at the same time, for the data containing S 11 , it is divided by selecting the S 21 data corresponding to the size parameters in each group of S 11 data.
[0015] Optionally, in step 3, an encoder is used to map the data into a latent representation instead of the random noise used in the traditional generative adversarial network.
[0016] Optionally, the number of branches of the multi-branch conditional generative adversarial network in step 4 is determined according to the number of divided data sets. At the same time, when generating S 21 response, the optimal frequency of S 21 is used as the conditional information to guide the direction of data generation; similarly, when generating S 11 response, the added conditional information is also the optimal frequency, but it has no other physical meaning, only represents its category.
[0017] Optionally, the formula for judging the number of monotonic changes in step 5 is as follows:
[0018] [f(x)-f(x-1)]×[f(x+1)-f(x)]<0
[0019] Where x is a certain frequency point, and f(x) is the S response value at the frequency point;
[0020] In the operation process of the spectral clustering algorithm, the real data and the generated data are clustered and analyzed.
[0021] Optionally, in the process of training the multi-branch random forest model in step 6, a reverse multi-branch random forest model is established for S 21 data, the input is S 21 response, and the output is the size parameters of the antenna; for S 11The data establishes a forward multi-branch random forest model, the input and output are opposite to the reverse model; and in the training process, the optimal model is selected by increasing ten data each time, the maximum iteration number is set to 30, and finally the optimal model is used to optimize the antenna.
[0022] The application provides an antenna optimization method based on a multi-branch conditional generative adversarial network. 21 The optimal frequency of the parameter divides the collected data set into multiple different groups; then, the data is mapped into a latent representation with prior knowledge by using an encoder; then, a conditional generative adversarial network with a corresponding number of branches is established according to the number of group classification of the data set, and the optimal frequency is added as the conditional information of the conditional generative adversarial network to guide the generation direction of the data; the established multi-branch conditional generative adversarial network generates the antenna size parameter and the corresponding S parameter response of the expected category by using the latent representation mapped by the encoder; then, a double filtering strategy composed of monotonicity analysis and spectral clustering algorithm is used to filter out the low-quality data generated; finally, based on the high-fidelity data reserved, a well-trained multi-branch random forest model is obtained to optimize the antenna. It has been verified that the proxy model established by the application is not only accurate, but also reduces the cost of collecting data and reduces the optimization time of the antenna. BRIEF DESCRIPTION OF DRAWINGS
[0023] In order to more clearly illustrate the technical solutions of the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.
[0024] Figure 1 It is a step flowchart of the antenna optimization method based on the multi-branch conditional generative adversarial network of the application.
[0025] Figure 2 It is a schematic diagram of the establishment process of the multi-branch conditional generative adversarial network of the application.
[0026] Figure 3 It is a flowchart of the operation of the forward and reverse multi-branch random forest of the application.
[0027] Figure 4 It is a structure schematic diagram of the wideband stacked patch antenna array of the specific embodiment of the application.
[0028] Figure 5 It is a comparison schematic diagram of the simulation S parameter corresponding to the size parameter in the generated data of the specific embodiment of the application at different optimal frequencies and the generated S parameter.
[0029] Figure 6 This is a specific embodiment of the invention where S is input at different optimal frequencies. 21 Curves and simulated S 21 Comparison of curves, and prediction of S 11 Curves and simulated S 11 A diagram showing the comparison of the curves. Detailed Implementation
[0030] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0031] Please see Figure 1 This invention provides an antenna optimization method based on a multi-branch conditional generative adversarial network, comprising the following steps:
[0032] S1: Collect the initial dataset for the antenna;
[0033] S2: According to S 21 The optimal frequency of the parameters is used to divide the dataset into multiple groups;
[0034] S3: Use an encoder to map data into a latent representation;
[0035] S4: Establish a multi-branch conditional generative adversarial network and use latent representations to generate antenna size parameters and corresponding response parameters;
[0036] S5: Use monotonicity analysis and spectral clustering algorithms to filter out low-quality generated data;
[0037] S6: Based on the retained high-fidelity data, train a multi-branch random forest model as a surrogate model to optimize the antenna.
[0038] This invention employs a multi-branch conditional generative adversarial network with prior knowledge to efficiently optimize antenna design. The multi-branch conditional generative adversarial network trained with a small amount of data generates a large amount of high-fidelity data, and this data is used to efficiently build a well-trained proxy model.
[0039] Specifically, the antenna dataset acquired in step S1 contains only a small amount of data, and it exists in the form of data pairs. The initial dataset includes the antenna's size parameters and the corresponding S... 11 and S 21 Response data.
[0040] During the grouping process in step S2, the initial dataset S is divided into groups.21 The data with the same optimal frequency point are divided into the same group, because the same group size parameter corresponds to a group of S 11 21 Response, therefore, for the data containing S 11 , is divided by selecting the S 21 data corresponding to the size parameter in each group S 11 .
[0041] In step S3, the data is first mapped to a latent representation using an encoder instead of random noise used in traditional GAN to provide prior knowledge and reduce the amount of training data.
[0042] In step S4, a multi-branch conditional GAN with a corresponding number of branches is established according to the number of divided data sets to generate each category of data, improving the quality of generated samples, as shown in Figure 2 . At the same time, when generating S 21 response, the optimal frequency of S 21 is used as conditional information to guide the direction of data generation; similarly, when generating S 11 response, the added conditional information is also the optimal frequency, but has no other physical meaning, only representing its category.
[0043] In step S5, a double filtering strategy containing monotonicity analysis and spectral clustering is adopted, in which the formula for judging the number of monotonicity changes is as follows:
[0044] [f(x)-f(x-1)]×[f(x+1)-f(x)]<0
[0045] Where x is a certain frequency point and f(x) is the S response value at that frequency point. In spectral clustering, real data and generated data are analyzed to filter out low-quality data.
[0046] In step S6, a reverse multi-branch random forest model is established for S 21 data, with S 21 response as input and size parameters as output; a forward multi-branch random forest model is established for S 11 data, with the input and output opposite to the reverse model, as shown in Figure 3 . The optimal model is selected by increasing ten data each time, and the maximum number of iterations is set to 30. Finally, the optimal model is used to optimize the antenna.
[0047] Further, the following further explanation is made in combination with the drawings and examples:
[0048] The present application takes the decoupling problem of a wideband stacked patch antenna array as an example to illustrate the effectiveness of the proposed method, uses the high-fidelity data generated by the conditional generative adversarial network to obtain a well-trained random forest model, and uses the model to optimize the size parameters of the antenna. In order to solve the coupling problem, the data collected by the embodiment when collecting data is the size parameters of the antenna and the corresponding S 11 and S 21 response data. As Figure 4 shown, the structure diagram of a wideband stacked patch antenna array is given, and Table 1 gives the size range of the collected data.
[0049] Table 1 Size range of wideband stacked patch antenna array collected data
[0050]
[0051] wherein L is the length of the branch, W is the width of the branch, Ang is the angle of the arc-shaped part of the branch, R is the radius of the reflector, S Y is the patch spacing along the Y-axis direction, and S X is the patch spacing along the X-axis direction.
[0052] The steps proposed by the present application are combined for design:
[0053] Step 1: Construct the wideband stacked patch antenna array model to be optimized, and use the electromagnetic simulation software high frequency structure simulator (HFSS) to simulate according to the above size range to collect its data in the working frequency band 1.80-2.20GHz. The present application collects a total of 175 data, and the frequency step is set to 0.04 when saving the data.
[0054] Step 2: First, divide the data set into multiple different sub-data sets according to the optimal frequency of the S 21 response data in the data set, and then select the S 11 response corresponding to the size parameters in each sub-data set to divide the S 11 response data set. Specifically, the optimal frequency can be divided into: 1.96, 2.00, 2.04, 2.08 and 2.12GHz, and each sub-data set contains 18, 36, 50, 52 and 19 data. They are divided into training set and test set according to the ratio of 8:2, and are standardized pretreated.
[0055] Step 3: According to the number of sub-data sets, a five-branch conditional generative adversarial network is established, each branch is a separate conditional generative adversarial network, which generates the corresponding antenna size data and the corresponding response data. The conditional information added in the network is S21 the optimal frequency point, only for S 11 The condition information only represents the category of the data. The encoder is used to map the data of each branch to the latent representation, and the conditional generative adversarial network generates the data of the corresponding category based on the latent representation, which contains the size parameter and the corresponding S parameter response.
[0056] Step 4: First, the monotonicity of the response data in the generated data is judged to filter out some generated random data. For the antenna, the monotonicity of the S parameter generally changes once or twice, so for the S 11 data, we will filter out the data whose monotonicity changes more than twice, and for the S 21 data, we will filter out the data whose monotonicity changes more than once. The spectral clustering algorithm is used to cluster and label the real data and the generated data, and the number of clusters is set to two. The data with different labels from the real data are deleted to filter out the generated low-fidelity data. As shown in Figure 5 , the size parameter in the generated data is simulated by HFSS to obtain the S parameter result, which is compared with the corresponding generated response in the generated data. It can be seen that the simulation result of the S parameter and the generated result only have slight differences in numerical value, and the curve trend is basically the same, indicating that the generated size data and the response data are matched, and the generated data has high fidelity.
[0057] Step 5: Based on the generated high-fidelity data of each category, a reverse five-branch random forest model is established for the S 21 data, and a forward five-branch random forest model is established for the S 11 data. Each branch is trained with the corresponding generated data. During the training, ten data are added each time, and the process is repeated 30 times. The model with the smallest root mean squared error (RMSE) is selected as the optimal model. As shown in Table 2, the smallest RMSE of each branch and the corresponding iteration number are shown. Among the maximum iteration number, the RMSE of each branch of the S 21 and S 11 is less than 0.18 and 0.12, respectively. The RMSE value of branch 1 can be as low as 0.0887 and 0.0251, indicating that the network trained based on the generated data has good prediction ability.
[0058] Table 2 RMSE and iteration number of each branch
[0059]
[0060]
[0061] Step 6: Based on the actual needs of the project, S 11 and S 21 Set an optimization objective and use the trained model to optimize the size parameters. First, set the desired S... 21 The response vector is input into the corresponding branch of the inverse multi-branch random forest model to obtain the corresponding size parameter. This size parameter is then input into S. 11 The predicted S is obtained by using the corresponding branch in the positive multi-branch random forest model. 11 The response data, if both optimization objectives are met, then HFSS is used to simulate each set of size parameters to determine whether the actual results achieve the optimization objectives. Specifically, the optimization objectives designed in this embodiment are: 1. Within 1.96-2.12 GHz, S 11 <-10dB, and the bandwidth less than -10dB is greater than 300MHz. 2. In S 21 At the optimal frequency, S 21 <-35dB.
[0062] Based on this optimization objective, different S 21 The response vector is first input into the corresponding branch of the multi-branch random forest model to obtain multiple sets of responses satisfying S. 21 The dimensional parameters of the design specifications are shown in Table 3. The prediction results of the forward multi-branch random forest model for these dimensional parameters are as follows: Figure 6 As shown, from branch 1 to branch 5, S 11 The frequency range with a value less than -10dB is 1.84-2.17GHz, corresponding to a bandwidth of 330MHz, which meets the first optimization requirement; for S 21 Regarding the parameters, the input values at each optimal frequency point are -40.0045, -40.8981, -40.0789, -40.8860, and -40.4634 dB, while the corresponding simulation results are -37.3326, -38.2373, -40.1692, -37.6771, and -38.5966 dB. Although there are still some deviations in the numerical values, it can be seen that their curve trends are basically consistent, and they also satisfy the second optimization objective. In particular, the S of the third branch... 21 The simulation value is -40.1692dB, which means that the optimal antenna isolation design has been achieved. Therefore, the value optimized by the third branch is selected as the optimal antenna size parameter.
[0063] Table 3 shows the optimized size parameters for each branch.
[0064]
[0065] Further, in order to prove that the present application can reduce the optimization time of the antenna, based on the number of generated samples required to train each branch model and the time required to simulate a sample, the total time required to obtain these data using HFSS simulation is calculated, as shown in Table 4. It can be seen that branch 5 requires the most time of 54h, and the total time required is 214h. But the time spent collecting the entire data set is 34.80h, saving 83.74% of the time. Specifically for each branch, they use 18, 36, 50, 52 and 19 data respectively, and the corresponding simulation time is 3.6, 7.2, 10, 10.4 and 3.8h, respectively, saving 90%, 85%, 73.68%, 72.63% and 92.59% of the time. The above experimental results show that the proposed antenna optimization method is accurate and efficient, not only can solve the problem of data scarcity caused by high data acquisition cost, but also can reduce the optimization time of the antenna.
[0066] Table 4 Number of generated samples and time required for simulation
[0067]
[0068] The above disclosure is only a preferred embodiment of the present application, of course, cannot be limited by this to limit the scope of the present application, those skilled in the art can understand that the above-mentioned embodiment can realize all or part of the process, and the equivalent changes made according to the claims of the present application still belong to the scope covered by the present application.
Claims
1. A method for antenna optimization based on multi-branch conditional generative adversarial network, characterized in that, The method comprises the following steps: Step 1: collecting an initial data set of the antenna; Step 2: According to S 21 The optimal frequency partition dataset of parameters is multiple groups; Step 3: mapping the data into a latent representation using an encoder; Step 4: establishing a multi-branch conditional generative adversarial network and using the latent representation to generate size parameters and corresponding response parameters of the antenna; Step 5: filtering out low-quality generated data using monotonicity analysis and spectral clustering algorithm; Step 6: training a multi-branch random forest model as a surrogate model to optimize the antenna based on the high-fidelity data retained; In step 6, the process of training the multi-branch random forest model, for S 21 data, a reverse multi-branch random forest model is built, with S 21 as input and the size parameters of the antenna as output; for S 11 data, a forward multi-branch random forest model is built, with the input and output being opposite to the reverse model; specifically, the desired S 21 response vector is input into the corresponding branch of the reverse multi-branch random forest model to obtain the corresponding size parameter, and then the size parameter is input into the S 11 forward multi-branch random forest model corresponding to the branch to obtain the predicted S 11 response data; During the training process, the optimal model is selected by increasing ten data each time, the maximum number of iterations is set to 30, and finally the optimal model is used to optimize the antenna.
2. The antenna optimization method based on the multi-branch conditional generative adversarial network according to claim 1, wherein The execution process of step 1, in particular, uses a simulator to obtain a small simulation data set of the antenna as an initial data set, which contains the size parameters of the antenna and the corresponding S 11 and S 21 response data.
3. The antenna optimization method based on the multi-branch conditional generative adversarial network according to claim 2, wherein In the grouping process of Step 2, the initial data set S 21 is divided into groups according to the optimal frequency points. The data with the same optimal frequency points are divided into the same group. For the data S 11 , the division is made by selecting the data S 21 with the size parameter corresponding to the size parameter of each group S 11 .
4. The antenna optimization method based on the multi-branch conditional generative adversarial network according to claim 3, wherein The number of branches of the multi-branch conditional generative adversarial network in step 4 is determined according to the number of the divided data sets; meanwhile, in the generation of S 21 , the optimal frequency of S 21 is used as the conditional information to guide the direction of data generation; in the generation of S 11 , the added conditional information is also the optimal frequency of S 21 , and the corresponding conditional information has no other physical meaning and only represents the category.
5. The antenna optimization method based on the multi-branch conditional generative adversarial network according to claim 4, wherein The formula for determining the number of monotonicity changes in step 5 is as follows: ； wherein, is a certain frequency point, is the S response value at the frequency point; During the operation process of the spectral clustering algorithm, the real data and the generated data are subjected to clustering analysis.