Optimization design method of wideband dual-polarized antenna based on adaptive evolutionary neural network

Abstract

The application relates to the technical field of antenna design, in particular to a wideband dual-polarized antenna optimization design method based on an adaptive evolutionary neural network, which comprises the following steps: an initial antenna model is established, a full-wave simulation is performed on the initial antenna model by using simulation software, and an initial data set is collected; a first network and a second network in the adaptive evolutionary neural network are trained by using the initial data set, and two network models are saved; expected antenna performance parameters are defined and input into the adaptive evolutionary neural network to obtain structure parameters; in the execution process, it is intelligently judged whether the simulation software needs to be called to perform full-wave simulation to generate new data pairs; then the simulation software is called to judge whether the structure parameters meet the defined expected performance parameters; if yes, the process is ended; otherwise, the newly generated data pairs are added to the initial data set to continue training the network to improve the network precision. The above steps are repeated, and the antenna structure design is completed after meeting the expectation.

Description

A Broadband Dual-Polarized Antenna Optimization Design Method Based on Adaptive Evolutionary Neural Networks Technical Field

[0001] This invention relates to the field of antenna design technology, and specifically to a broadband dual-polarized antenna optimization design method based on an adaptive evolutionary neural network. Background Technology

[0002] With the development of next-generation mobile communication technologies such as 5G and the Internet of Things, the demand for broadband antennas is rapidly increasing. Traditional antenna design uses simulation software to model complex antennas and optimize many parameters until performance requirements are met. This process has many problems, such as exceptionally long runtime and high memory consumption.

[0003] In recent years, researchers in the field of computational electromagnetics have attempted to accelerate and optimize the antenna modeling process using machine learning methods and artificial neural networks. A common approach is to use machine learning and artificial neural networks to build surrogate models and then use these trained models to optimize the antenna. However, a well-fitting surrogate model often requires training on a large dataset. Collecting massive datasets is also extremely time-consuming, which is unacceptable for efficient antenna design. Summary of the Invention

[0004] The purpose of this invention is to provide a broadband dual-polarized antenna optimization design method based on adaptive evolutionary neural networks, which aims to solve the technical problem of large data volume required by the surrogate model in the existing antenna optimization modeling process.

[0005] To achieve the above objectives, this invention provides a broadband dual-polarized antenna optimization design method based on an adaptive evolutionary neural network, comprising the following steps:

[0006] Step 1: Establish an initial antenna model, use simulation software to perform a full-wave simulation of the initial antenna model, and collect the initial dataset at the same time;

[0007] Step 2: Train the first and second networks in the adaptive evolutionary neural network using the initial dataset, and save the two network models;

[0008] Step 3: Define the desired antenna performance parameters and input the antenna performance parameters into the adaptive evolutionary neural network to obtain the structural parameters;

[0009] Step 4: Call the simulation software to determine whether the structural parameters meet the performance parameters defined in Step 3.

[0010] If the condition is met, proceed to step 5.

[0011] If the conditions are not met, the newly generated data pairs in step 3 will be added to the initial dataset to continue training the adaptive evolutionary neural network;

[0012] Step 5: Repeat steps 2 to 4 to obtain the structural parameters that meet the desired antenna performance parameters, and then complete the antenna structure design.

[0013] The initial dataset consists of a small number of data pairs of the form (X, Y), where X represents antenna structural parameters and Y represents antenna performance parameters.

[0014] The antenna performance parameters involve one or more of the following: return loss, reflection coefficient, antenna gain, antenna pattern, and radiation efficiency. After collecting the initial dataset, in order to eliminate the influence of dimensions and enable the network to converge quickly, the initial dataset is 0-1 normalized.

[0015] The first network is used to train the mapping relationship between performance parameters and structural parameters, specifically by inputting performance parameters to obtain structural parameters.

[0016] The second network is used to train the mapping joint between structural parameters and performance parameters, specifically by inputting structural parameters to obtain performance parameters.

[0017] In the process of defining the desired antenna performance parameters and inputting these parameters into the adaptive evolutionary neural network to obtain structural parameters, the system intelligently determines whether to call simulation software to perform full-wave simulation to generate new data pairs. Specifically, this includes the following steps:

[0018] Step 31: Input the defined desired antenna performance parameters into the first network to obtain the structural parameters SP. (i) ;

[0019] Step 32: SP (i) SP is obtained by performing random perturbation (i+1) ;

[0020] Step 33: Transfer SP (i+1) The new performance parameters are obtained by inputting them into the second network.

[0021] Step 34: Determine if the error between the expected performance parameter and the new performance parameter is less than the threshold. If it is less, call the simulation software according to SP. (i+1) Perform a full-wave simulation; otherwise, return to step 32.

[0022] The process of calling simulation software to determine whether the structural parameters meet the performance parameters defined in step 3 includes the following steps:

[0023] Step 41: Call the simulation software according to SP (i)and SP (i+1) Full-wave simulation is performed to obtain the actual performance parameters;

[0024] Step 42: Determine whether the desired antenna performance parameters are met. If they are met, the process ends; otherwise, add the newly generated data pairs from Step 3 to the initial dataset.

[0025] This invention provides a broadband dual-polarized antenna optimization design method based on an adaptive evolutionary neural network. The method involves establishing an initial antenna model, performing a full-wave simulation on the initial model using simulation software, and simultaneously collecting an initial dataset. The first and second networks of the adaptive evolutionary neural network are trained using the initial dataset, and both network models are saved. Desired antenna performance parameters are defined and input into the adaptive evolutionary neural network to obtain structural parameters. During execution, the system intelligently determines whether to call the simulation software for full-wave simulation to generate new data pairs. Then, the simulation software is called to determine whether the structural parameters meet the defined desired performance parameters. If they do, the process ends; otherwise, the newly generated data pairs are added to the initial dataset to continue training the network to improve its accuracy. These steps are repeated until the structural parameters that meet the antenna performance parameters are obtained, thus completing the antenna structure design. This method reduces the number of simulation software calls, decreases the amount of data required for the surrogate model, and improves the efficiency of antenna optimization design. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 is a flowchart illustrating a broadband dual-polarized antenna optimization design method based on an adaptive evolutionary neural network according to the present invention.

[0028] Figure 2 is a diagram of the adaptive deep neural network topology of the present invention.

[0029] Figure 3 is a flowchart of adaptive deep neural network modeling in a specific embodiment of the present invention.

[0030] Figure 4 is a schematic diagram of the structure of a broadband dual-polarized antenna in a specific embodiment of the present invention.

[0031] Figure 5 shows the S-values ​​of three sets of optimization targets for the broadband dual-polarized antenna optimized using different methods in a specific embodiment of the present invention. 11 curve. Detailed Implementation

[0032] 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.

[0033] Please refer to Figure 1. This invention provides a broadband dual-polarized antenna optimization design method based on an adaptive evolutionary neural network, comprising the following steps:

[0034] S1: Establish an initial antenna model, use simulation software to perform full-wave simulation on the initial antenna model, and collect the initial dataset at the same time;

[0035] S2: Train the first and second networks in the adaptive evolutionary neural network using the initial dataset, and save the two network models;

[0036] S3: Define the desired antenna performance parameters and input the antenna performance parameters into the adaptive evolutionary neural network to obtain structural parameters;

[0037] S4: Call the simulation software to determine whether the structural parameters meet the performance parameters defined in step S3.

[0038] If the condition is met, proceed to step S5.

[0039] If the conditions are not met, the newly generated data pairs in step S3 will be added to the initial dataset to continue training the adaptive evolutionary neural network.

[0040] S5: Repeat steps S2 to S4 to obtain the structural parameters that meet the desired antenna performance parameters, and then complete the antenna structure design.

[0041] The following provides further explanation in conjunction with the specific implementation steps:

[0042] S1. Establish an initial antenna model and perform a full-wave simulation of the initial antenna model using simulation software. Collect a small dataset consisting of data pairs of the form (X, Y), where X represents the antenna structural parameters and Y represents the antenna performance parameters.

[0043] Specifically, the performance parameters include the antenna's return loss, reflection coefficient, antenna gain, antenna pattern, and radiation efficiency. In this example, electromagnetic simulation software is used to simulate the structural parameters of the initial antenna model to obtain the return loss (S...). 11 After collecting the dataset, in order to eliminate the influence of units and enable the network to converge quickly, the dataset is standardized to 0-1 according to the following formula.

[0044]

[0045] Where x represents an element in the dataset.

[0046] The antenna structure is shown in Figure 4. Sensitivity analysis of the antenna was performed in the simulation software HFSS to obtain the results for S... 11 The five structural parameters with the greatest impact are: H, Ang, PIN, d, and r. Based on the design of experiment (DOE) method and the actual physical meaning of the antenna, 25 data points were collected as the initial dataset as shown in Table 1, which took 2 hours and 3 minutes. The frequency range of the data sampling points was 1.7 to 2.7 GHz.

[0047] Table 1. Range of values ​​for optimization parameters for the initial dataset.

[0048]

[0049]

[0050] S2 uses the initial dataset from S1 to train the first and second networks in the adaptive evolutionary neural network and saves both network models.

[0051] Specifically, the method involves constructing a first network F1 and a second network F2 according to the topology diagram of the adaptive deep neural network shown in Figure 2, training them using the initial dataset, and saving the model with the minimum loss function. This invention uses mean squared error (MSE) as the network metric to find the optimal weight parameters. The loss functions of F1 and F2 are defined as follows:

[0052] LossFirstnetwork=MSE(predicted SP,real SP)

[0053] LossSecondnetwork=MSE(predicted S 11 ,real S 11 )

[0054] SP represents the structural parameters of the antenna.

[0055] S3 defines the desired antenna performance parameters, which are then input into an adaptive evolutionary neural network composed of the first and second networks in S2 to obtain the structural parameters. During algorithm execution, the algorithm intelligently determines whether to call simulation software to perform full-wave simulation and generate new data pairs.

[0056] The specific method is as follows:

[0057] S31 inputs the defined desired antenna performance parameters into the first network to obtain the structural parameters SP. (i)

[0058] Specifically, the desired antenna performance parameter is defined as S. To better measure the stability of the algorithm, this embodiment involves three sets of optimization objectives:

[0059] Case 1: S 11 (f = 2.1 GHz) <= -22 dB, and S 11 A bandwidth less than -18dB is greater than 500MHz.

[0060] Case 2: S 11 (f = 2.4 GHz) <= -25 dB, and S 11 A bandwidth less than -18dB is greater than 700MHz.

[0061] Case 3: S 11 (f = 2.15 GHz) <= -34 dB, and S 11 A bandwidth less than -20dB is greater than 600MHz. Inputting this performance parameter into the first network yields the structural parameter SP. (i)

[0062] S32 to SP (i) SP is obtained by performing random perturbation (i+1) .

[0063] Specifically, the random perturbation is performed according to the following formula:

[0064] SP (i+1) =SP (i) +α·FR

[0065] Where FR represents the feasible region of random perturbations of structural parameters (usually set as the smallest unit of change of structural parameters), α represents a random variable between -1 and 1, and when SP (i+1) If the value is not in the feasible region, select the maximum or minimum value that is closest to the value in the feasible region for replacement.

[0066] S33 will SP (i+1) The input is fed into the second network to obtain new performance parameters.

[0067] Specifically, the SP generated after random perturbation (i+1) The new performance parameter S' is obtained by inputting it into the second network.

[0068] S34 determines whether the error between the expected performance parameter and the new performance parameter is less than a threshold. If it is less, it calls the simulation software based on SP. (i+1) Perform a full-wave simulation; otherwise, return to S32.

[0069] Specifically, if |S-S'|≤δ, then SP (i) and SP (i+1) If the values ​​are not substituted into the simulation software for calculation, otherwise return to S32 to re-perturb the random variables and obtain a new SP. (i+1) δ represents a predefined error threshold; when it is less than this value, the SP is considered to be in a state of error under the current circumstances. (i+1) This step is crucial. By using a second network to verify the structural parameters after random perturbation, we can determine whether simulation software calculations are necessary. This approach yields data pairs that effectively improve the network's fitting performance while reducing the number of simulation software calls, thus increasing optimization efficiency.

[0070] S4 calls the simulation software to determine whether the structural parameters meet the performance parameters defined in S3. If they do, the process ends; otherwise, the newly generated data pairs in S3 are added to the original dataset to continue training the network and improve its accuracy.

[0071] The specific method is as follows:

[0072] S41 calls the simulation software according to SP (i) and SP (i+1) Full-wave simulation is performed to obtain the actual performance parameters.

[0073] Specifically, as shown in the flowchart in Figure 3, when |S-S'|≤δ, the simulation software is called according to SP. (i) and SP (i+1) Full-wave simulation is performed to obtain the actual performance parameters S* and S**. Here, EM represents the full-wave simulation performed using simulation software.

[0074] S42 determines whether the expected performance parameters are met. If they are met, the process ends; otherwise, the newly generated data pairs in S3 are added to the original dataset.

[0075] Specifically, determine whether S* and S** satisfy the desired performance parameter S. If so, the process ends; otherwise, collect (SP) data. (i) (S*) and (SP) (i+1) Add the S**) data pairs to the initial dataset and return to step S2.

[0076] S5 repeats S2 to S4 until the structural parameters that meet the antenna performance parameters are obtained, thus completing the antenna structure design.

[0077] Furthermore, the present invention provides a specific embodiment for comparison and illustration:

[0078] In this embodiment, based on the optimization objective, optimization is first performed using the optimization method in simulation software. Then, the optimization results and optimization time of the proposed method (Adaptive Evolutionary Neural Network, AENN) are compared with those of Artificial Neural Network (ANN), Extreme Learning Machine (ELM), and Gaussian process regression (GPR). For fair comparison, when using ANN for optimization, only the original AENN is replaced with ANN, while the rest of the workflow remains unchanged. For the ELM and GPR machine learning methods, uniform sampling is performed using simulation software according to Table 1, collecting 3125 sets of data as the dataset. Each time, n*25 data points are randomly selected from the dataset as the training set and (n*25) / 10 data points as the test set for model training. Each model is trained 10 times, and the model with the highest accuracy on the test set is used for antenna optimization. The obtained optimization results are then substituted into the simulation software for calculation. If the optimization objective is not met, a new dataset is selected, where n = 1, 2, 3... 125.

[0079] The optimized structural parameters after different methods are shown in Table 2, and the optimization time is shown in Table 3. The S-values ​​after satisfying the optimization objective are also shown in Table 3. 11 The parameters are shown in Figure 5. As can be seen from Figure 5, although various methods can ultimately satisfy the optimization objective, the method of this invention achieves better optimization results, obtaining a smaller S value while still satisfying the optimization objective. 11Parameters. As can be observed from Table 3, the method of this invention can achieve the optimization target in a very short time compared to other advanced methods (since the dataset was pre-collected in Case 1, only the time required for algorithm iterations is counted in Case 2 and Case 3). The effectiveness of AENN in this invention's method can be effectively verified by comparing it with the ANN method. Using a second network to verify the results of the first network can significantly reduce the number of simulation software calls while obtaining high-fidelity data more conducive to network evolution, thereby reducing the optimization time. For example, in Case 2, AENN found the target solution in only 6 iterations, while ANN required 66 iterations. The ELM and GPR machine learning methods require more datasets to train efficient surrogate models; for example, in Case 3, ELM requires 275 sets of data, and GPR requires 660 sets. Generally, once a machine learning model is trained, it can be used as a surrogate model for optimization. As with the method of this invention, only a very small amount of data needs to be pre-collected for antenna parameter optimization. This aligns with practical applications, as an antenna model often has multiple different optimization objectives. Therefore, the aforementioned machine learning methods and the method of this invention are more suitable for multiple different optimization objectives. However, ELM and GPR require too much pre-collected data, which contradicts the goal of efficient antenna optimization design. The experiments above demonstrate that the method of this invention can significantly improve optimization efficiency in practical applications of antenna parameter optimization and possesses strong algorithmic stability, with the number of iterations less than 10 for all three optimization objectives.

[0080] Table 2 Results of structural parameter optimization using different methods

[0081]

[0082]

[0083] Table 3. Optimization time required for different methods

[0084]

[0085] The above description discloses only one preferred embodiment of the present invention, and should not be construed as limiting the scope of the present invention. Those skilled in the art will understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A broadband dual-polarized antenna optimization design method based on adaptive evolutionary neural networks, characterized in that, The process includes the following steps: Step 1: Establish an initial antenna model, perform a full-wave simulation of the initial antenna model using simulation software, and collect an initial dataset; Step 2: Train the first and second networks in the adaptive evolutionary neural network using the initial dataset, and save the two network models; Step 3: Define the desired antenna performance parameters and input the antenna performance parameters into the adaptive evolutionary neural network to obtain structural parameters; During the process of defining the desired antenna performance parameters and inputting the antenna performance parameters into the adaptive evolutionary neural network to obtain structural parameters, the system will intelligently determine whether it is necessary to call the simulation software to perform a full-wave simulation to generate new data pairs, specifically including the following steps: Step 31: Input the defined desired antenna performance parameters into the first network to obtain the structural parameters SP. (i) Step 32: SP (i) SP is obtained by performing random perturbation (i+1) Step 33: Transfer SP (i+1) The input is fed into the second network to obtain new performance parameters; Step 34: Determine if the error between the expected performance parameter and the new performance parameter is less than the threshold. If it is less, call the simulation software according to SP. (i+1) Perform full-wave simulation; otherwise, return to step 32. Step 4: Call the simulation software to determine whether the structural parameters meet the performance parameters defined in step 3. If they do, jump to step 5. If not, add the newly generated data pairs from step 3 to the initial dataset and continue training the adaptive evolutionary neural network. The process of calling the simulation software to determine whether the structural parameters meet the performance parameters defined in step 3 includes the following steps: Step 41: Call the simulation software according to SP... (i) and SP (i+1) Perform full-wave simulation to obtain the actual performance parameters; Step 42: Determine whether the desired antenna performance parameters are met. If they are met, the process ends; otherwise, add the newly generated data pairs from Step 3 to the initial dataset; Step 5: Repeat Steps 2 to 4 to obtain the structural parameters that meet the desired antenna performance parameters, and then complete the antenna structure design.

2. The broadband dual-polarized antenna optimization design method based on adaptive evolutionary neural network as described in claim 1, characterized in that, The initial dataset consists of a small number of data pairs of the form (X, Y), where X represents antenna structural parameters and Y represents antenna performance parameters.

3. The broadband dual-polarized antenna optimization design method based on adaptive evolutionary neural network as described in claim 2, characterized in that, The antenna performance parameters involve one or more of the following: return loss, reflection coefficient, antenna gain, antenna pattern, and radiation efficiency. After collecting the initial dataset, in order to eliminate the influence of dimensions and enable the network to converge quickly, the initial dataset is normalized to 0-1.

4. The broadband dual-polarized antenna optimization design method based on adaptive evolutionary neural network as described in claim 1, characterized in that, The first network is used to train the mapping relationship between performance parameters and structural parameters, specifically by inputting performance parameters to obtain structural parameters; the second network is used to train the mapping joint between structural parameters and performance parameters, specifically by inputting structural parameters to obtain performance parameters.

Images