Antenna robustness design method and device based on hybrid deep learning
By combining deep learning and genetic algorithms, an antenna response substitution model is constructed, which solves the problem of high computational complexity in traditional robustness optimization and achieves efficient and accurate antenna robustness design.
Patent Information
- Application Number
- CN202610610625.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-06
- Publication Date
- 2026-06-26
AI Technical Summary
Traditional robust optimization methods have high computational complexity and large computational load, making it difficult to converge to the global optimum quickly. Furthermore, traditional deep learning models cannot accurately characterize high-frequency nonlinear resonance characteristics, failing to meet the accuracy requirements of antenna design.
A hybrid deep learning-based approach is adopted to construct an antenna response substitution model by obtaining the Gaussian distribution curve of high-sensitivity design variables. A genetic algorithm is then used for robust optimization, which reduces the dimensionality of variables, decreases the amount of computation, and improves computational efficiency and accuracy.
It significantly shortens the antenna design cycle, reduces computational complexity, improves the accuracy and efficiency of robustness optimization, avoids premature convergence, and finds the optimal solution with high robustness.
Smart Images
Figure CN122287379A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of wireless communication technology, and in particular to an antenna robustness design method and apparatus based on hybrid deep learning. Background Technology
[0002] Antennas, as crucial components in wireless communication, perform spatial conversion of energy and information through the radiation and reception of electromagnetic waves, and have significant application value in various fields such as modern communication, radar, and navigation. Due to limitations in current manufacturing precision, random errors are unavoidable in antenna dimensions, leading to performance deviations and decreased yield rates in batch-produced antennas. Therefore, it is necessary to improve antenna robustness—that is, the ability of an antenna to function normally despite manufacturing errors. Higher robustness results in better design performance and economic benefits.
[0003] Traditional robust optimization methods require optimization of all design variables. While they can accurately simulate manufacturing uncertainties, they fail to consider the differences in the contributions of each variable to the antenna response. This leads to an exponential increase in computational complexity with the dimensionality of the variables, wasting resources and making it prone to getting trapped in local optima. This method relies on full-wave electromagnetic (EM) simulation to evaluate performance, with design cycles lasting weeks or even months. Furthermore, when using the Monte Carlo method to simulate manufacturing errors, each sample requires a re-simulation, causing the computational load to increase exponentially with the number of samples, making it difficult to converge quickly to the global optimum in high-dimensional spaces.
[0004] Data-driven alternative models have become an important trend in alleviating the burden of simulation. However, traditional deep learning models have limitations when dealing with broadband complex electromagnetic responses. They are difficult to accurately characterize high-frequency nonlinear resonance characteristics, and the single mean square error loss function is prone to numerical oscillations, which cannot meet the accuracy requirements.
[0005] Therefore, a technical solution is urgently needed to solve the above problems. Summary of the Invention
[0006] Therefore, it is necessary to provide an antenna robustness design method and device based on hybrid deep learning to address the above-mentioned technical problems.
[0007] A robust antenna design method based on hybrid deep learning, the method comprising the following steps: Obtain antenna structure parameters, and use the ratio of antenna structure design variables to antenna wavelength as a screening criterion to screen antenna design variables and obtain potential high-sensitivity design variables; Within the antenna manufacturing error range, each design variable among the potential high-sensitivity design variables is sampled to obtain the Gaussian distribution curves of different design variables in the operating frequency band. The mean, standard deviation and probability density function corresponding to the Gaussian distribution curves of each design variable are statistically analyzed to perform sensitivity analysis. Design variables whose sensitivity meets the preset requirements are selected as high-sensitivity design variables. Combining a feature-modulation hybrid deep learning model with highly sensitive design variables as input, S The parameters are used as output to construct an antenna response substitution model; A robust objective function is constructed, and the robustness calculation is performed using the antenna response as a substitute model. Within the preset optimization range, the antenna parameters are robustly optimized using a genetic algorithm, and the optimal high-sensitivity design variable parameters are output.
[0008] In one embodiment, the feature-modulation hybrid deep learning model includes a multilayer perceptron branch and a one-dimensional convolutional neural network branch; The multilayer perceptron branch includes multiple cascaded fully connected layers, activation layers, batch normalized layers, and random deactivation layers with a dropout rate of 15%. The one-dimensional convolutional neural network branch is driven by a conditional encoder and a feature modulation layer.
[0009] In one embodiment, antenna structure parameters are obtained, and the antenna design variables are screened using the ratio of antenna structure design variables to antenna wavelength as a screening criterion to obtain potential high-sensitivity design variables, including: Obtain antenna structure parameters and construct screening criteria based on the ratio of antenna structure design variables to antenna wavelength: ; in, To select indicators, This refers to the number of operating frequency bands of the multi-frequency antenna; These are design variables for the antenna structure; The antenna wavelength; Select the one that satisfies The design variables are considered as potential high-sensitivity design variables, among which, To screen indicators The preset filtering threshold.
[0010] In one embodiment, each design variable among the potential high-sensitivity design variables is sampled within the antenna manufacturing error range to obtain Gaussian distribution curves of different design variables in the operating frequency band, including: Within the antenna manufacturing tolerance range, each design variable among the potential high-sensitivity design variables is sampled, and the corresponding values for the samples are obtained through electromagnetic simulation. S Parametric curves; For each sample S The maximum response value of the parameter curve at the operating frequency is used as the response point for robustness assessment. The response point is: ; in, For the first iResponse points of each operating frequency band; The first in a multi-frequency antenna i One operating frequency band; This is due to antenna manufacturing errors; For potential high-sensitivity design variables; The probability density distribution between samples of different sizes and their corresponding response points obtained from statistical sampling is used to construct Gaussian distribution curves of different design variables in the operating frequency band.
[0011] In one embodiment, the probability density distribution is calculated according to the following formula: ; in, It is a probability density distribution; Standard deviation; This is the mean.
[0012] In one embodiment, the mean, standard deviation, and probability density function corresponding to the Gaussian distribution curves of each design variable are statistically analyzed to perform sensitivity analysis. Design variables whose sensitivity meets preset requirements are selected as high-sensitivity design variables, including: Calculate the mean, standard deviation, and probability density function of the Gaussian distribution curves for each design variable, and then calculate the Gaussian distribution curves. S The probability that the maximum value of the parameter response is less than or equal to -10: ; in, Different operating frequency bands of the antenna in the Gaussian distribution curve S The probability that the maximum value of the parameter response is less than or equal to -10; Different operating frequency bands of the antenna in the Gaussian distribution curve S Maximum value of parameter response; Based on the statistical results and calculated probabilities, design variables are selected as high-sensitivity design variables according to preset standards.
[0013] In one embodiment, the antenna response substitution model is: ; in, The original model was costly; A low-cost alternative model; It is a hybrid deep learning model.
[0014] In one embodiment, the robustness objective function is: ; in, The indicators are considered acceptable. The result obtained by fitting a Gaussian distributionS Parameter response estimate.
[0015] In one embodiment, a robustness calculation is performed using an antenna response replacement model. Within a preset optimization range, a genetic algorithm is used to optimize the antenna parameters for robustness, outputting the optimal high-sensitivity design variable parameters, including: Optimization scope is Let the population size be N , No. t The population of generations can be represented as: ; in, For the first t The first generation of the population i Individual; The fitness function is used to evaluate the quality of each individual and select superior individuals for breeding; the fitness function is: ; in, The fitness function; To ensure that the number of antennas meeting design specifications accounts for a certain percentage of the total during manufacturing. K The proportion; The selection probability in the filtering is: ; in, For the first i The probability of an individual's choice; By exchanging partial gene segments among individuals through crossover operations, new gene combinations are generated. Random variations are introduced through mutation operations to maintain population diversity. After multiple generations of iterative evolution, the optimal, highly sensitive design variable parameters are output. ; in, Parameters for designing the optimal high sensitivity.
[0016] An antenna robustness design device based on hybrid deep learning, the device comprising: The preprocessing module is used to obtain antenna structure parameters and uses the ratio of antenna structure design variables to antenna wavelength as a screening index to screen the antenna design variables and obtain potential high-sensitivity design variables. The design variable confirmation module is used to sample each design variable among the potential high-sensitivity design variables within the antenna manufacturing error range, obtain the Gaussian distribution curves of different design variables in the operating frequency band, and statistically analyze the mean, standard deviation and probability density function corresponding to the Gaussian distribution curves of each design variable to perform sensitivity analysis and select the design variables whose sensitivity meets the preset requirements as high-sensitivity design variables. Alternative model building modules are used to combine feature-modulated hybrid deep learning models with highly sensitive design variables as input. S The parameters are used as output to construct an antenna response substitution model; The parameter optimization module is used to construct a robust objective function, perform robustness calculations by replacing the antenna response model with the antenna response, and optimize the antenna parameters robustness using a genetic algorithm within a preset optimization range, outputting the optimal high-sensitivity design variable parameters.
[0017] The aforementioned antenna robustness design method and device based on hybrid deep learning simulates manufacturing process errors by combining sampling methods within the antenna manufacturing error range, thereby obtaining design variables and... S Sensitivity analysis was performed on the Gaussian distribution curve of the statistical mapping relationship between parameter responses to identify variables more sensitive to manufacturing errors, thereby reducing the dimensionality of variables and decreasing the complexity and computational load of subsequent antenna robustness optimization. An antenna response substitution model was constructed based on a feature-modulation hybrid deep learning model to replace traditional electromagnetic simulation, providing efficient computational support for subsequent genetic algorithm optimization, significantly shortening the optimization cycle, and ensuring the accuracy of response prediction while avoiding distortion of optimization results due to model errors. By constructing a robustness objective function and using a genetic algorithm to optimize antenna parameter robustness, a highly robust optimal solution was found, avoiding premature convergence of the gradient method in multi-peak landscapes caused by random errors. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating an antenna robustness design method based on hybrid deep learning in one embodiment. Figure 2 This is a schematic diagram of the structure of a WLAN dual-band monopole antenna in one embodiment, wherein (a) is a top view of the WLAN dual-band monopole antenna and (b) is a front view of the WLAN dual-band monopole antenna; Figure 3 Here are Gaussian distribution curves of different design variables in one embodiment at the operating frequency band, where (a) is a Gaussian distribution curve from 2.4 GHz to 2.5 GHz and (b) is a Gaussian distribution curve from 5.1 GHz to 7.2 GHz. Figure 4 Here is a structural diagram of a feature-modulated hybrid deep learning model in one embodiment; Figure 5The following figures illustrate the modeling and simulation results of the antenna response substitution model for different samples in one embodiment: (a) shows the modeling and simulation results of the antenna response substitution model for sample 1; (b) shows the modeling and simulation results of the antenna response substitution model for sample 2; (c) shows the modeling and simulation results of the antenna response substitution model for sample 3; and (d) shows the modeling and simulation results of the antenna response substitution model for sample 4. Figure 6 This is a schematic diagram of the genetic algorithm optimization process in one embodiment; Figure 7 This is a structural block diagram of an antenna robustness design device based on hybrid deep learning in one embodiment. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0020] In one embodiment, such as Figure 1 As shown, a robust antenna design method based on hybrid deep learning is provided, the method comprising the following steps: Step 201: Obtain antenna structure parameters. Use the ratio of antenna structure design variables to antenna wavelength as a screening criterion to screen the antenna design variables and obtain potential high-sensitivity design variables.
[0021] Step 202: Within the antenna manufacturing error range, sample each design variable among the potential high-sensitivity design variables, obtain the Gaussian distribution curves of different design variables in the operating frequency band, and statistically analyze the mean, standard deviation and probability density function corresponding to the Gaussian distribution curves of each design variable. Select the design variables whose sensitivity meets the preset requirements as high-sensitivity design variables.
[0022] Step 203: Combine the feature-modulation-based hybrid deep learning model with high-sensitivity design variables as input. S The parameters are used as output to construct an antenna response substitution model.
[0023] Step 204: Construct a robust objective function, use the antenna response as a substitute model for robustness calculation, and use a genetic algorithm to optimize the robustness of antenna parameters within a preset optimization range, outputting the optimal high-sensitivity design variable parameters.
[0024] In the aforementioned antenna robustness design method based on hybrid deep learning, sampling methods are combined within the antenna manufacturing error range to simulate manufacturing process errors and obtain design variables. SSensitivity analysis was performed on the Gaussian distribution curve of the statistical mapping relationship between parameter responses to identify variables more sensitive to manufacturing errors, thereby reducing the dimensionality of variables and decreasing the complexity and computational load of subsequent antenna robustness optimization. An antenna response substitution model was constructed based on a feature-modulation hybrid deep learning model to replace traditional electromagnetic simulation, providing efficient computational support for subsequent genetic algorithm optimization, significantly shortening the optimization cycle, and ensuring the accuracy of response prediction while avoiding distortion of optimization results due to model errors. By constructing a robustness objective function and using a genetic algorithm to optimize antenna parameter robustness, a highly robust optimal solution was found, avoiding premature convergence of the gradient method in multi-peak landscapes caused by random errors.
[0025] Reference Figure 2 In one embodiment, the antenna is a WLAN dual-band monopole antenna. Specifically, the operating frequencies of the WLAN dual-band monopole antenna are respectively... and , It is 2.45GHz. It is 5.49GHz.
[0026] It is understood that the method described in this invention is applied to multi-band antennas, including but not limited to WLAN dual-band monopole antennas.
[0027] In one embodiment, the feature-modulation hybrid deep learning model includes a multilayer perceptron branch and a one-dimensional convolutional neural network branch; The multilayer perceptron branch includes multiple cascaded fully connected layers, activation layers, batch normalized layers, and random deactivation layers with a dropout rate of 15%. The one-dimensional convolutional neural network branch is driven by a conditional encoder and a feature modulation layer.
[0028] Specifically, such as Figure 4 As shown, a feature-modulated hybrid deep learning model is used to extract high-sensitivity design variable features using a multilayer perceptron (MLP) branch to generate a preliminary antenna. SThe parameter curves are used; data augmentation is performed by adding small random Gaussian noise to the high-sensitivity design variables in the training set to improve the model's robustness and generalization ability; the design variables with feature perturbations are input into a multilayer perceptron (MLP) for global coarse prediction; in the MLP branch, fully connected layers are responsible for mapping low-dimensional design variables to the target dimension, constructing basic mapping relationships, with 64, 128, and 256 neurons respectively; nonlinear activation layers endow the network with nonlinear ability to fit complex electromagnetic responses; batch normalization layers accelerate model convergence by standardizing the output features of the previous layer; 15% random deactivation layers effectively prevent overfitting by randomly shielding some neurons. The MLP branch outputs a coarse prediction through deep nonlinear mapping. S Parameter curve results.
[0029] Curve local feature refinement and residual compensation are performed by driving a one-dimensional convolutional neural network branch (1D-CNN) through a conditional encoder and a feature modulation layer (FiLM); a coarse-grained model is generated in the multilayer perceptron branch. S Simultaneously with the parametric curve results, the input high-sensitivity design variables are processed by a conditional encoder and mapped into a high-dimensional modulation vector through a two-layer fully connected network. At the same time, the coarse S-parameter curves generated by the multilayer perceptron branch are expanded in channel dimension and fed into the CNN branch (a 1D-CNN branch). The CNN branch uses multiple layers of convolutional kernels of different scales to capture local high-frequency features, with the number of channels in each feature map layer set to 32, 64, 64, and 32 respectively.
[0030] It's important to note that the feature modulation layer enables the convolutional network to maintain real-time awareness and control over the global physical dimensions of the input antenna while extracting subtle local features. Furthermore, the feature modulation layer is embedded within each cascaded module of the CNN, located after the convolution and batch normalization operations and before the non-linear activation function in each layer. For each convolutional feature layer of the neural network (CNN), the FiLM layer utilizes two parallel linear mapping matrices to dynamically generate a scaling factor that matches the number of channels in the current feature map based on the modulation vector. With offset factor Subsequently, a channel-wise affine transformation is performed on the convolutional feature map to achieve precise modulation and offset correction of local features. The CNN branch outputs a 301-dimensional fine residual sequence, which is then element-wise added to the coarse curve generated by the MLP branch to obtain the final high-precision antenna. S Parameter prediction curve.
[0031] It should also be noted that during the model training phase of the feature-modulated hybrid deep learning model, a combined loss function consisting of mean square error and first-order derivative error is introduced to update the network weights: ; ; ; in, This is the combined loss function; N This represents the total number of frequency sampling points within the target frequency band. These are the weighting coefficients for the derivative loss; In the first i Frequency points The actual location obtained from EM simulation S Parameter values; In the first i Frequency points The predicted value output by the replacement model is used. The first part of the combined loss function is the basic mean squared error (MSE) between the predicted curve and the true S-parameter curve; the second part is the derivative loss, which approximates the local slope of the curve by calculating the difference between adjacent frequency sampling points in the predicted curve, and calculates the mean squared error between it and the local slope of the true curve. In the backpropagation calculation, this derivative loss is assigned a weight coefficient of 0.3, which is added to the basic MSE loss to form the total penalty term. This combined loss function effectively constrains the shape and trend of the predicted curve, eliminates numerical fluctuations that are not physically governed, and ensures that the model output shape is highly consistent with the true value.
[0032] In one embodiment, step 201 involves obtaining antenna structure parameters, using the ratio of antenna structure design variables to antenna wavelength as a screening criterion to filter antenna design variables and obtain potential high-sensitivity design variables, including: Obtain antenna structure parameters and construct screening criteria based on the ratio of antenna structure design variables to antenna wavelength: ; in, To select indicators, This refers to the number of operating frequency bands of the multi-frequency antenna; These are design variables for the antenna structure; The antenna wavelength; Select the one that satisfies The design variables are considered as potential high-sensitivity design variables, among which, To screen indicators The preset filtering threshold.
[0033] Specifically, when the antenna is a WLAN dual-band monopole antenna, select and The design variables are considered as potential high-sensitivity design variables, among which, For the operating frequency is The corresponding screening criteria at that time For the operating frequency is The corresponding screening criteria at that time, after confirming the screening criteria When the value is numerical, it is the same as The numerical ratio should roughly follow and The ratio of .
[0034] It should be noted that the ratio of antenna structure design variables to antenna wavelength determines the degree to which dimensional fluctuations affect antenna electrical performance, and therefore serves as a screening criterion. When the ratio is small, even small changes can easily cause significant shifts in antenna resonance characteristics, impedance matching, and other performance aspects. The screening criteria correspond to the operating frequency; when an antenna operates at multiple frequencies, there is a corresponding number of screening criteria. The selection threshold for the range of screening criteria is preset and set based on experience.
[0035] For a dual-band monopole antenna in WLAN, the antenna wavelength is calculated using the following formula: ; in, The first in a multi-frequency antenna i One operating frequency band; The speed at which electromagnetic waves propagate in free space. The antenna wavelength was obtained as , .choose and The design variables were used as potential high-sensitivity design variables to obtain... x =[ h , s , l , w , l 1, l 2, r 1, r 2, d (corresponding to respectively) Figure 2 (each position in the text), corresponding value x =[1.52, 3.50, 14.00, 2.00, 4.25, 7.00, 10.25,18.50, 11.00](mm).
[0036] In this embodiment, the antenna structure parameters are preprocessed to complete the initial screening, thereby reducing the dimensionality of design variables.
[0037] In one embodiment, step 202 involves sampling each design variable among the potential high-sensitivity design variables within the antenna manufacturing error range to obtain Gaussian distribution curves of different design variables in the operating frequency band, including: Within the antenna manufacturing tolerance range, each design variable among the potential high-sensitivity design variables is sampled, and the corresponding values for the samples are obtained through electromagnetic simulation. S Parametric curves; For each sample S The maximum response value of the parameter curve at the operating frequency is used as the response point for robustness assessment. The response point is: ; in, For the first i Response points of each operating frequency band; The first in a multi-frequency antenna i One operating frequency band; This is due to antenna manufacturing errors; For potential high-sensitivity design variables; The probability density distribution between samples of different sizes and their corresponding response points obtained from statistical sampling is used to construct Gaussian distribution curves of different design variables in the operating frequency band.
[0038] Specifically, for a WLAN dual-band monopole antenna, the operating frequency bands are as follows: and , The frequency range is 2.4GHz to 2.5GHz. The range is from 5.1 GHz to 7.2 GHz; Specifically, using 0.1mm as the antenna manufacturing tolerance, the sample... ± Uniform sampling is performed, and the corresponding samples are obtained through electromagnetic simulation. S The parameter curves were generated using 21 samples for each design variable. (Refer to...) Figure 3 The figure shows the Gaussian distribution curves of different design variables in the operating frequency band.
[0039] In one embodiment, the probability density distribution is calculated according to the following formula: ; in, It is a probability density distribution; Standard deviation; This is the mean.
[0040] In one embodiment, step 202 involves statistically analyzing the mean, standard deviation, and probability density function corresponding to the Gaussian distribution curves of each design variable, performing sensitivity analysis, and selecting design variables whose sensitivity meets preset requirements as high-sensitivity design variables, including: Calculate the mean, standard deviation, and probability density function of the Gaussian distribution curves for each design variable, and then calculate the Gaussian distribution curves. SThe probability that the maximum value of the parameter response is less than or equal to -10: ; in, Different operating frequency bands of the antenna in the Gaussian distribution curve S The probability that the maximum value of the parameter response is less than or equal to -10; Different operating frequency bands of the antenna in the Gaussian distribution curve S Maximum value of parameter response; Based on the statistical results and calculated probabilities, design variables are selected as high-sensitivity design variables according to preset standards.
[0041] Understandably, by analyzing Gaussian distribution curves, different Gaussian distribution parameters can be obtained. Analyzing the mean, standard deviation, and probability density function corresponding to each design variable curve allows for further evaluation of the antenna's performance variation and reliability under the influence of manufacturing errors. This improves the effectiveness of robust antenna design and optimization. Based on the statistical results, variables with higher sensitivity are selected as high-sensitivity design variables for robust optimization. x 1.
[0042] Specifically, for WLAN dual-band monopole antennas, the 5.1GHz to 7.2GHz frequency band should be selected. Furthermore, design variables with a standard deviation greater than 0.45 are considered high-sensitivity design variables. In the Gaussian distribution curve, the first... i Response points of each operating frequency band S The probability that the maximum value of the parameter response is less than or equal to -10.
[0043] Specifically, for the WLAN dual-band monopole antenna, the mean, standard deviation, and probability density function corresponding to the Gaussian distribution curves of each design variable were statistically analyzed, and the results are shown in Tables 1 and 2: Table 1 Key information of Gaussian distribution curves from 2.4GHz to 2.5GHz
[0044] Table 2 Key Information of Gaussian Distribution Curve from 5.1GHz to 7.2GHz
[0045] From Table 1, Table 2 and Figure 3 We can obtain, This represents the probability that an antenna will meet performance requirements under the influence of a certain variable. The smaller the probability value, the higher the risk that the design variable will cause the antenna performance to fail to meet requirements, and the stronger the disturbance to the antenna performance. When this variable fluctuates slightly... SThe parameter response is more likely to exceed the acceptable threshold of -10dB, thus making it more sensitive to the impact on antenna performance. The standard deviation, as a statistic reflecting the degree of fluctuation of a random variable, indicates that a larger value means the manufacturing error of the design variable has a greater impact. S The stronger the influence of the discreteness of the parameter response, the more likely small fluctuations in the variable will cause [a problem / effect]. S The parameter response exhibits a wider range of shifts, further amplifying the uncertainty in antenna performance. Statistical analysis reveals significant differences in the statistical characteristics of each design variable across the 2.4GHz-2.5GHz and 5.1GHz-7.2GHz frequency bands: in the lower frequency bands, the variables... The percentage is generally close to 100%, indicating that most design variables in this frequency band... S The parameter distribution is generally within the acceptable range, and its performance response to variable fluctuations has tended to stabilize; while the high-frequency range... Significant differences exist between variables; therefore, the selection of highly sensitive design variables should prioritize statistical characteristics within the high-frequency range. Based on this, further analysis should be conducted incorporating the fluctuation of standard deviation: selecting variables within the high-frequency range... Design variables with less than 90% and a standard deviation greater than or equal to 0.45 are used as high-sensitivity design variables for robust optimization. x 1. Obtain x 1=[ h , l 1, l 2, d ], corresponding to the initial value x 1=[1.52,4.25,7.00,11.00](mm).
[0046] In this embodiment, sensitivity analysis is performed on potential highly sensitive variables in the antenna, and corresponding statistical analysis methods are adopted to evaluate the impact of manufacturing errors on antenna performance, providing a basis for the robust design of the antenna, thereby screening out highly sensitive design variables that have a significant impact on antenna performance.
[0047] In one embodiment, the robustness of the antenna under initial design variables was calculated. Given manufacturing errors, the design can be categorized as either a qualified product (antenna meeting design specifications) or a defective product (antenna not meeting design specifications). The robustness calculation is defined as follows: ; in, A qualified indicator indicates a high-sensitivity design variable. x 1. Manufacturing errors exist At that time, does the antenna performance meet the design specifications? To ensure that the number of antennas meeting design specifications accounts for a certain percentage of the total during manufacturing. KThe initial robustness of the WLAN dual-band monopole antenna was calculated to be 84%.
[0048] In one embodiment, the antenna response substitution model is: ; in, The original model was costly; A low-cost alternative model; It is a hybrid deep learning model.
[0049] Specifically, in x 1±3 Within the range, Monte Carlo sampling was used to generate 3000 sample points (i.e., a dataset), with high-sensitivity design variables. For input, S Using the parameters as output, the dataset is divided into 80% training set and 20% test set, resulting in the antenna response replacement model: ; The mean squared error of the obtained test set was 0.696. The modeling performance was evaluated using the test set. Further, four samples were randomly selected to compare the modeling results and simulation results of the hybrid deep learning model. Figure 5 As shown, the fitting results are in good agreement with the true values.
[0050] Understandably, the trained alternative model can complete the training under any combination of highly sensitive variables within milliseconds. S Parametric curve prediction, the predicted curve is This model effectively avoids the high computational cost caused by relying on electromagnetic simulation in traditional robust optimization, and provides an efficient performance evaluation tool for the subsequent optimization process.
[0051] In this embodiment, the constructed antenna response substitution model focuses on highly sensitive variables, which reduces training complexity while improving adaptability to processing error scenarios, achieving a balance between accuracy and efficiency.
[0052] In one embodiment, the robustness objective function is: ; in, The indicators are considered acceptable. The result obtained by fitting a Gaussian distribution S Parameter response estimate.
[0053] In one embodiment, in step 204, robustness calculation is performed using an antenna response replacement model. Within a preset optimization range, a genetic algorithm is used to optimize the robustness of antenna parameters, outputting the optimal high-sensitivity design variable parameters, including: Optimization scope is Let the population size be N , No. t The population of generations can be represented as: ; in, For the first t The first generation of the population i Individual; The fitness function is used to evaluate the quality of each individual and select superior individuals for breeding; the fitness function is: ; in, The fitness function; To ensure that the number of antennas meeting design specifications accounts for a certain percentage of the total during manufacturing. K The proportion; The selection probability in the filtering is: ; in, For the first i The probability of an individual's choice; By exchanging partial gene segments among individuals through crossover operations, new gene combinations are generated. Random variations are introduced through mutation operations to maintain population diversity. After multiple generations of iterative evolution, the optimal, highly sensitive design variable parameters are output. ; in, Parameters for designing the optimal high sensitivity.
[0054] It should be noted that, since manufacturing errors have a relatively small fluctuation range for variables, the optimization range is set to simulate the error range generated during manufacturing processes. The fitness function is used to evaluate the quality of each individual, and superior individuals are selected for reproduction. In this process, the higher the fitness value, the better the quality of the solution. The selection operation in the process uses methods such as roulette wheel selection and tournament selection based on fitness to select superior individuals for reproduction.
[0055] Crossover operations (such as single-point crossover, multi-point crossover, or uniform crossover) exchange partial gene fragments between individuals to generate new gene combinations. Simultaneously, mutation operations introduce random variations to maintain population diversity and prevent the algorithm from getting trapped in local optima. After multiple generations of iterative evolution, the individual with the highest fitness in the population gradually approaches the global optimum. In the optimization, the population size is set to 20, and the number of iterations is 20. The changes in the average fitness value and the optimal fitness value in each generation are as follows: Figure 6 As shown. The final optimal solution is... x =[1.3676, 4.2696, 6.8279, 11.1519] (mm), with an optimal robustness of 100%. Electromagnetic simulation was performed to verify the robustness, which is consistent with the results of the genetic algorithm. The robustness of the initial size is 84%, indicating that the antenna robustness improved by 16% after optimization.
[0056] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.
[0057] In one embodiment, such as Figure 7 As shown, an antenna robustness design device based on hybrid deep learning is provided, the device comprising: The preprocessing module 901 is used to obtain antenna structure parameters and to filter the antenna design variables by using the ratio of antenna structure design variables to antenna wavelength as a screening index to obtain potential high-sensitivity design variables.
[0058] The design variable confirmation module 902 is used to sample each design variable among the potential high-sensitivity design variables within the antenna manufacturing error range, obtain the Gaussian distribution curves of different design variables in the operating frequency band, and statistically analyze the mean, standard deviation and probability density function corresponding to the Gaussian distribution curves of each design variable to perform sensitivity analysis and select the design variables whose sensitivity meets the preset requirements as high-sensitivity design variables.
[0059] Alternative model building module 903 is used to combine a feature-modulated hybrid deep learning model with a highly sensitive design variable as input. S The parameters are used as output to construct an antenna response substitution model.
[0060] The parameter optimization module 904 is used to construct a robust objective function, perform robustness calculations by replacing the antenna response model with the antenna response, optimize the antenna parameters robustness using a genetic algorithm within a preset optimization range, and output the optimal high-sensitivity design variable parameters.
[0061] Specific limitations regarding the antenna robustness design optimization device based on sensitivity and neural networks can be found in the limitations of the antenna robustness design optimization method based on sensitivity and neural networks mentioned above, and will not be repeated here. Each module in the aforementioned antenna robustness design optimization device based on sensitivity and neural networks can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.
[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0063] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A robust antenna design method based on hybrid deep learning, characterized in that, The method includes the following steps: Obtain antenna structure parameters, and use the ratio of antenna structure design variables to antenna wavelength as a screening criterion to screen antenna design variables and obtain potential high-sensitivity design variables; Within the antenna manufacturing error range, each design variable among the potential high-sensitivity design variables is sampled to obtain the Gaussian distribution curves of different design variables in the operating frequency band. The mean, standard deviation and probability density function corresponding to the Gaussian distribution curves of each design variable are statistically analyzed to perform sensitivity analysis. Design variables whose sensitivity meets the preset requirements are selected as high-sensitivity design variables. Combining a feature-modulation hybrid deep learning model with highly sensitive design variables as input, S The parameters are used as output to construct an antenna response substitution model; A robust objective function is constructed, and the robustness calculation is performed using the antenna response as a substitute model. Within the preset optimization range, the antenna parameters are robustly optimized using a genetic algorithm, and the optimal high-sensitivity design variable parameters are output.
2. The antenna robustness design method based on hybrid deep learning according to claim 1, characterized in that, The feature-modulation hybrid deep learning model includes a multilayer perceptron branch and a one-dimensional convolutional neural network branch. The multilayer perceptron branch includes multiple cascaded fully connected layers, activation layers, batch normalized layers, and random deactivation layers with a dropout rate of 15%. The one-dimensional convolutional neural network branch is driven by a conditional encoder and a feature modulation layer.
3. The antenna robustness design method based on hybrid deep learning according to claim 1, characterized in that, Obtain antenna structure parameters, and use the ratio of antenna structure design variables to antenna wavelength as a screening criterion to filter antenna design variables and obtain potential high-sensitivity design variables, including: Obtain antenna structure parameters and construct screening criteria based on the ratio of antenna structure design variables to antenna wavelength: in, To select indicators, This refers to the number of operating frequency bands of the multi-frequency antenna; These are design variables for the antenna structure; The antenna wavelength; Select the one that satisfies The design variables are considered as potential high-sensitivity design variables, among which, To screen indicators The preset filtering threshold.
4. The antenna robustness design method based on hybrid deep learning according to claim 3, characterized in that, Within the antenna manufacturing tolerance range, each design variable among the potential high-sensitivity design variables is sampled to obtain Gaussian distribution curves of different design variables at the operating frequency band, including: Within the antenna manufacturing tolerance range, each design variable among the potential high-sensitivity design variables is sampled, and the corresponding values for the samples are obtained through electromagnetic simulation. S Parametric curves; For each sample S The maximum response value of the parameter curve at the operating frequency is used as the response point for robustness assessment. The response point is: in, For the first i Response points of each operating frequency band; The first in a multi-frequency antenna i One operating frequency band; This is due to antenna manufacturing errors; For potential high-sensitivity design variables; The probability density distribution between samples of different sizes and their corresponding response points obtained from statistical sampling is used to construct Gaussian distribution curves of different design variables in the operating frequency band.
5. The antenna robustness design method based on hybrid deep learning according to claim 4, characterized in that, The probability density distribution is calculated according to the following formula: in, It is a probability density distribution; Standard deviation; This is the mean.
6. The antenna robustness design method based on hybrid deep learning according to claim 5, characterized in that, Analyze the mean, standard deviation, and probability density function of the Gaussian distribution curves for each design variable, perform sensitivity analysis, and select design variables whose sensitivity meets the preset requirements as high-sensitivity design variables, including: Calculate the mean, standard deviation, and probability density function of the Gaussian distribution curves for each design variable, and then calculate the Gaussian distribution curves. S The probability that the maximum value of the parameter response is less than or equal to -10: in, Different operating frequency bands of the antenna in the Gaussian distribution curve S The probability that the maximum value of the parameter response is less than or equal to -10; Different operating frequency bands of the antenna in the Gaussian distribution curve S Maximum value of parameter response; Based on the statistical results and calculated probabilities, design variables are selected as high-sensitivity design variables according to preset standards.
7. The antenna robustness design method based on hybrid deep learning according to claim 6, characterized in that, The antenna response substitution model is as follows: in, The original model was costly; A low-cost alternative model; It is a hybrid deep learning model.
8. The antenna robustness design method based on hybrid deep learning according to claim 7, characterized in that, The robustness objective function is: in, The indicators are considered acceptable. The result obtained by fitting a Gaussian distribution S Parameter response estimate.
9. The antenna robustness design method based on hybrid deep learning according to claim 8, characterized in that, Robustness calculations are performed using an antenna response replacement model. Within a preset optimization range, a genetic algorithm is used to optimize the antenna parameters for robustness, outputting the optimal high-sensitivity design variable parameters, including: Optimization scope is Let the population size be N , No. t The population of generations can be represented as: in, For the first t The first generation of the population i Individual; The fitness function is used to evaluate the quality of each individual and select superior individuals for breeding; the fitness function is: in, The fitness function; To ensure that the number of antennas meeting design specifications accounts for a certain percentage of the total during manufacturing. K The proportion; The selection probability in the filtering is: in, For the first i The probability of an individual's choice; By exchanging partial gene segments among individuals through crossover operations, new gene combinations are generated. Random variations are introduced through mutation operations to maintain population diversity. After multiple generations of iterative evolution, the optimal, highly sensitive design variable parameters are output. in, Parameters for designing the optimal high sensitivity.
10. An antenna robustness design device based on hybrid deep learning, characterized in that, The device includes: The preprocessing module is used to obtain antenna structure parameters and uses the ratio of antenna structure design variables to antenna wavelength as a screening index to screen the antenna design variables and obtain potential high-sensitivity design variables. The design variable confirmation module is used to sample each design variable among the potential high-sensitivity design variables within the antenna manufacturing error range, obtain the Gaussian distribution curves of different design variables in the operating frequency band, and statistically analyze the mean, standard deviation and probability density function corresponding to the Gaussian distribution curves of each design variable to perform sensitivity analysis and select the design variables whose sensitivity meets the preset requirements as high-sensitivity design variables. Alternative model building modules are used to combine feature-modulated hybrid deep learning models with highly sensitive design variables as input. S The parameters are used as output to construct an antenna response substitution model; The parameter optimization module is used to construct a robust objective function, perform robustness calculations by replacing the antenna response model with the antenna response, and optimize the antenna parameters robustness using a genetic algorithm within a preset optimization range, outputting the optimal high-sensitivity design variable parameters.