Ultra-wideband antenna design method and system based on deep learning, and medium

Abstract

The invention discloses an ultra-wideband antenna design method and system based on deep learning and a medium, and the method comprises the steps: obtaining a design target of a target ultra-wideband antenna, and extracting structure parameters related to the antenna performance according to antenna design instance data; preparing sample data according to the structure parameters and preprocessing the sample data, building an agent model based on an NAR dynamic network in combination with a multi-layer sensor, training the agent model by using the sample data, and configuring agent model parameters; and obtaining an optimal structure parameter of the target ultra-wideband antenna by using the trained proxy model, performing simulation verification on the output optimal structure parameter, and configuring a design scheme of the target ultra-wideband antenna. According to the method, the NAR dynamic network is combined with the multi-layer perceptron to establish the proxy model, so that optimization tasks of different structural parameters can be adapted, the efficiency is improved on the premise of ensuring the performance, and the design speed and precision of the ultra-wideband antenna are improved.

Description

Technical Field

[0001] The present invention relates to the field of antenna design technology, and more specifically, to an ultra-wideband antenna design method, system and medium based on deep learning. Background Art

[0002] Antenna is a device that converts electromagnetic waves into electric current or electric current into electromagnetic waves. It is an indispensable part of wireless communication system. In the era of booming wireless communication, antennas have to shoulder more and more tasks, and the system's requirements for antennas are also increasing. How to quickly design antennas that meet the requirements has become a task worth exploring. The system's requirements for antenna performance are getting higher and higher. When designing antennas, it is necessary to comprehensively consider various aspects of performance and perform multi-objective optimization on the antenna. With the development and progress of artificial intelligence, the use of proxy models and deep learning algorithms to optimize antenna design has become an important research direction in the field of antennas.

[0003] Traditional antenna design methods use mathematical equations to calculate the electromagnetic field based on the complexity of the structure. For simple designs such as square patches, circular patches, and wire antennas, mathematical equations already exist and are derived from Maxwell's equations, which can be used to obtain the performance of the antenna at the considered frequency. However, for the optimization design of some complex antennas with multiple parameters, the derivation of equations is very difficult and complicated; high-fidelity electromagnetic simulation software such as HFSS, CST, and FEKO are used to analyze the performance of the antenna (S11 parameters, bandwidth, gain, etc.) and optimize the parameters of the antenna, but this method is to perform full-wave simulation of the antenna, and each simulation requires a lot of calculations, which is time-consuming and labor-intensive, and takes up a lot of memory. Therefore, how to use deep learning methods to build proxy models to improve efficiency while designing antennas that meet target requirements. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes an ultra-wideband antenna design method, system and medium based on deep learning, with the aim of proposing a deep learning-based proxy model to optimize the time-consuming problem in the antenna design process, thereby achieving higher accuracy while ensuring performance.

[0005] A first aspect of the present invention provides an ultra-wideband antenna design method based on deep learning, comprising:

[0006] Acquire a design goal of a target ultra-wideband antenna, extract antenna design example data based on the design goal, and extract structural parameters related to antenna performance according to the antenna design example data;

[0007] Prepare sample data and perform preprocessing according to the structural parameters, build a proxy model based on the NAR dynamic network combined with a multi-layer perceptron, use the sample data to train the proxy model, and configure proxy model parameters;

[0008] The trained proxy model is used to obtain the optimal structural parameters of the target ultra-wideband antenna, and the optimal structural parameters are simulated and verified to obtain a design solution of the target ultra-wideband antenna.

[0009] In this solution, antenna design example data is extracted based on the design goal, and structural parameters related to antenna performance are extracted according to the antenna design example data, specifically:

[0010] The design goal is segmented and the antenna design keyword library is pre-trained to process the word vector, keywords are extracted according to similarity calculation, different keywords and their corresponding values ​​are determined, and the design goal keywords and corresponding values ​​are used to construct a retrieval task:

[0011] Based on the retrieval task, antenna design instance data that meets the preset requirements is extracted from the historical ultra-wideband antenna design data, the retrieved antenna design instance data is structured, and a heterogeneous graph of different design target keywords is constructed based on semantic association and data association;

[0012] Extracting a meta-path starting from the design target keyword in the heterogeneous graph, constructing an adjacency matrix under two relationships based on semantic association and data association in the meta-path, and performing node feature aggregation based on the adjacency matrix under the two relationships using a graph attention network;

[0013] The weights of the two relationships are obtained through attention at the relationship level. The total weight of the keyword nodes in the meta-path is obtained based on the weights of the two relationships. The keyword nodes that meet the preset weight threshold are screened and further screened based on the type labels of the keyword nodes.

[0014] According to the screening results, structural parameter categories related to antenna performance are obtained, and the structural parameter categories are used to perform data integration in antenna design example data.

[0015] In this solution, an agent model is constructed based on the NAR dynamic network combined with a multi-layer perceptron, specifically:

[0016] Obtain data samples after integration of data of each structural parameter category, perform data fusion on the data samples using a Copula function, obtain a data sample set based on Copula tail correlation fusion, and divide the data sample set into training samples and test samples;

[0017] Initialize the network parameters of the NAR dynamic network, optimize the delay order and the number of neurons in the hidden layer based on the improved genetic algorithm, use the BP algorithm to train the NAR dynamic network, adjust the connection weights of the NAR dynamic network according to the error data, and obtain the optimized NAR dynamic network through iterative training;

[0018] The output data of the optimized NAR dynamic network is obtained and imported into the multi-layer perceptron for training. The model parameters of the NAR dynamic network and the multi-layer perceptron are fine-tuned according to the training results. After stopping the training, the test error is obtained according to the test sample. When the test error is less than the preset error threshold, the proxy model with the configured parameters is output.

[0019] In this scheme, the network parameters of the NAR dynamic network are initialized, and the delay order and the number of neurons in the hidden layer are optimized based on the improved genetic algorithm, specifically:

[0020] Initializing the network parameters of the NAR dynamic network, mapping the network parameters using a Circle chaotic sequence to generate a chaotic population, and mapping the obtained chaotic population to genetic population individuals to generate a chaotic initialization population;

[0021] Generating reverse positions of individuals in the chaotic initialization population, constructing a reverse population, merging the chaotic initialization population and the reverse population, and selecting dominant individuals through evaluation to generate a final population;

[0022] The improved genetic algorithm is used to optimize the individuals in the population. The average accuracy of the test samples is used as the fitness function to calculate the fitness of the individuals in the population. In the iterative process, adaptive mutations and dynamic weights that change with the number of iterations are introduced to increase the search ability for excellent individuals.

[0023] The fitness of individuals is continuously calculated through iteration, and the individuals are selected, crossed and adaptively mutated to generate the next generation population. After completing the set number of iterations, the optimal delay order and the optimal number of hidden layer neurons of the NAR dynamic network are obtained.

[0024] In this solution, the output data of the optimized NAR dynamic network is obtained and imported into the multi-layer perceptron for training. The model parameters of the NAR dynamic network and the multi-layer perceptron are fine-tuned according to the training results. Specifically:

[0025] Obtaining output data of the NAR dynamic model during the training process, using the output data as input of the multi-layer perceptron, initializing the network structure of the multi-layer perceptron, and determining network parameters through iterative training;

[0026] Using multiple fully connected layers in a multi-layer perceptron to perform feature dimension reduction on the output data, using a self-attention mechanism to assign weights to the reduced-dimensional output data, and compressing the weighted features through a fully connected layer to obtain deep features;

[0027] The NAR dynamic model is combined with the multi-layer perceptron to build an overall model, the root mean square error between the test data and the predicted output of the overall model is calculated as the test error, and the model parameters of the overall model are adjusted using back propagation according to the test error;

[0028] The goal is to minimize the test error. When it is less than the preset threshold, the proxy model with configured parameters is output.

[0029] In this scheme, the optimal structural parameters of the target ultra-wideband antenna are obtained, specifically:

[0030] Importing structural parameters corresponding to the design objectives of the target ultra-wideband antenna into the optimized proxy model, obtaining predicted optimal structural parameters, and obtaining simulation data based on the optimal structural parameters;

[0031] Perform simulation calculation according to the sample data corresponding to the structural parameters, obtain simulation results of the sample data, calculate the degree of deviation between the simulation data of the optimal structural parameters and the simulation results of the sample data, and normalize the degree of deviation;

[0032] When the normalized deviation degree is less than a preset threshold, a design scheme of the target ultra-wideband antenna is generated according to the optimal structural parameters.

[0033] A second aspect of the present invention provides an ultra-wideband antenna design system based on deep learning, the system comprising a design target acquisition unit, a structural parameter confirmation unit, a proxy model unit, a proxy model training unit and a simulation verification unit;

[0034] The design target acquisition unit is responsible for acquiring the design target of the target ultra-wideband antenna;

[0035] The structural parameter confirmation unit is responsible for extracting structural parameters related to antenna performance from antenna design example data according to design objectives;

[0036] The proxy model unit is responsible for constructing a proxy model based on the NAR dynamic network combined with a multi-layer perceptron, configuring the proxy model parameters through training, and obtaining the optimal structural parameters of the target ultra-wideband antenna using the proxy model;

[0037] The agent model training unit is responsible for training the agent model using the improved genetic algorithm to obtain the optimal model parameters;

[0038] The simulation verification unit is responsible for performing simulation verification on the optimal structural parameters, and obtaining the design scheme of the target ultra-wideband antenna after verification.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention utilizes NAR dynamic network in combination with multi-layer perceptron to establish a proxy model that can adapt to optimization tasks of different structural parameters. The training process of the proxy model is optimized by an improved genetic algorithm, which solves the problem of difficulty in training samples for multi-dimensional structural parameter antennas. The efficiency is improved while ensuring performance, the design speed and accuracy of ultra-wideband antennas are improved, and the labor cost of ultra-wideband antenna design is further reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the drawings required for use in the embodiments or exemplary descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to the drawings without paying creative work.

[0042] Figure 1 A flow chart of an ultra-wideband antenna design method based on deep learning is shown;

[0043] Figure 2 A flowchart of an embodiment for constructing an agent model is shown;

[0044] Figure 3 A flow chart of obtaining optimal structural parameters of a target ultra-wideband antenna according to an embodiment is shown;

[0045] Figure 4 A block diagram of a deep learning based ultra-wideband antenna design system is shown. DETAILED DESCRIPTION

[0046] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0047] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0048] Figure 1 A flow chart of an ultra-wideband antenna design method based on deep learning is shown.

[0049] like Figure 1 As shown, this embodiment provides an ultra-wideband antenna design method based on deep learning, including:

[0050] S102, obtaining a design goal of a target ultra-wideband antenna, extracting antenna design example data based on the design goal, and extracting structural parameters related to antenna performance according to the antenna design example data;

[0051] S104, preparing sample data according to the structural parameters and performing preprocessing, building a proxy model based on the NAR dynamic network combined with a multi-layer perceptron, training the proxy model using the sample data, and configuring proxy model parameters;

[0052] S106, using the trained proxy model to obtain optimal structural parameters of the target ultra-wideband antenna, performing simulation verification on the optimal structural parameters, and obtaining a design solution for the target ultra-wideband antenna.

[0053] It should be noted that the ultra-wideband antenna can match the characteristics within a specific working frequency band and work continuously and stably within the working bandwidth. The ultra-wideband antenna coexists with other devices working in narrowband systems and works stably together. The ultra-wideband antenna is improved to have a stopband notch function or other filtering characteristics to filter out the interference of narrowband signals. The basic parameters of the ultra-wideband antenna include: working bandwidth, input impedance, return loss and voltage standing wave ratio, radiation pattern, antenna efficiency and gain, etc.

[0054] The design goal of the target ultra-wideband antenna is obtained, such as fractal structure, size and position information, and the design goal is segmented and the antenna design keyword library is pre-trained to process the word vector. The antenna design keyword library is obtained by accessing the antenna design related database and the related field knowledge graph for pre-training. Keywords are extracted according to the similarity calculation through the pre-trained antenna design keyword library. If the similarity between the word vector corresponding to the design goal and the keyword vector in the keyword library meets the standard, it is marked as a keyword. Through traversal calculation, different keywords and their corresponding values ​​are determined, and the design goal keywords and corresponding values ​​are used to construct a retrieval task: based on the retrieval task, antenna design instance data that meets the preset requirements are extracted from the historical ultra-wideband antenna design data.

[0055] The retrieved antenna design instance data are structured, and the semantic association and data association are obtained by calculating the Pearson correlation coefficient and mRMR algorithm using the structured instance data. A heterogeneous graph of different design target keywords is constructed based on the semantic association and data association. In the heterogeneous graph, the design target keywords extracted from the structured instance data are used as nodes, and the edge structure between the design target keyword nodes is designed through the semantic association and data association.

[0056] A meta-path starting from the design target keyword is extracted in the heterogeneous graph by random walk, and an adjacency matrix under two relationships is constructed in the meta-path based on semantic association and data association. A graph attention network is used to aggregate node features based on the adjacency matrix under the two relationships. The aggregation feature of the k-th layer of neighboring nodes is obtained by aggregating the embedding of the k-th layer of neighboring nodes by the mean aggregation method. The aggregation feature is embedded and spliced ​​with the risk feature of the k-1th layer of the design target keyword node to obtain the embedding representation of the design target keyword node under the relationship graph. A nonlinear transformation is performed through the graph attention mechanism, and the weight under the relationship is obtained after averaging after the activation function and multiplication with the learnable parameters. The weights under the two relationships are obtained through the attention at the relationship level. The total weight of the keyword nodes in the meta-path is obtained according to the sum of the weights of the two relationships, and the keyword nodes that meet the preset weight threshold are screened, and the keyword nodes are screened again according to the type label of the keyword nodes. The structural parameter category related to the antenna performance is obtained according to the screening result, and the instance data is extracted from the antenna design instance data by using the structural parameter category to realize sample data integration.

[0057] Figure 2 A flowchart of an embodiment for constructing an agent model is shown.

[0058] According to an embodiment of the present invention, an agent model is constructed based on a NAR dynamic network combined with a multi-layer perceptron, specifically:

[0059] S202, obtaining data samples after integration of data of various structural parameter categories, performing data fusion on the data samples using a Copula function, obtaining a data sample set based on Copula tail correlation fusion, and dividing the data sample set into training samples and test samples;

[0060] S204, initializing the network parameters of the NAR dynamic network, optimizing the delay order and the number of neurons in the hidden layer based on the improved genetic algorithm, training the NAR dynamic network using the BP algorithm, adjusting the connection weights of the NAR dynamic network according to the error data, and obtaining the optimized NAR dynamic network through iterative training;

[0061] S206, obtaining the output data of the optimized NAR dynamic network, importing it into the multi-layer perceptron for training, fine-tuning the model parameters of the NAR dynamic network and the multi-layer perceptron according to the training results, obtaining the test error according to the test sample after stopping the training, and outputting the proxy model after the parameter configuration when the test error is less than the preset error threshold.

[0062] It should be noted that the Copula function is used to establish a data dependency model between data samples and characterize the tail correlation between data samples, which can further optimize the multivariate analysis in antenna design. The delay order of the NAR dynamic network enables the network to have a memory function. A reasonable delay order can improve the network performance. The network parameters of the NAR dynamic network are initialized, including the delay order and the number of hidden layer neurons. The Circle chaotic sequence is used to map the network parameters to generate a chaotic population. The obtained chaotic population is mapped to the genetic population individuals to generate a chaotic initialization population. The calculation formula of the chaotic individual X after mapping is: X = X L +(X L -X U )×α, where X U , X L are the upper and lower limits of the individual dimension respectively, and α is the chaotic variable.

[0063] Generate the reverse position of individuals in the chaotic initialization population to construct a reverse population. The individuals in the reverse population are represented by X f =X L +X U -X, merge the chaotic initialization population and the reverse population, and generate the final population by evaluating and selecting the dominant individuals; optimize the individuals in the population through the improved genetic algorithm, calculate the fitness of the individuals in the population according to the average accuracy of the test samples as the fitness function, and introduce adaptive mutations and dynamic weights that change with the number of iterations during the iteration process to increase the search ability of excellent individuals; the adaptive mutation expands the individual selection range, and performs comparative calculations of fitness in a larger range, which is conducive to individual search for better solutions. In addition, the dynamic weight changes with the number of iterations, and the fitness search calculation of the individual is optimized. The calculation formula of the dynamic weight λ(i) is: λ(i)=λ c -i(λ c -λ z ) / T max ,λ c is the initial dynamic weight, λ z is the final dynamic weight, i is the number of iterations, T max is the maximum number of iterations; the fitness of individuals is continuously calculated through iterations, and individuals are selected, crossed, and adaptively mutated to generate the next generation population. After completing the set number of iterations, the optimal delay order and the optimal number of hidden layer neurons of the NAR dynamic network are obtained.

[0064] Preferably, an elite strategy is introduced in the mutation operation of the genetic algorithm, and the direction of individual mutation is clarified according to the elite individual combined with differential update, and the optimal fitness individual is selected in combination with the reverse learning strategy. A random factor is introduced on the basis of the two strategies, and a random factor threshold is preset. When the random factor is greater than or equal to the preset random factor threshold, the elite strategy is used to mutate the optimal fitness individual, and when the random factor is less than the preset random factor threshold, the reverse learning strategy is used to update the optimal fitness individual; wherein the individual X after the elite strategy mutation g It is expressed as: in They represent the individual with the highest fitness, the individual with the second highest fitness, and the individual with the third highest fitness, respectively. β1β2 represents a random constant. The optimal fitness individual is updated using the reverse learning strategy, which is expressed as: Where X U , X L are the upper and lower limits of the individual dimension respectively, and θ1θ2 represent random constants.

[0065] The BP algorithm is used to train the NAR dynamic network. The BP algorithm adjusts the network weights through error data. First, the error vector between the true value and the predicted value is calculated, and the connection weights of the network are adjusted according to the error vector. The network is trained iteratively, and the termination conditions of the iteration are that the global error of the network meets the error requirements and the number of training times reaches the maximum value, ensuring that the NAR dynamic network meets the stability requirements.

[0066] It should be noted that the multilayer perceptron is a feedforward neural network composed of multiple fully connected neural networks. The output data of the NAR dynamic model during the training process is obtained, and the output data is used as the input of the multilayer perceptron to initialize the network structure of the multilayer perceptron, determine the network parameters through iterative training, and update the network parameters through the back propagation algorithm until the network converges; use multiple fully connected layers in the multilayer perceptron to reduce the feature dimension of the output data, use the self-attention mechanism to assign weights to the reduced output data, and compress the weighted features through the fully connected layer to obtain deep features; and import the deep features into the last fully connected layer to obtain the predicted output, combine the NAR dynamic model with the multilayer perceptron to build the overall model, calculate the root mean square error between the test data and the predicted output of the overall model as the test error, and use back propagation to adjust the model parameters of the overall model according to the test error; with the goal of minimizing the test error, when it is less than the preset threshold, the proxy model with parameter configuration is output.

[0067] Figure 3 A flow chart of an embodiment for obtaining optimal structural parameters of a target ultra-wideband antenna is shown.

[0068] According to an embodiment of the present invention, the optimal structural parameters of the target ultra-wideband antenna are obtained, specifically:

[0069] S302, importing structural parameters corresponding to the design objectives of the target ultra-wideband antenna into the optimized proxy model, obtaining predicted optimal structural parameters, and obtaining simulation data based on the optimal structural parameters;

[0070] S304, performing simulation calculation according to the sample data corresponding to the structural parameters, obtaining simulation results of the sample data, calculating the degree of deviation between the simulation data of the optimal structural parameters and the simulation results of the sample data, and normalizing the degree of deviation;

[0071] S306: When the normalized deviation degree is less than a preset threshold, a design scheme of the target ultra-wideband antenna is generated according to the optimal structural parameters.

[0072] It should be noted that HFSS is called to implement the design and simulation of the ultra-wideband antenna, and the VB script is generated through sample data to control the HFSS software, and a three-dimensional model is generated in HFSS, and then the simulation is solved and the output data is output to obtain the corresponding simulation results. The degree of deviation between the simulation data of the optimal structural parameters and the simulation results of the sample data is calculated by one or more metric functions, for example, the distance vectors between the simulation data are generated using Euclidean distance, Chebyshev distance, Canberra distance and Mahalanobis distance, all distance vector representations are spliced ​​and aggregated, and the spliced ​​and aggregated distance vectors are averaged and pooled to obtain the final degree of deviation.

[0073] Figure 4 A block diagram of a deep learning based ultra-wideband antenna design system is shown.

[0074] The second embodiment of the present invention provides an ultra-wideband antenna design system 4 based on deep learning, which includes a design target acquisition unit 401, a structure parameter confirmation unit 402, a proxy model unit 403, a proxy model training unit 404 and a simulation verification unit 405;

[0075] The design target acquisition unit is responsible for acquiring the design target of the target ultra-wideband antenna;

[0076] The structural parameter confirmation unit is responsible for extracting structural parameters related to antenna performance from antenna design example data according to design objectives;

[0077] The proxy model unit is responsible for constructing a proxy model based on the NAR dynamic network combined with a multi-layer perceptron, configuring the proxy model parameters through training, and obtaining the optimal structural parameters of the target ultra-wideband antenna using the proxy model;

[0078] The agent model training unit is responsible for training the agent model using the improved genetic algorithm to obtain the optimal model parameters;

[0079] The simulation verification unit is responsible for performing simulation verification on the optimal structural parameters, and obtaining the design scheme of the target ultra-wideband antenna after verification.

[0080] A third embodiment of the present invention provides a computer-readable storage medium, which includes an ultra-wideband antenna design method program based on deep learning. When the ultra-wideband antenna design method program based on deep learning is executed by a processor, the steps of the ultra-wideband antenna design method based on deep learning are implemented.

[0081] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0082] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0083] Those skilled in the art can understand that: all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0084] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0085] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for designing ultra-wideband antennas based on deep learning, characterized in that: The following steps are involved: Acquire a design goal of a target ultra-wideband antenna, extract antenna design example data based on the design goal, and extract structural parameters related to antenna performance according to the antenna design example data; Prepare sample data and perform preprocessing according to the structural parameters, build a proxy model based on the NAR dynamic network combined with a multi-layer perceptron, use the sample data to train the proxy model, and configure proxy model parameters; The trained proxy model is used to obtain the optimal structural parameters of the target ultra-wideband antenna, and the optimal structural parameters are simulated and verified to obtain a design solution of the target ultra-wideband antenna.

2. The ultra-wideband antenna design method based on deep learning according to claim 1, characterized in that: Extracting antenna design example data based on the design goal, and extracting structural parameters related to antenna performance according to the antenna design example data, specifically: The design goal is segmented and the antenna design keyword library is pre-trained to process the word vector, keywords are extracted according to similarity calculation, different keywords and their corresponding values ​​are determined, and the design goal keywords and corresponding values ​​are used to construct a retrieval task: Based on the retrieval task, antenna design instance data that meets the preset requirements is extracted from the historical ultra-wideband antenna design data, the retrieved antenna design instance data is structured, and a heterogeneous graph of different design target keywords is constructed based on semantic association and data association; Extracting a meta-path starting from the design target keyword in the heterogeneous graph, constructing an adjacency matrix under two relationships based on semantic association and data association in the meta-path, and performing node feature aggregation based on the adjacency matrix under the two relationships using a graph attention network; The weights of the two relationships are obtained through attention at the relationship level. The total weight of the keyword nodes in the meta-path is obtained based on the weights of the two relationships. The keyword nodes that meet the preset weight threshold are screened and further screened based on the type labels of the keyword nodes. According to the screening results, structural parameter categories related to antenna performance are obtained, and the structural parameter categories are used to perform data integration in antenna design example data.

3. The ultra-wideband antenna design method based on deep learning according to claim 1, characterized in that: The agent model is constructed based on the NAR dynamic network combined with a multi-layer perceptron. Specifically: Obtain data samples after integration of data of each structural parameter category, perform data fusion on the data samples using a Copula function, obtain a data sample set based on Copula tail correlation fusion, and divide the data sample set into training samples and test samples; Initialize the network parameters of the NAR dynamic network, optimize the delay order and the number of neurons in the hidden layer based on the improved genetic algorithm, use the BP algorithm to train the NAR dynamic network, adjust the connection weights of the NAR dynamic network according to the error data, and obtain the optimized NAR dynamic network through iterative training; The output data of the optimized NAR dynamic network is obtained and imported into the multi-layer perceptron for training. The model parameters of the NAR dynamic network and the multi-layer perceptron are fine-tuned according to the training results. After stopping the training, the test error is obtained according to the test sample. When the test error is less than the preset error threshold, the proxy model with the configured parameters is output.

4. The ultra-wideband antenna design method based on deep learning according to claim 3, characterized in that: Initialize the network parameters of the NAR dynamic network, and optimize the delay order and the number of neurons in the hidden layer based on the improved genetic algorithm, specifically: Initializing the network parameters of the NAR dynamic network, mapping the network parameters using a Circle chaotic sequence to generate a chaotic population, mapping the obtained chaotic population to genetic population individuals to generate a chaotic initialization population; Generating reverse positions of individuals in the chaotic initialization population, constructing a reverse population, merging the chaotic initialization population and the reverse population, and selecting dominant individuals through evaluation to generate a final population; The improved genetic algorithm is used to optimize the individuals in the population. The average accuracy of the test samples is used as the fitness function to calculate the fitness of the individuals in the population. In the iterative process, adaptive mutations and dynamic weights that change with the number of iterations are introduced to increase the search ability for excellent individuals. The fitness of individuals is continuously calculated through iteration, and individuals are selected, crossed and adaptively mutated to generate the next generation population. After completing the set number of iterations, the optimal delay order and the optimal number of hidden layer neurons of the NAR dynamic network are obtained.

5. The ultra-wideband antenna design method based on deep learning according to claim 3, characterized in that: Obtain the output data of the optimized NAR dynamic network, import it into the multi-layer perceptron for training, and fine-tune the model parameters of the NAR dynamic network and the multi-layer perceptron according to the training results. Specifically: Obtaining output data of the NAR dynamic model during the training process, using the output data as input of the multi-layer perceptron, initializing the network structure of the multi-layer perceptron, and determining network parameters through iterative training; Using multiple fully connected layers in a multi-layer perceptron to perform feature dimension reduction on the output data, using a self-attention mechanism to assign weights to the reduced-dimensional output data, and compressing the weighted features through a fully connected layer to obtain deep features; The NAR dynamic model is combined with the multi-layer perceptron to build an overall model, the root mean square error between the test data and the predicted output of the overall model is calculated as the test error, and the model parameters of the overall model are adjusted using back propagation according to the test error; The goal is to minimize the test error. When it is less than the preset threshold, the proxy model with configured parameters is output.

6. The ultra-wideband antenna design method based on deep learning according to claim 1, characterized in that: Obtain the optimal structural parameters of the target ultra-wideband antenna, specifically: Importing structural parameters corresponding to the design objectives of the target ultra-wideband antenna into the optimized proxy model, obtaining predicted optimal structural parameters, and obtaining simulation data based on the optimal structural parameters; Perform simulation calculation according to the sample data corresponding to the structural parameters, obtain simulation results of the sample data, calculate the degree of deviation between the simulation data of the optimal structural parameters and the simulation results of the sample data, and normalize the degree of deviation; When the normalized deviation degree is less than a preset threshold, a design scheme of the target ultra-wideband antenna is generated according to the optimal structural parameters.

7. A deep learning-based ultra-wideband antenna design system, characterized in that: Implementing the ultra-wideband antenna design method based on deep learning as described in any one of claims 1 to 6, the system includes a design target acquisition unit, a structure parameter confirmation unit, a proxy model unit, a proxy model training unit and a simulation verification unit; The design target acquisition unit is responsible for acquiring the design target of the target ultra-wideband antenna; The structural parameter confirmation unit is responsible for extracting structural parameters related to antenna performance from antenna design example data according to design objectives; The proxy model unit is responsible for constructing a proxy model based on the NAR dynamic network combined with a multi-layer perceptron, configuring the proxy model parameters through training, and using the proxy model to obtain the optimal structural parameters of the target ultra-wideband antenna; The agent model training unit is responsible for training the agent model using the improved genetic algorithm to obtain the optimal model parameters; The simulation verification unit is responsible for performing simulation verification on the optimal structural parameters, and obtaining the design scheme of the target ultra-wideband antenna after verification.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes an ultra-wideband antenna design method program based on deep learning. When the ultra-wideband antenna design method program based on deep learning is executed by a processor, the steps of the ultra-wideband antenna design method based on deep learning as described in any one of claims 1 to 6 are implemented.

Images