Antenna optimization design method and device based on semi-supervised width learning system

Abstract

The invention discloses an antenna optimization design method and device based on a semi-supervised width learning system, the method comprising the steps of: constructing a structural model of an antenna to be optimized; generating a data set consisting of antenna size parameters and corresponding return loss values; dividing the data set into an initial training set, a test set and an unlabeled sample data set; using the initial training set to train a convolution-width learning system and a convolution-stacked width learning system; inputting the unlabeled sample data set into the trained convolution-width learning system and the trained convolution-stacked width learning system to predict pseudo-labeled data, and adding the pseudo-labeled data to a labeled sample training data set; using the updated labeled sample training data set to cross-train and coordinately train the convolution-width learning system and the convolution-stacked width learning system until a learning system that meets the set rules is obtained. The invention can reduce the computational cost in the antenna optimization design process.

Description

Technical Field

[0001] The invention belongs to the technical field of antenna optimization design, and in particular relates to an antenna optimization design method based on a semi-supervised width learning system. Background Art

[0002] When using a proxy model to model and optimize the antenna design, the first task is to simulate the antenna size parameters in electromagnetic simulation software such as HFSS to obtain data for training and testing the proxy model. Traditionally, the method to obtain this data is to continuously modify the antenna size parameters in HFSS and simulate, which is a time-consuming and complicated task, especially when solving the optimization design problem of antennas with complex structures. Due to the significant increase in the dimension of the size parameters, the simulation software has higher requirements for computer hardware and is more time-consuming.

[0003] In the patent application with application number: 202310629245.X and application name: Dual-network sample expanded antenna design optimization method based on dual-fidelity evaluation, a dual-network sample expanded antenna design optimization method based on dual-fidelity evaluation is disclosed. The BP neural network architecture is adopted to design a dual neural network structure of regression prediction neural network and classification neural network. The data is comprehensively evaluated in two dimensions of global fidelity and local fidelity, normalized and weighted respectively, and the data that meets the requirements is used as the expanded data source. The neural network is retrained to realize the modeling of a planar printed antenna. The dual networks of this patent are a regression network and a classification network. The examples are all shallow BP neural network architectures, and the amount of data to be evaluated is large. Especially when there are many structural parameters of the antenna, the accuracy and efficiency of network modeling are limited. Summary of the invention

[0004] In order to overcome the above technical defects, the present invention provides an antenna optimization design method and device based on a semi-supervised width learning system, which can reduce the computational cost in the antenna optimization design process.

[0005] The present invention is achieved through the following scheme:

[0006] An antenna optimization design method based on a semi-supervised width learning system comprises the following steps:

[0007] Step 1: construct a structural model of the antenna to be optimized and set the size range that needs to be optimized for the antenna to be optimized;

[0008] Step 2: Generate a data set consisting of antenna size parameters and corresponding return loss values; divide the data set into an initial training set, a test set, and an unlabeled sample data set;

[0009] Step 3: Use the initial training set to train the convolution-width learning system;

[0010] Step 4: Use the initial training set to train the convolution-stack width learning system;

[0011] Step 5: Input the unlabeled sample data set into the trained convolution-width learning system and the trained convolution-stacked width learning system to predict and obtain pseudo-labeled data, and add the pseudo-labels to the labeled sample training data set;

[0012] Step 6. Use the updated labeled sample training data set of the convolution-width learning system and the updated labeled sample training data set of the convolution-stacked width learning system to cross-train and collaboratively train the convolution-width learning system and the convolution-stacked width learning system until a learning system that meets the set rules is obtained.

[0013] As a further improvement of the present invention, after step 6, the following steps are further included:

[0014] Step 7: Optimize the design of the antenna based on the learning system that meets the set rules, simulate the antenna structure, compare the prediction results of the learning system with the simulation results, verify the performance of the learning system, and obtain the final design result of the antenna.

[0015] As a further improvement of the present invention, the step five comprises:

[0016] Selecting a number of data from the unlabeled sample data set and inputting them into the trained convolution-width learning system for prediction to obtain first pseudo-labeled data, adding the first pseudo-labeled data to the first labeled sample training set, and obtaining a first sample training set updated by the convolution-width learning system;

[0017] Select some data from the unlabeled sample data set and input them into the trained convolution-stack width learning system for prediction to obtain second pseudo-labeled data, and add the second pseudo-labeled data to the second labeled sample training set to obtain the second sample training set updated by the convolution-stack width learning system.

[0018] As a further improvement of the present invention, the step of cross-training the convolution-width learning system and the convolution-stack width learning system using the labeled sample training data set updated by the convolution-width learning system and the labeled sample training data set updated by the convolution-stack width learning system includes:

[0019] The convolution-width learning system is trained using the second sample training set to obtain third pseudo-labeled data, and the convolution-width learning system is tested using the test set to obtain a first error between the two;

[0020] The convolution-stack width learning system is trained using the first sample training set to obtain fourth pseudo-labeled data, and the convolution-stack width learning system is tested using the test set to obtain a second error between the two;

[0021] Add the pseudo-labeled data with smaller errors to the initial training set, delete this part of data from the unlabeled sample data set, and update the unlabeled sample data set.

[0022] As a further improvement of the present invention, the steps of collaboratively training the convolution-width learning system and the convolution-stack width learning system using the labeled sample training data set updated by the convolution-width learning system and the labeled sample training data set updated by the convolution-stack width learning system include:

[0023] The convolution-width learning system is trained using the sample training set that has been updated through cross-training, and the convolution-width learning system is tested using the test set to obtain the third error between the two.

[0024] The convolution-stack width learning system is trained using the sample training set that has been updated through cross-training, and the convolution-stack width learning system is tested using the test set to obtain the fourth error between the two;

[0025] If the smaller value of the third error and the fourth error meets the set rules, the training is completed.

[0026] As a further improvement of the present invention, in step three, the convolution operator of the convolutional neural network is used to extract features from the initial training set, and the extracted data features are fused with the original initial training set to form enhanced data, which is input into the width learning system to obtain the convolution-width learning system.

[0027] As a further improvement of the present invention, in step four, the convolution operator of the convolutional neural network is used to extract features from the initial training set, and the extracted data features are fused with the original initial training set to form enhanced data, which is input into the width learning system. Then, according to the set number of feature windows, feature nodes, and enhanced nodes of the width learning system, feature nodes and enhanced nodes are generated in sequence, and a feature node layer and an enhanced node layer are generated. The output of the output layer of the underlying width learning system is calculated using the grid search method and the ridge regression algorithm, and the underlying output prediction label is recorded; the stacking layer is calculated, that is, the underlying output is used as the input of the upper layer network, and a new width learning system is constructed. The expected output is the labeled data of the original input sample data, so as to obtain a convolution-stacked width learning system.

[0028] As a further improvement of the present invention, if the set rules are not satisfied after cross-training and collaborative training, the process returns to step three.

[0029] The present invention also provides a computer device, including a processor and a storage device, wherein the storage device stores a program code, and the processor executes the program code to execute the above-mentioned antenna optimization design method.

[0030] The present invention also provides a computer-readable storage medium, characterized in that at least one instruction is stored in the computer-readable storage medium, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned antenna optimization design method.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows: compared with the traditional neural network training which requires a large amount of sample data, the present invention adopts the convolution-width learning system and the convolution-stacked width learning system, which are used as proxy models to process antenna size parameters and performance such as return loss, which has the advantage of being more efficient. In this process, the convolution-width learning system and the convolution-stacked width learning system are used as two different learners to train each other. The use of the collaborative training method can make full use of cheap unlabeled data, greatly saving the computational cost in the antenna optimization design process. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings, wherein:

[0033] Figure 1 A flow chart of the antenna optimization design method of the present invention;

[0034] Figure 2 This is the HFSS modeling diagram of the SIW aperture-coupled microstrip antenna;

[0035] Figure 3 It is the structural diagram of SIW aperture coupled microstrip antenna;

[0036] Figure 4 It is a return loss curve fitting diagram of the SIW aperture coupled microstrip antenna optimized by the antenna optimization design method of the present invention;

[0037] Figure 5 It is a schematic diagram of the structure of the computer device described in the present invention. DETAILED DESCRIPTION

[0038] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0039] It should be noted that similar reference numerals and letters represent similar items in the following figures, so once an item is defined in one figure, it does not need to be further defined and explained in the subsequent figures. At the same time, in the description of the present invention, the serial numbers of each step are only used to distinguish between steps, and do not mean that each step needs to be strictly executed in the order of the serial numbers.

[0040] Semi-supervised learning (SSL) makes full use of unlabeled sample data for continuous learning and training, which can solve the confusion of not being able to make full use of these data when processing unlabeled data. Therefore, the inventive concept of the present invention is to introduce the idea of ​​semi-supervised learning SSL to help the optimal design of antennas.

[0041] The present invention provides an antenna optimization design method based on a semi-supervised width learning system, such as Figure 1 As shown, the steps include:

[0042] Step 1: construct a structural model of the antenna to be optimized, and set the size range that needs to be optimized for the antenna to be optimized; for example, the structural model of the antenna to be optimized can be constructed in electromagnetic simulation software HFSS.

[0043] Step 2, generate a data set consisting of antenna size parameters and corresponding return loss values; divide the data set into an initial training set, a test set, and an unlabeled sample data set. Exemplarily, call the HFSS-MATLAB-API joint simulation program to generate a small number of data sets consisting of antenna size parameters and corresponding return loss values, separate them according to a certain ratio, and use them as the initial training set and test set respectively; use the remaining data without generating labels as the unlabeled sample data set, in which the antenna size parameters are used as input and the return loss values ​​corresponding to the antenna size parameters are used as output. Since it is an unlabeled sample, the computational cost of this part of the data set can be ignored. The main computational cost of this step is the time required for the HFSS-MATLAB-API joint simulation, and the time to obtain the unlabeled sample data set can be ignored.

[0044] Step 3: Use the initial training set to train the convolution-width learning system C-BLS to obtain the trained convolution-width learning system C-BLS trained0 ; The convolution operator of the convolutional neural network is used to extract features from the initial training set, and the extracted data features are fused with the original initial training set to form enhanced data, which is input into the width learning system to obtain the convolution-width learning system. Based on this, the feature extraction capability of the convolutional neural network can be fully utilized, and the enhanced data obtained is conducive to improving the modeling accuracy.

[0045] Step 4: Use the initial training set to train the convolution-stack width learning system CS-BLS to obtain the trained convolution-stack width learning system CS-BLS trained0 ; Use the convolution operator of the convolutional neural network to extract features from the initial training set, and fuse the extracted data features with the original initial training set to form enhanced data, which is input into the width learning system. Then, according to the number of feature windows, feature nodes, and enhanced nodes of the width learning system, feature nodes and enhanced nodes are generated in sequence, and feature node layers and enhanced node layers are generated. The output of the bottom-level width learning system output layer is calculated using the grid search method and the ridge regression algorithm, and the bottom-level output prediction label is recorded; perform stacking layer calculations, that is, use the bottom-level output as the input of the previous layer of the network, and construct a new width learning system. The expected output is the labeled data of the original input sample data, so as to obtain the convolution-stacked width learning system. The given stacked width learning system form can further improve the model accuracy.

[0046] Step 5: Input the unlabeled sample data set into the trained convolution-width learning system C-BLS trained0 , trained convolution-stack width learning system CS-BLS trained0 Make predictions to obtain pseudo-labeled data, and add the pseudo-labels to the labeled sample training data set.

[0047] Specifically, some data are selected from the unlabeled sample data set and input into the trained convolution-width learning system C-BLS. trained0 The first pseudo-labeled data is predicted and added to the first labeled sample training set to obtain the first sample training set Train updated by the convolution-width learning system. C-BLS ;

[0048] Select some data from the unlabeled sample data set and input it into the trained convolution-stack width learning system CS-BLS trained0 The second pseudo-labeled data is predicted and added to the second labeled sample training set to obtain the second sample training set Train updated by the convolution-stack width learning system. CS-BLS .

[0049] Step 6: Use Convolution-Breadth Learning System C-BLS trained0 Updated labeled sample training dataset Train C-BLS , Convolution-Stack Width Learning System CS-BLS trained0 Updated labeled sample training dataset Train CS-BLS Convolution-Breadth Learning System C-BLS trained0 , Convolution-Stack Width Learning System CS-BLStrained0 Carry out cross-training and collaborative training until a learning system that meets the set rules is obtained.

[0050] Cross-Training:

[0051] Use the second sample training set Train CS-BLS Convolution-Breadth Learning System CS-BLS trained0 The third pseudo-labeled data is obtained by training, and the convolution-width learning system is tested with the test set to obtain the first error RMSE between the two. CS-BLS0 .

[0052] Use the first sample training set Train C-BLS Convolution-Stack Width Learning System CS-BLS trained0 The fourth pseudo-labeled data is obtained by training, and the convolution-stack width learning system is tested using the test set to obtain the second error RMSE between the two. C-BLS0 ;

[0053] Compare the first error RMSE CS-BLS0 The second error RMSE C-BLS0 The size of the first error RMSE is used to add pseudo-labeled data with smaller errors to the initial training set, and delete this part of data from the unlabeled sample data set to update the unlabeled sample data set. CS-BLS0 If the value of is smaller, the third pseudo-labeled data is added to the initial training set, the labeled sample training set is updated, the selected data is deleted from the unlabeled sample data set, and the unlabeled sample data set is updated.

[0054] Collaborative training:

[0055] The convolution-width learning system C-BLS is trained using a labeled sample training set that has been updated through cross-training. trained0 Train and use the test set to train the convolution-width learning system C-BLS trained0 Test and get the third error RMSE between the two C-BLS1 .

[0056] The convolution-stack width learning system CS-BLS is trained using the updated labeled sample training set trained0 Train and use the test set to train the convolution-stack width learning system CS-BLS trained0 Test and get the fourth error RMSE between the two CS-BLS1 .

[0057] If the third error RMSE C-BLS1 , the fourth error RMSE CS-BLS1If the smaller value in satisfies the set rule, the training is completed, and the set rule is the preset expected error value.

[0058] If the set rules are not met after cross-training and collaborative training, return to step three until a learning system that meets the expected value error appears.

[0059] Step 7: Optimize the design of the antenna based on the learning system that meets the set rules, and simulate the antenna structure using the electromagnetic simulation software HFSS. Compare the predicted results of the learning system with the simulation results to verify the performance of the learning system and obtain the final design result of the antenna.

[0060] Next, the present invention is further explained by taking the SIW aperture coupled antenna as an example, as follows:

[0061] When optimizing the design of SIW aperture-coupled microstrip antennas using traditional supervised learning modeling, the required labeled sample data set is relatively large due to its complex structure, which causes the problem of long calculation time. The present invention utilizes an improved model of a width learning system BLS: a convolution-width learning system C-BLS and a convolution-stacked width learning system CS-BLS, introduces the idea of ​​collaborative training to realize a semi-supervised width learning system based on the convolution-width learning system C-BLS and the convolution-stacked width learning system CS-BLS, and applies it to the optimization design of SIW aperture-coupled microstrip antennas. On the one hand, the feature extraction capability of the convolutional neural network CNN and the efficient learning capability of the width learning system BLS are fully utilized. On the other hand, based on the collaborative semi-supervised learning idea, cheap unlabeled data is fully utilized, the time required to collect labeled data during antenna model simulation training is reduced, and the optimization efficiency of the antenna is improved.

[0062] The SIW aperture coupling antenna is composed of two dielectric layers, the microstrip patch is located on the upper dielectric plate, and a gap is opened on the waveguide surface between the two dielectric plates to realize the coupling feeding function. The electromagnetic simulation software HFSS modeling diagram and top view structure diagram are shown in the figure. Figure 2 , Figure 3 The design index of the antenna is: when the operating frequency is 28.5GHz, its return loss value S 11 Less than -15dB, -10dB bandwidth is about 10GHz.

[0063] Step 1: Model the SIW aperture-coupled antenna in the electromagnetic simulation software HFSS. Set the dielectric layer material of the antenna to RT Duroid 5870, the thickness of the upper and lower layers of the substrate is 1.575mm, the length is 16mm, and the width is 10mm. The dielectric constant is 2.33, and the loss tangent is 0.002. Figure 2 and Figure 3 shown.

[0064] Step 2: As shown in Table 1, determine the input and output of the semi-supervised width learning system. The seven size variables X=[d, s, ys, L slot ,W slot ,L mpa ,W mpa ] as input, the size parameter names and their corresponding optimization design interval values ​​are listed in Table 1, and the output is the corresponding return loss parameter sweep value at 20.5GHz~37.5GHz. The sweep interval is 0.2GHz, so there are 86 values ​​in total, recorded as According to the range of size variable parameters, 200 sets of data are generated according to the Latin hypercube sampling method, of which 100 sets of data are used to obtain the corresponding return loss values ​​through the HFSS-MATLAB-API joint simulation program to form a labeled sample data set, and the remaining 100 sets are used as unlabeled sample data sets. Then 70 sets of the labeled sample data set are used as training samples, and the remaining 30 sets are used as test samples.

[0065] Table 1 Dimensional parameters that need to be optimized for SIW aperture-coupled microstrip antennas

[0066]

[0067] Step 3: Initial training of the convolution-width learning system C-BLS. Use the initial training set to perform preliminary training on the convolution-width learning system C-BLS. 70 groups of sample data with a dimension of 1×7 are transformed into 1×3 after two layers of 1×3 convolution kernels, and the number of channels is set to 5. After fusion and enhancement with the original input data, 70 groups of 1×10 enhanced sample input data are obtained. The number of windows is set to [1,30], [1,30], [1,30], and the sparse regularization parameter is 2. -30 , the reduction parameter of the enhanced node is 0.8, and the trained convolution-width learning system C-BLS is obtained trained0 .

[0068] Step 4: Initial training of the convolution-stack width learning system CS-BLS. Use the initial training set to perform preliminary training on the convolution-stack width learning system CS-BLS. Similarly, 70 groups of sample data with a dimension of 1×7 are transformed into 1×3 after passing through two layers of 1×3 convolution kernels, and the number of channels is set to 5. After fusion and enhancement with the original input data, 70 groups of 1×10 enhanced sample input data are obtained. Set the number of bottom-level windows to [1,30], [1,30], [1,30], and the number of upper-level windows to [1,20], [1,20], [1,20], and the sparse regularization parameter to 2. -30 , the reduction parameter of the enhanced node is 0.8. The trained convolution-width learning system CS-BLS is obtained trained0 .

[0069] Step 5: Preliminary update of the training set. Randomly select 10 groups from the unlabeled sample data set and put them into the convolution-width learning system C-BLS trained0 and Convolution-Stack Width Learning System CS-BLS trained0 Make predictions, get the corresponding pseudo labels, and add these data to the labeled sample training data set to get the updated training sample set Train. C-BLS and Train CS-BLS , the dataset size is now 80.

[0070] Step 6: Update the sample set. Use the updated training sets to interactively train the model, that is, use the second sample training set Train CS-BLS Convolution-Breadth Learning System C-BLS trained0 For training, use the first sample training set Train C-BLS Convolution-Stack Width Learning System CS-BLS trained0 After training, the test set is used for testing to obtain the first error RMSE CS-BLS0 and the second error RMSE C-BLS0 , compare the sizes of the two errors, add the pseudo-labeled data generated by the model with the smaller error value to the initial training set, update the labeled sample training set, delete the selected data from the unlabeled sample data set, and update the unlabeled sample data set. At this time, the updated labeled sample training set size is 80, and the updated unlabeled sample data set size is 90.

[0071] Co-training: Use the updated labeled sample training set to train the convolution-width learning system C-BLS again trained0 and Convolution-Stack Width Learning System CS-BLS trained0 After training, we get the test errors of the two learning systems: the third error RMSE C-BLS1 And the fourth error RMSE CS-BLS1 ; Determine the third error RMSE C-BLS1 And the fourth error RMSE CS-BLS1 Whether the smaller value in meets the expected error value set by the system; if not, return to step 3 and repeat to obtain the root mean square error RMSE of the convolution-width learning system. C-BLSi and convolution-stack width learning system root mean square error RMSE CS-BLSi , until RMSE C-BLSi or RMSE CS-BLSiA model that meets the expected error value appears in the training, and the training is completed. The updated labeled sample training set size is 70+10*i, where i represents the number of times unlabeled data is added. At this time, the updated unlabeled sample data set size is 100-10*i. This process makes full use of unlabeled data, so that the overall performance of the system can be improved through coordinated training at a lower computational cost.

[0072] Step 7: Model verification and antenna optimization. The antenna is optimized and designed based on the trained model. The specific dimensions are shown in Table 2. It is then simulated in the electromagnetic simulation software HFSS. The comparison between the model prediction results and the electromagnetic simulation software HFSS simulation results is shown in the figure below. Figure 4 As shown, “FZ HFSS” represents the HFSS simulated return loss curve, and “Co-SSBLS” is the return loss curve fitted by the model. At 28.5GHz, the return loss value is lower than -15dB, and the -10dB bandwidth reaches 10GHz, which meets the design requirements.

[0073] Table 2 Dimensional parameter values ​​of optimized SIW aperture-coupled microstrip antenna

[0074]

[0075]

[0076] Compared with the traditional neural network training which requires a large amount of sample data, the present invention adopts the convolution-width learning system and the convolution-stacked width learning system, which are used as proxy models to process antenna size parameters and performance, such as small sample data such as return loss, and have the advantage of being more efficient. Using the idea of ​​the collaborative training method in semi-supervised learning SSL, on the one hand, the convolutional neural network CNN used has a strong feature extraction capability, and the semi-supervised learning SSL used has an efficient learning capability; on the other hand, the convolution-width learning system and the convolution-stacked width learning system are used as two different learners for mutual training. The use of the collaborative training method can make full use of cheap unlabeled data, greatly saving the computational cost in the antenna optimization design process.

[0077] The present invention also provides a computer device, such as Figure 5 As shown, it includes a processor and a storage, the storage stores program codes, and the processor executes the program codes to perform the above antenna optimization design method.

[0078] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the embodiments of the present application can be implemented with hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein the communication media include any media that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that a general or special-purpose computer can access.

[0079] This embodiment provides a computer-readable storage medium, in which at least one instruction is stored. The at least one instruction is loaded and executed by a processor to implement the above antenna optimization design method.

[0080] Optionally, the computer readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a solid state drive (SSD), or an optical disk, etc. Among them, the random access memory may include a resistance random access memory (ReRAM) and a dynamic random access memory (DRAM).

[0081] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An antenna optimization design method based on a semi-supervised width learning system, characterized in that: Includes steps: Step 1: construct a structural model of the antenna to be optimized and set the size range that needs to be optimized for the antenna to be optimized; Step 2: Generate a data set consisting of antenna size parameters and corresponding return loss values; Divide the dataset into an initial training set, a test set, and an unlabeled sample dataset; Step 3: Using the initial training set to train the convolution-width learning system, including: using the convolution operator of the convolutional neural network to extract features from the initial training set, and fusing the extracted data features with the initial training set to form enhanced data, which is input into the width learning system to obtain the convolution-width learning system; Step 4: Use the initial training set to train the convolution-stack width learning system; Step 5: Input the unlabeled sample data set into the trained convolution-width learning system and the trained convolution-stacked width learning system for prediction to obtain pseudo-labeled data, and add the pseudo-labeled data to the labeled sample training data set, including: Selecting a number of data from the unlabeled sample data set and inputting them into the trained convolution-width learning system for prediction to obtain first pseudo-labeled data, adding the first pseudo-labeled data into the first labeled sample training set, and obtaining a first sample training set updated by the convolution-width learning system; selecting a number of data from the unlabeled sample data set and inputting them into the trained convolution-stacked width learning system for prediction to obtain second pseudo-labeled data, adding the second pseudo-labeled data into the second labeled sample training set, and obtaining a second sample training set updated by the convolution-stacked width learning system; Step 6: Use the updated labeled sample training data set of the convolution-width learning system and the updated labeled sample training data set of the convolution-stacked width learning system to perform cross-training and collaborative training on the convolution-width learning system and the convolution-stacked width learning system until a learning system that meets the set rules is obtained; The step of cross-training the convolution-width learning system and the convolution-stack width learning system using the labeled sample training data set updated by the convolution-width learning system and the labeled sample training data set updated by the convolution-stack width learning system includes: The convolution-width learning system is trained using the second sample training set to obtain third pseudo-labeled data, and the convolution-width learning system is tested using the test set to obtain a first error between the two; The convolution-stack width learning system is trained using the first sample training set to obtain fourth pseudo-labeled data, and the convolution-stack width learning system is tested using the test set to obtain a second error between the two; Add the pseudo-labeled data with smaller errors to the initial training set, update the labeled sample training set, delete this part of data from the unlabeled sample data set, and update the unlabeled sample data set; The step of collaboratively training the convolution-width learning system and the convolution-stack width learning system using the labeled sample training data set updated by the convolution-width learning system and the labeled sample training data set updated by the convolution-stack width learning system includes: The convolution-width learning system is trained using the labeled sample training set that has been updated through cross-training, and the convolution-width learning system is tested using the test set to obtain the third error between the two. The convolution-stack width learning system is trained using the labeled sample training set that has been updated through cross-training, and the convolution-stack width learning system is tested using the test set to obtain the fourth error between the two; If the smaller value of the third error and the fourth error satisfies the set rule, the training is completed; if the set rule is not satisfied after cross-training and collaborative training, the process returns to step three; In the step four, the convolution operator of the convolutional neural network is used to extract features from the initial training set, and the extracted data features are fused with the initial training set to form enhanced data, which is input into the width learning system. Then, according to the set number of feature windows, feature nodes, and enhanced nodes of the width learning system, feature nodes and enhanced nodes are generated in sequence, and feature node layers and enhanced node layers are generated. The output of the output layer of the underlying width learning system is calculated using the grid search method and the ridge regression algorithm, and the underlying output prediction label is recorded; the stacking layer is calculated, that is, the underlying output is used as the input of the upper layer network, and a new width learning system is constructed. The expected output is the labeled data of the original input sample data, so as to obtain a convolution-stacked width learning system.

2. The antenna optimization design method according to claim 1, characterized in that: After step six, the method further includes the following steps: Step 7: Optimize the design of the antenna based on the learning system that meets the set rules, simulate the antenna structure, compare the prediction results of the learning system with the simulation results, verify the performance of the learning system, and obtain the final design result of the antenna.

3. A computer device, characterized in that: It comprises a processor and a storage, wherein the storage stores a program code, and the processor executes the program code to execute the antenna optimization design method according to claim 1 or 2.

4. A computer-readable storage medium, characterized in that: The readable storage medium stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the antenna optimization design method as claimed in claim 1 or 2.

Images