Cooperative antenna design method

Abstract

The invention discloses a collaborative antenna design method, which comprises the following steps of: establishing a joint simulation model according to design requirements of an antenna to be optimized; randomly selecting N groups of antenna size parameter combinations, and obtaining marked data samples and unmarked data samples; selecting a model training data set and a model test data set from the marked data samples; training a student T process regression device and a convolutional neural network regression device by using the model training data set to obtain an initial model; selecting a model test data set to evaluate the performance of the two models to obtain a high-quality model; part of unmarked data samples are selected and substituted into the high-quality model to obtain a prediction result, then the part of samples are added into a model training data set, a student T process regression device and a convolutional neural network regression device are trained again, and a cooperative semi-supervised agent model is established after continuous iterative training; after iteration is ended, the model test data set is used for testing; and performing optimization and simulation by adopting a global optimization algorithm to obtain antenna performance indexes, and completing optimization design.

Description

Technical Field

[0001] The present invention belongs to the technical field of antenna design, and in particular relates to a collaborative antenna design method. Background Art

[0002] The design of microwave engineering antennas is primarily influenced by multiple factors, including the upgrading of communication needs, innovations in materials and processes, developments in electromagnetic theory and simulation technology, the need for multi-scenario applications, and the improvement of testing and standardization systems. With the development of wireless communication technology, low-frequency spectrum resources are becoming increasingly scarce. Traditional antenna designs are unable to meet the requirements of wide bandwidth, high gain, and controllable directionality, prompting the development of microwave antennas towards high frequency bands, miniaturization, and multi-bands. Modern communication systems are constantly increasing the requirements for data transmission rate and capacity. As key components for signal transmission and reception, microwave antennas must possess high gain, low loss, and wide bandwidth to support high-speed, high-capacity data transmission.

[0003] In the field of antenna design, a global algorithm based on swarm intelligence is generally used in conjunction with full-wave electromagnetic simulation software, where the full-wave electromagnetic simulation software serves as the fitness evaluation function for the global optimization algorithm. This approach is universal, but its major drawback is its high time complexity, making it difficult to implement in practical antenna engineering. An effective solution is to replace the electromagnetic simulation software with a proxy model, which can significantly reduce time complexity. Therefore, in the initial stages of antenna optimization design, it is crucial to create a high-precision proxy model. However, this process often requires collecting data samples for model training and validation, which can generally be obtained using high-precision electromagnetic simulation tools. However, the prominent problem with this task is that it is time-consuming, resulting in low efficiency. Summary of the Invention

[0004] In response to the problems existing in the existing technology, the present invention provides a collaborative antenna design method, a collaborative agent model construction method based on convolutional neural networks and Student's T process, and improves the antenna modeling accuracy through iterative training, providing an efficient approach for antenna design and optimization.

[0005] The technical solution of the present invention is achieved as follows:

[0006] A collaborative antenna design method comprises the following steps:

[0007] S1. Based on the design requirements of the antenna to be optimized, a joint simulation model is established using Python programming language and 3D electromagnetic simulation software.

[0008] S2. Randomly select N groups of antenna size parameter combinations, and obtain N0 groups of labeled data samples according to the joint simulation model, and the remaining N-N0 groups are unlabeled data samples;

[0009] S3, select N from N0 group of labeled data samples T group as the model training data set, and the rest N0-N T group as the model testing dataset;

[0010] S4. Using N T The model training dataset is used to train the Student T Process regressor and the Convolutional Neural Network regressor respectively. The fully connected part of the Convolutional Neural Network regressor is replaced by the Student T Process, which is expressed as CNN-STP. After training, the STP model and CNN-STP model are obtained.

[0011] S5, using N0-N T The performance of the trained STP model and CNN-STP model is evaluated by using a set of model test data sets. If the evaluation error meets the evaluation threshold requirement, a collaborative semi-supervised proxy model of the antenna is established; otherwise, N unlabeled data samples are selected from N-N0 groups. t Bring a group of samples into the model with smaller evaluation error to obtain the corresponding prediction results. t Group prediction results and N T The group model training data sets are combined to form a new training data set, and the STP model and CNN-STP model are retrained until the set evaluation threshold requirements are met;

[0012] S6. After the iteration is terminated, the two models are trained using the updated training data set, and then the N0-N T Set up a model test data set to test the model and verify the performance of the model;

[0013] S7. Use a global optimization algorithm to obtain the antenna size parameters that meet the design requirements after optimization, use three-dimensional electromagnetic simulation software to simulate and obtain the antenna performance indicators, and complete the optimization design.

[0014] Furthermore, in S1, according to the design requirements of the antenna to be optimized, the design variables of the antenna to be optimized are obtained, and the design variables of the antenna to be optimized include the length of the excitation array, the distance between the excitation array and the reflector, the distance between the excitation array and the director, and the distance between the directors; then, a joint simulation model is established using the Python programming language and three-dimensional electromagnetic simulation software.

[0015] Furthermore, in S2, the Latin hypercube sampling method is used to randomly select N groups of antenna size parameter combinations, and N0 groups of electromagnetic performance response values are obtained according to the joint simulation model, and each group of electromagnetic performance response selects 1 frequency point.

[0016] Furthermore, in S4, the fully connected layer of the convolutional neural network regressor is replaced by the Student T process, which is expressed as CNN-STP; T The model training data set is used to train the student T process regressor to obtain the STP model and the corresponding mean square error; T The group model training data set is used to train the CNN-STP model to obtain the CNN-STP model and the corresponding mean square error.

[0017] Furthermore, in said S5, if the evaluation error does not meet the evaluation threshold requirement, then N unlabeled data samples are selected from N-N0 groups. t A group of unlabeled data samples are brought into the STP model or CNN-STP model with small evaluation error after training to make predictions, and the corresponding prediction results are obtained and compared with N T The training data sets of the two models are combined to form a new training data set, and the STP model and CNN-STP model are trained again, using the N0-N T The performance of the trained STP model and CNN-STP model is evaluated using a model test dataset. The performance of the model is evaluated using the mean square error:

[0018] If the mean square error of the model with the smaller test error meets the set evaluation threshold requirement, the collaborative semi-supervised agent model of the antenna is obtained;

[0019] If the mean square error of the model with smaller test error does not meet the set evaluation threshold requirement, the prediction result obtained by the model prediction output is compared with N T The group model training data sets are combined to form a new training data set for iterative training until the minimum test mean square error of the two models meets the set evaluation threshold range, and a collaborative semi-supervised agent model of the antenna is established; the evaluation threshold is the mean square error threshold.

[0020] Furthermore, in the S5, N0-N T The performance of the trained STP model and CNN-STP model is evaluated by the group model test data set. The specific method is as follows: T The trained STP model and CNN-STP model are tested on the group model test data set, and the corresponding mean square errors are calculated and obtained respectively. The two mean square errors are compared with the set mean square error threshold to determine whether the mean square error meets the set mean square error threshold range;

[0021] If the mean square error meets the set mean square error threshold range, a collaborative semi-supervised agent model of the antenna is established.

[0022] Furthermore, in S5, if the mean square error does not meet the set mean square error threshold range, the mean square errors corresponding to the two models are compared, and the model with the smaller mean square error is selected as the high-quality model, and N is selected. t A set of unlabeled data samples are brought into the high-quality model to generate a pseudo-labeled dataset, and the pseudo-labeled dataset is merged into N T A model training data set is formed to obtain an updated model training data set, and the STP model and the CNN-STP model are trained again using the updated model training data set to obtain the trained STP model and the CNN-STP model.

[0023] Furthermore, in said S5, according to N0-N T The group model test data set is used to evaluate the trained STP model and CNN-STP model respectively, and the corresponding mean square error is calculated respectively. The mean square error corresponding to the updated model is compared with the set mean square error threshold range, and it is judged again whether the mean square error meets the set mean square error threshold range.

[0024] Furthermore, in S6, all pseudo-labeled datasets after the iteration is terminated are added to the initial N T The model training data set is collected to obtain the completed model training data set, and the STP model and CNN-STP model are trained respectively to obtain the completed STP model and CNN-STP model. T The trained STP model and CNN-STP model were tested using the group's model test dataset to verify the performance of the model and obtain a collaborative model.

[0025] Furthermore, in S7, a particle swarm algorithm is used as a global optimization algorithm, and a collaborative model is used as a fitness evaluation function of the antenna for optimization design. The optimization design results are simulated and verified using three-dimensional electromagnetic simulation software to obtain antenna performance indicators and complete the optimization design.

[0026] Compared with the prior art, the present invention achieves the following beneficial effects:

[0027] The present invention provides a collaborative antenna design method that uses semi-supervised learning technology to assist in the process of creating a model and improve modeling efficiency. In the semi-supervised learning framework, only a small number of labeled data samples are required in combination with a large number of unlabeled data samples to perform effective model training, which significantly reduces the consumption of time and resources. This strategy not only reduces the workload in the data sample preparation stage, but also improves the training effect of the model under conditions where data samples are limited. The introduction of semi-supervised learning methods into the field of antenna optimization design relies on a small amount of known information to guide the model. By continuously predicting and utilizing unlabeled data to improve its own performance, the model achieves effective self-improvement. This method is particularly suitable for application scenarios with limited resources or that require rapid development.

[0028] A collaborative semi-supervised learning method is implemented through convolutional neural networks and student T processes. By exploring and utilizing the differences between two different models, the understanding of data sample characteristics is deepened. Through the collaborative work of the two models, the accuracy of model prediction and the robustness of the system are improved. At the same time, by effectively tapping the potential of unlabeled data, the dependence on electromagnetic simulation resources is reduced, the design process is accelerated, and the maximum efficiency is achieved under limited resources, providing an efficient approach for antenna design and optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a flow chart of a collaborative antenna design method provided by the present invention;

[0030] Figure 2 Schematic diagram of an antenna to be modeled in a collaborative antenna design method provided by the present invention;

[0031] Figure 3 This is a HFSS model diagram of an antenna to be modeled in a collaborative antenna design method provided by the present invention;

[0032] Figure 4 This is a schematic diagram of modeling results of a collaborative antenna design method provided by the present invention. DETAILED DESCRIPTION

[0033] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0034] Example

[0035] like Figures 1 to 4, a collaborative antenna design method, comprising the following steps:

[0036] S1. Based on the design requirements of the antenna to be optimized, a joint simulation model is established using the Python programming language and the 3D electromagnetic simulation software HFSS.

[0037] According to the design requirements of the microstrip Yagi antenna to be optimized, the design variables of the antenna to be optimized are obtained. The design variables of the antenna to be optimized include the length of the excitation array, the distance between the excitation array and the reflector, the distance between the excitation array and the director, and the distance between the directors. Then, a joint simulation model is established using the Python programming language and three-dimensional electromagnetic simulation software.

[0038] The microstrip Yagi antenna combines the compact integration advantages of the microstrip antenna with the wideband characteristics of the traditional Yagi antenna, providing an efficient and easy-to-integrate wireless communication solution. Considering that the parameters that have a greater impact on the performance of the Yagi antenna are mainly: the length d of the excitation array r , the distance between the excitation array and the reflector g1, the distance between the excitation array and the director g2, and the distance between the directors g3, so the four parameters X = [d r Using the Python programming language and the electromagnetic simulation software HFSS, a Python-HFSS co-simulation model was established to automatically obtain the antenna's electrical performance. The simulation frequency sweep range was set from 1.8 GHz to 2.2 GHz. The design specifications were a bandwidth of at least 400 MHz and a return loss of less than -10 dB at a center frequency of 2.0 GHz.

[0039] like Figure 2 The microstrip Yagi antenna shown is a schematic diagram of the antenna to be modeled.

[0040] S2. Randomly select N groups of antenna size parameter combinations, and obtain N0 groups of labeled data samples according to the joint simulation model, and the remaining (N-N0) groups are unlabeled data samples;

[0041] The Latin hypercube sampling method is used to randomly select N groups of antenna size parameter combinations, and N0 groups of electromagnetic performance response values are obtained according to the joint simulation model. That is, N0 groups of labeled data samples are obtained, and l frequency point is selected for each group of electromagnetic performance response.

[0042] Specifically, the Latin hypercube sampling method was used to randomly select 100 groups of antenna size parameter combinations, and 40 groups of electromagnetic performance response values were obtained based on the joint simulation model. The frequency sweep step of each group of electromagnetic performance response was 0.01 GHz, that is, 41 frequency sampling points were taken for each group of electrical performance, and the remaining 60 groups were unlabeled sample data.

[0043] S3, select N from N0 group of labeled data samples T The group is used as the model training data set, and the rest (N0-N T ) group as the model testing data set;

[0044] Specifically, 30 groups of 40 labeled data samples are selected as model training data sets, and the remaining 10 groups are used as model testing data sets.

[0045] S4. Using N T The model training dataset is used to train the student T process regressor and the convolutional neural network regressor respectively, and the STP model and CNN-STP model and the corresponding mean square error are obtained;

[0046] The fully connected layer of the convolutional neural network regressor is replaced by the Student T process, denoted as CNN-STP; T The model training data set is used to train the student T process regressor to obtain the STP model and the corresponding mean square error; T The convolutional neural network regressor is trained using the group model training data set to obtain the CNN-STP model and the corresponding mean square error.

[0047] Based on N T A training dataset was used to train a Student T-Process (STP) regressor and a convolutional neural network regressor. The fully connected part of the convolutional neural network regressor was replaced with a Student T-Process, representing a CNN-STP model. This model then lightweighted the CNN. After preliminary training, the initial STP model and CNN-STP model were obtained.

[0048] Specifically, the STP model and CNN-STP model were trained based on 30 sets of training data sets to obtain the initial STP model and CNN-STP model. The corresponding mean square errors are shown in Table 1:

[0049] Table 1

[0050]

[0051] S5, using N0-N T The performance of the trained STP model and CNN-STP model is evaluated by using a set of model test data sets. If the evaluation error meets the evaluation threshold requirement, a collaborative semi-supervised proxy model of the antenna is established; otherwise, N unlabeled data samples are selected from N-N0 groups. t Bring a group of samples into the model with smaller evaluation error to obtain the corresponding prediction results. t Group prediction results and N TThe group model training data sets are combined to form a new training data set to retrain the STP model and CNN-STP model until the set evaluation threshold requirements are met;

[0052] Furthermore, if the evaluation error does not meet the evaluation threshold requirement, N unlabeled data samples are selected from N-N0 groups. t A group of unlabeled data samples are brought into the STP model or CNN-STP model with small evaluation error after training for prediction, and the corresponding prediction results and N are obtained. T The model training data sets are combined to form a new training data set to retrain the STP model and CNN-STP model, and the N0-N T The performance of the trained STP model and CNN-STP model is evaluated using a model test dataset. The performance of the model is evaluated using the mean square error:

[0053] If the mean square error of the model with the smaller test error meets the set evaluation threshold requirement, the collaborative semi-supervised agent model of the antenna is obtained;

[0054] If the mean square error of the model with smaller test error does not meet the set evaluation threshold requirement, the prediction result obtained by the model prediction output is compared with N T The group model training data sets are combined to form a new training data set for iterative training until the minimum test mean square error of the two models meets the set evaluation threshold range, and a collaborative semi-supervised agent model of the antenna is established; the evaluation threshold is the mean square error threshold.

[0055] Furthermore, using N0-N T The performance of the trained STP model and CNN-STP model is evaluated by the group model test data set. The specific method is as follows: T The trained STP model and CNN-STP model are tested on the group model test data set, and the corresponding mean square errors are calculated and obtained respectively. The two mean square errors are compared with the set mean square error threshold to determine whether the mean square error meets the set mean square error threshold range;

[0056] If the mean square error meets the set mean square error threshold range, a collaborative semi-supervised agent model of the antenna is established.

[0057] Furthermore, if the mean square error does not meet the set mean square error threshold range, the mean square errors of the two models are compared, and the model with the smaller mean square error is selected as the high-quality model, and N is selected. t A set of unlabeled data samples are brought into the high-quality model to generate a pseudo-labeled dataset, and the pseudo-labeled dataset is merged into N TA model training data set is formed to obtain an updated model training data set, and the STP model and the CNN-STP model are trained again using the updated model training data set to obtain the trained STP model and the CNN-STP model.

[0058] Furthermore, according to N0-N T The group model test data set is used to evaluate the trained STP model and CNN-STP model respectively, and the corresponding mean square error is calculated respectively. The mean square error corresponding to the updated model is compared with the set mean square error threshold range, and it is judged again whether the mean square error meets the set mean square error threshold range.

[0059] For example, the STP model and the CNN-STP model are tested with 10 groups of test samples, and the corresponding mean square error is calculated. The performance of the model is judged based on the mean square error results.

[0060] The mean square error threshold is set to 0.6300. If the calculated corresponding mean square error is less than the mean square error threshold, the design requirements are met, that is, it is within the mean square error threshold range. Thus, a collaborative semi-supervised agent model of the antenna is established.

[0061] If the calculated mean square error is greater than the mean square error threshold, the mean square errors corresponding to the two models are compared, and the model with the smaller mean square error is selected as the high-quality model. Five groups of unlabeled samples are selected from the 60 groups of unlabeled samples to predict and generate corresponding pseudo labels. Subsequently, these five groups of pseudo-labeled data sets are incorporated into the original 30 groups of model training data sets to obtain an updated model training data set, i.e., 35 groups. The STP model and CNN-STP model are trained again using the updated model training data sets to obtain an updated STP model and CNN-STP model.

[0062] Then, 10 groups of test samples were used to test the updated STP model and CNN-STP model, and the corresponding mean square errors were calculated respectively. The mean square errors corresponding to the updated models were compared with the set mean square error threshold range, and it was determined again whether the mean square errors met the set mean square error threshold range.

[0063] Iterative training is performed in this way. Table 2 shows the error value at each iteration. The error obtained at the 8th iteration is 0.6266, which is less than the mean square error threshold of 0.6300 and meets the design requirements.

[0064] Table 2

[0065]

[0066] S6. After the iterative training is terminated, the two models are trained using the updated model training data set, and then the N0-N T The model test data set of the group is tested to verify the performance of the model;

[0067] Specifically, all pseudo-label datasets after the iteration are added to the initial N T The model training data set is collected to obtain the completed model training data set, and the STP model and CNN-STP model are trained respectively to obtain the completed STP model and CNN-STP model. T The trained STP model and CNN-STP model were tested using the group's model test dataset to verify the performance of the model and obtain a collaborative model.

[0068] For example, after the iteration is terminated, there are now 35 groups of pseudo-label data sets corresponding to unlabeled samples, which together with the initial 30 groups of model training data sets constitute the updated model training data set, that is, 65 groups. The STP model and CNN-STP model are trained separately to obtain the completed STP model and CNN-STP model. Then, 10 groups of model test data sets are used to test their performance, obtain the corresponding mean square error, verify the performance of the model, and obtain the collaborative model.

[0069] S7. Use a global optimization algorithm to obtain the antenna size parameters that meet the design requirements after optimization, use three-dimensional electromagnetic simulation software to simulate and obtain the antenna performance indicators, and complete the optimization design.

[0070] Specifically, the particle swarm algorithm is used as the global optimization algorithm, and the collaborative model obtained by S6 is used as the fitness evaluation function of the antenna for optimization design. The optimization design results are simulated and verified using the three-dimensional electromagnetic simulation software HFSS to obtain the antenna performance indicators and complete the optimization design.

[0071] like Figure 4 , where the solid line is the simulation result of the three-dimensional electromagnetic simulation software HFSS, and the dotted line is the prediction result of the collaborative model. The two are highly consistent, achieving the optimization design goal.

[0072] The present invention discloses a collaborative semi-supervised modeling method based on CNN and STP for the optimization design of antennas. In this process, the adoption of the collaborative semi-supervised method enables the model to have higher accuracy with fewer labeled training samples.

[0073] Based on the disclosure and teachings of the above description, those skilled in the art may also make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and modifications and variations of the present invention should also fall within the scope of protection of the claims of the present invention. In addition, although certain specific terms are used in this description, these terms are only for convenience of description and do not constitute any limitation to the present invention.

Claims

1. A collaborative antenna design method, characterized in that: The following steps are involved: S1. Based on the design requirements of the antenna to be optimized, a joint simulation model is established using Python programming language and 3D electromagnetic simulation software. S2. Randomly select N groups of antenna size parameter combinations, and obtain N0 groups of labeled data samples according to the joint simulation model, and the remaining N-N0 groups are unlabeled data samples; S3, select N from N0 group of labeled data samples T group as the model training data set, and the rest N0-N T group as the model testing dataset; S4. Using N T The model training dataset is used to train the Student T Process regressor and the Convolutional Neural Network regressor respectively. The fully connected part of the Convolutional Neural Network regressor is replaced by the Student T Process, which is expressed as CNN-STP. After training, the STP model and CNN-STP model are obtained. S5, using N0-N T The performance of the trained STP model and CNN-STP model is evaluated by using a set of model test data sets. If the evaluation error meets the evaluation threshold requirement, a collaborative semi-supervised proxy model of the antenna is established; otherwise, N unlabeled data samples are selected from N-N0 groups. t Bring a group of samples into the model with smaller evaluation error to obtain the corresponding prediction results. t Group prediction results and N T The group model training data sets are combined to form a new training data set to retrain the STP model and CNN-STP model until the set evaluation threshold requirements are met; S6. After the iteration is terminated, the two models are trained using the updated training data set, and then the N0-N T Set up a model test data set to test the model and verify the performance of the model; S7. Use a global optimization algorithm to obtain the antenna size parameters that meet the design requirements after optimization, use three-dimensional electromagnetic simulation software to simulate and obtain the antenna performance indicators, and complete the optimization design.

2. A collaborative antenna design method according to claim 1, characterized in that: In S1, the design variables of the antenna to be optimized are obtained according to the design requirements of the antenna to be optimized, and then a joint simulation model is established using Python programming language and three-dimensional electromagnetic simulation software.

3. The collaborative antenna design method according to claim 1, wherein: In S2, the Latin hypercube sampling method is used to randomly select N groups of antenna size parameter combinations, and N0 groups of electromagnetic performance response values are obtained according to the joint simulation model, and each group of electromagnetic performance response selects 1 frequency point.

4. The collaborative antenna design method according to claim 1, wherein: In S4, the fully connected layer of the convolutional neural network regressor is replaced by the Student T process, which is expressed as CNN-STP; T The model training data set is used to train the student T process regressor to obtain the STP model and the corresponding mean square error; T The group model training data set is used to train the CNN-STP model to obtain the CNN-STP model and the corresponding mean square error.

5. The collaborative antenna design method according to claim 1, wherein: In S5, if the evaluation error does not meet the evaluation threshold requirement, select N from the N-N0 groups of unlabeled data samples. t A group of unlabeled data samples are brought into the STP model or CNN-STP model with small evaluation error after training to make predictions, and the corresponding prediction results are obtained and compared with N T The training data sets of the two models are combined to form a new training data set, and the STP model and CNN-STP model are trained again, using the N0-N T The performance of the trained STP model and CNN-STP model is evaluated using a model test dataset. The performance of the model is evaluated using the mean square error: If the mean square error of the model with the smaller test error meets the set evaluation threshold requirement, the collaborative semi-supervised agent model of the antenna is obtained; If the mean square error of the model with smaller test error does not meet the set evaluation threshold requirement, the prediction result obtained by the model prediction output is compared with N T The group model training data sets are combined to form a new training data set for iterative training until the minimum test mean square error of the two models meets the set evaluation threshold range, and a collaborative semi-supervised agent model of the antenna is established; the evaluation threshold is the mean square error threshold.

6. The collaborative antenna design method according to claim 5, wherein: In the S5, N0-N T The performance of the trained STP model and CNN-STP model is evaluated by the group model test data set. The specific method is as follows: T The trained STP model and CNN-STP model are tested on the group model test data set, and the corresponding mean square errors are calculated and obtained respectively. The two mean square errors are compared with the set mean square error threshold to determine whether the mean square error meets the set mean square error threshold range; If the mean square error meets the set mean square error threshold range, a collaborative semi-supervised agent model of the antenna is established.

7. The collaborative antenna design method according to claim 6, wherein: In S5, if the mean square error does not meet the set mean square error threshold range, the mean square errors of the two models are compared, and the model with the smaller mean square error is selected as the high-quality model, and N is selected. t A set of unlabeled data samples are brought into the high-quality model to generate a pseudo-labeled dataset, and the pseudo-labeled dataset is merged into N T A model training data set is formed to obtain an updated model training data set, and the STP model and the CNN-STP model are trained again using the updated model training data set to obtain the trained STP model and the CNN-STP model.

8. The collaborative antenna design method according to claim 7, wherein: In the above S5, according to N0-N T The group model test data set is used to evaluate the trained STP model and CNN-STP model respectively, and the corresponding mean square error is calculated respectively. The mean square error corresponding to the updated model is compared with the set mean square error threshold range, and it is judged again whether the mean square error meets the set mean square error threshold range.

9. The collaborative antenna design method according to claim 8, wherein: In S6, all pseudo-labeled datasets after the iteration are added to the initial N T The model training data set is collected to obtain the completed model training data set, and the STP model and CNN-STP model are trained respectively to obtain the completed STP model and CNN-STP model. T The trained STP model and CNN-STP model were tested using the group's model test dataset to verify the performance of the model and obtain a collaborative model.

10. The collaborative antenna design method according to claim 9, wherein: In S7, a particle swarm algorithm is used as a global optimization algorithm, and a collaborative model is used as a fitness evaluation function of the antenna for optimization design. The optimization design results are simulated and verified using three-dimensional electromagnetic simulation software to obtain antenna performance indicators and complete the optimization design.

Images