Antenna design method, device, equipment, medium and product

Through the trained target prediction network and target reinforced deep learning model, the antenna design is automated, and the problems of inefficient antenna design and relying on manual experience in the existing technology are solved, and efficient and accurate antenna design is achieved.

CN119004964APending Publication Date: 2024-11-22LINKZHILIAN (CHONGQING) TECH CO LTD +2
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411025313.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing antenna design methods are inefficient and rely on manual experience, making it difficult to quickly and efficiently deal with designs with complex structures and high performance requirements, and there are problems that both design accuracy and efficiency are difficult to take into account.

Method used

The trained target prediction network and target reinforcement deep learning model are adopted to automatically output antenna sizes that meet design requirements by imitating antenna design ideas and optimization parameters, reducing dependence on professional qualities and improving design efficiency and accuracy.

Benefits of technology

The automated process of antenna design is realized, the design efficiency and accuracy are improved, and the dependence on designers' professional qualities is reduced, making antenna design more efficient, flexible and highly adaptable.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004964A_ABST
    Figure CN119004964A_ABST
Patent Text Reader

Abstract

The invention provides an antenna design method and device, equipment, a medium and a product, and the method comprises the steps: obtaining a trained target prediction network which is used for simulating a target antenna design thought to output a first antenna size meeting the target antenna design thought; obtaining a plurality of historical antenna sizes, and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size; according to the plurality of historical antenna sizes and the first antenna reflection coefficient and the first antenna gain corresponding to each historical antenna size, a trained target enhanced deep learning model is obtained, the active work network of the target enhanced deep learning model is a target prediction network, and the weight of the active work network is synchronized with the weight of the target prediction network; and inputting an antenna design demand into the target reinforcement deep learning model to obtain a target antenna size meeting the antenna design demand. According to the embodiment of the invention, the antenna design efficiency can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of communication technology, and in particular to an antenna design method, device, equipment, medium and product. Background Art

[0002] With the rapid advancement of wireless communication technology, antenna design, as a core element in the development of wireless communication systems, has a crucial impact on the optimization of the overall performance of communication systems and the improvement of market competitiveness. However, current antenna design mainly relies on manual operation and the deep experience of experts. This traditional design model is not only time-consuming and inefficient, but also highly dependent on the professional quality and skills of designers. Especially when dealing with antenna design with complex structures, this design method that relies on manual and expert experience is even more stretched and requires a lot of human resources, resulting in low efficiency of antenna design. Summary of the invention

[0003] The present application provides an antenna design method, device, equipment, medium and product, which can improve the efficiency of antenna design.

[0004] In a first aspect, an embodiment of the present application provides an antenna design method, the method comprising:

[0005] Obtaining a trained target prediction network, where the target prediction network is used to imitate a target antenna design idea to output a first antenna size that meets the target antenna design idea;

[0006] Acquire multiple historical antenna sizes, and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size;

[0007] According to multiple historical antenna sizes, and the first antenna reflection coefficient and the first antenna gain corresponding to each historical antenna size, a trained target reinforcement deep learning model is obtained, the main action network of the target reinforcement deep learning model is the target prediction network, and the weight of the main action network is synchronized with the weight of the target prediction network;

[0008] The antenna design requirements are input into the target reinforcement deep learning model to obtain the target antenna size that meets the antenna design requirements. The antenna design requirements include the antenna gain requirement range and the antenna reflection coefficient requirement range.

[0009] In a second aspect, the present application provides an antenna design device, the device comprising:

[0010] A first acquisition module is used to acquire a trained target prediction network, where the target prediction network is used to imitate a target antenna design idea to output a first antenna size that meets the target antenna design idea;

[0011] A second acquisition module is used to acquire multiple historical antenna sizes, and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size;

[0012] A first determination module is used to obtain a trained target reinforcement deep learning model according to the multiple historical antenna sizes and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size, wherein a main action network of the target reinforcement deep learning model is the target prediction network, and a weight of the main action network is synchronized with a weight of the target prediction network;

[0013] The second determination module is used to input the antenna design requirements into the target reinforcement deep learning model to obtain a target antenna size that meets the antenna design requirements. The antenna design requirements include an antenna gain requirement range and an antenna reflection coefficient requirement range.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, the electronic device comprising: a processor and a memory storing computer program instructions;

[0015] When the processor executes the computer program instructions, the antenna design method in any one of the embodiments of the first aspect is implemented.

[0016] In a fourth aspect, an embodiment of the present application provides a computer storage medium having computer program instructions stored thereon, and when the computer program instructions are executed by a processor, an antenna design method as in any one of the embodiments in the first aspect is implemented.

[0017] In a fifth aspect, an embodiment of the present application provides a computer program product. When the instructions in the computer program product are executed by a processor of an electronic device, the electronic device executes the antenna design method in any one of the embodiments in the first aspect above.

[0018] In an antenna design method, device, equipment, medium and product provided in an embodiment of the present application, firstly, through a trained target prediction network, it is possible to imitate the target antenna design idea and quickly output a preliminary antenna size that meets the design idea. Secondly, by combining multiple historical antenna sizes and their corresponding performance data, an efficient target reinforcement deep learning model is trained. This model can not only autonomously learn and optimize antenna design parameters, but also the weights of its main action network are synchronized with the target prediction network, ensuring the consistency and accuracy of the design idea and the target prediction network. Therefore, in practical applications, users only need to provide specific antenna design requirements, and the target reinforcement deep learning model can quickly output the target antenna size that meets the design requirements. This automated design process not only improves the efficiency of antenna design, but also reduces the dependence on the professional qualities of designers, making antenna design more efficient, accurate and flexible, thereby improving the efficiency of antenna design. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solution of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0020] Figure 1 is a flowchart of an antenna design method provided by an embodiment of the present application;

[0021] Figure 2 It is a schematic diagram of the training process of the target prediction network provided by an embodiment of the present application;

[0022] Figure 3 It is a schematic diagram of the training process of a target reinforcement deep learning model provided by an embodiment of the present application;

[0023] Figure 4 It is another training process diagram of the target reinforcement deep learning model provided by one embodiment of the present application;

[0024] Figure 5 is another flowchart of an antenna design method provided by an embodiment of the present application;

[0025] Figure 6 is a schematic diagram of a target antenna size obtained using a target reinforcement deep learning model provided by an embodiment of the present application;

[0026] Figure 7 is another schematic diagram of a target antenna size obtained by using a target reinforcement deep learning model provided by an embodiment of the present application;

[0027] Figure 8 It is a reflection coefficient curve diagram of a design without deep learning and a design based on deep learning provided by an embodiment of the present application;

[0028] Fig. 9 It is a graph of antenna gain without deep learning design and based on deep learning design provided by an embodiment of the present application;

[0029] Fig.10 is a structural schematic diagram of an antenna design device provided in an embodiment of the present application;

[0030] Fig.11 It is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to more clearly understand the above-mentioned objectives, features and advantages of the present disclosure, the scheme of the present disclosure will be further described below. It should be noted that the embodiments of the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0032] In the following description, many specific details are set forth to facilitate a full understanding of the present disclosure, but the present disclosure may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present disclosure, rather than all of the embodiments.

[0033] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprising a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0034] In the rapid development of wireless communication technology, antenna design is a core link, and its design efficiency directly affects the overall performance and market competitiveness of the communication system. However, the existing antenna design methods face the challenge of low efficiency. The traditional design process is long and relies on the experience and intuition of the designer. It seems to be unable to design antennas with complex structures and high performance requirements. Although optimization methods based on machine learning and deep reinforcement learning have been proposed, such as antenna topology pixel design assisted by machine learning optimization and array antenna design system based on deep reinforcement learning, these methods still have significant shortcomings. The former makes it difficult to balance design accuracy and efficiency due to the pixelization of the antenna, while the latter is inefficient due to the simulation of complex electromagnetic environments and the training of the initial random strategy network. Therefore, seeking an efficient, accurate and adaptable antenna design method has become a technical problem that needs to be solved urgently in the field of wireless communications.

[0035] In order to solve the problems existing in the prior art, the embodiments of the present application provide an antenna design method, device, equipment, medium and product.

[0036] The present application embodiment provides an antenna design method, device, equipment, medium and product. The antenna design method provided by the present application embodiment is first introduced below. Figure 1 As shown, the method specifically comprises the following steps:

[0037] S100, obtaining a trained target prediction network, where the target prediction network is used to imitate a target antenna design concept to output a first antenna size that meets the target antenna design concept.

[0038] Optionally, in a feasible implementation of the present application, first, a large amount of historical data and expert knowledge about antenna design can be collected, including but not limited to key parameters such as size, frequency response, radiation pattern, etc. of different antenna structures. These data will be used as training samples to build a neural network model that can understand and imitate antenna design ideas.

[0039] Next, a target prediction network is constructed using a deep learning algorithm (such as a convolutional neural network, a recurrent neural network, or their variants). This target prediction network will be trained to predict and output the corresponding antenna size based on the input parameters (such as frequency, bandwidth, gain requirements, etc.). During the training process, the target prediction network will continuously learn and optimize its internal parameters (weights and biases) to minimize the difference between the predicted output and the actual antenna performance. Supervised learning methods can also be used. In supervised learning, each input in the training sample corresponds to a known output (ie, "label"), which can be obtained by experts based on experience or experimental measurements. During the training process, the target prediction network will try to predict these labels and adjust the internal parameters of the target prediction network by calculating the error between the predicted output and the actual label. This process will be repeated many times until the target prediction network can accurately predict the output of most samples.

[0040] In order to further improve the generalization ability and robustness of the network, data augmentation techniques can also be used to expand the number and diversity of training samples. For example, new training samples can be generated by rotating, scaling, translating, and other transformations of the original data. In addition, regularization techniques can be used to prevent network overfitting, that is, when the network relies too much on certain features in the training samples and ignores other features, some randomness is introduced to reduce the complexity of the network.

[0041] Finally, after sufficient training and verification, a trained target prediction network is obtained. This target prediction network can imitate the design idea of ​​the target antenna and quickly output the antenna size that meets the design requirements based on the input parameters.

[0042] S200: Acquire multiple historical antenna sizes, and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size.

[0043] Optionally, in a feasible implementation of the present application, first, data can be collected from the historical records of antenna design. These data can be obtained from publicly published academic papers, research reports and experimental data. These data contain detailed information of various antenna designs, such as antenna size, frequency response, reflection coefficient, gain, etc.

[0044] Specifically, collect the dimensional data of different antenna designs, including length, width, height, caliber, etc. This data can be directly extracted from the design documents. Then obtain the reflection coefficient and gain data corresponding to each antenna size. This data can be obtained through experimental measurement or simulation software. The reflection coefficient describes the antenna's ability to reflect incident waves, while the antenna gain measures the antenna's radiation ability in the direction of maximum radiation.

[0045] However, in practical applications, there may be a problem of insufficient actual antenna data, which limits the learning and generalization capabilities of the deep learning model during training. In order to overcome the challenge of insufficient data, in an embodiment of the present application, electromagnetic simulation software (such as CST) can be used to augment the data set. Electromagnetic simulation software can simulate the electromagnetic behavior of the antenna and predict its performance. Electromagnetic simulation software can generate a series of antenna models with different sizes and performance parameters, and simulate their performance under various conditions.

[0046] Specifically, we can first set a series of target antenna sizes and performance parameter ranges, such as a series of attenuation rates from low to high and a series of bandwidth data from narrow to wide. Then, we use electromagnetic simulation software to create corresponding antenna models and simulate the electromagnetic performance of these models under different frequencies, polarizations, directions, etc. Through simulation, we can obtain key data such as reflection coefficient, gain, and radiation pattern of each antenna model.

[0047] Next, the simulated data is integrated with the existing actual antenna design data to form a richer and more comprehensive data set. This data set not only contains the performance data of the actual antenna, but also expands the coverage and diversity of the data through simulation software. Such a data set can better reflect the actual needs and challenges of antenna design and provide a more comprehensive and reliable foundation for the training of deep learning models.

[0048] By collating and summarizing previous data and using electromagnetic simulation software to expand the data set, we can obtain a rich data set containing a range of attenuation rates from low to high and a range of bandwidth data from narrow to wide. This data set will provide strong support for the subsequent training target reinforcement deep learning model, enabling the model to better learn and understand the laws of antenna design and output antenna size and performance parameters that better meet the design requirements.

[0049] S300, according to the multiple historical antenna sizes, and the first antenna reflection coefficient and the first antenna gain corresponding to each historical antenna size, a trained target reinforcement deep learning model is obtained, the main action network of the target reinforcement deep learning model is the target prediction network, and the weight of the main action network is synchronized with the weight of the target prediction network.

[0050] Optionally, in a feasible implementation of the present application, a deep deterministic policy gradient (DDPG) algorithm can be used to train a main action network suitable for antenna design by strengthening deep learning technology. First, a main action network is constructed based on the target prediction network pre-trained in step 100. This network can generate effective parameter adjustment actions, thereby improving learning efficiency. Next, the state function, action function, loss function of reflection coefficient and gain, and total loss function are defined. These functions together constitute the target indicators for antenna design optimization. By simulating the real environment through electromagnetic simulation software, the antenna size data can be used as input to obtain the corresponding antenna reflection coefficient and gain, and these results are used to calculate the loss function and reward function.

[0051] During the training process, a learning mechanism based on cache space can be adopted. Specifically, the data of each pair of state, action, reward and next state can be stored in the cache space first, and when the cache space is full, a certain amount of data can be randomly sampled from it as learning samples. Through the target action network and the target evaluation network, the target value function can be obtained, and it can be compared with the main value function obtained by the main evaluation network, and the mean square error can be calculated as the error function. Subsequently, the parameters of the main evaluation network and the main action network are updated using the gradient descent method to minimize the error function.

[0052] In order to maintain the stability and convergence of the model, the sliding average assignment update method can be used to update the parameters of the target action network and the target evaluation network. This method can smoothly transition the parameter update of the model and avoid learning instability caused by excessive parameter changes.

[0053] After repeated iterative training, a weighted main action network is finally generated (that is, the trained target reinforcement deep learning model includes the weighted main action network). This weighted main action network can automatically output antenna size parameters according to the given antenna design requirements. Compared with traditional antenna design methods, this antenna design method based on reinforcement deep learning has a higher degree of automation and better optimization effect. It can not only quickly find antenna size parameters that meet the design requirements, but also show better performance in practical applications. Therefore, this method has important application value and development prospects in the field of antenna design.

[0054] S400, inputting the antenna design requirements into the target reinforcement deep learning model to obtain a target antenna size that meets the antenna design requirements, wherein the antenna design requirements include an antenna gain requirement range and an antenna reflection coefficient requirement range.

[0055] Optionally, in the embodiment of the present application, the antenna design requirements may include the following aspects in addition to the antenna gain requirement range and the antenna reflection coefficient requirement range:

[0056] Communication frequency band: Determine the frequency band range in which the antenna works. The antenna design methods for different frequency bands will be different. Functional requirements: Determine the functions that the antenna needs to achieve, such as directional antenna, omnidirectional antenna or sector coverage antenna. Depending on the application requirements, the antenna's directivity, polarization mode, etc. may need to be considered. Performance requirements: Determine the performance requirements of the antenna in terms of bandwidth, gain, return loss, etc. Installation environment: Determine the installation environment of the antenna, such as indoors, outdoors, antenna height, etc. This will affect the size, shape and material selection of the antenna, as well as environmental adaptability requirements such as wind resistance and rain resistance. Reflection coefficient: The reflection coefficient is the ratio of the reflected voltage at the antenna input port to the incident voltage value, reflecting the antenna's ability to transmit electromagnetic energy. The lower the reflection coefficient, the higher the radiation efficiency of the antenna and the greater the effective distance.

[0057] Optionally, in a feasible implementation of the present application, first, the antenna design requirements need to be clarified, including the required range of antenna gain and the required range of antenna reflection coefficient. These requirements reflect the user's expectations for antenna performance and are also the benchmark for guiding the model output results.

[0058] Next, these design requirements are converted into an input form that the model can understand. Specifically, the antenna gain requirement range and the antenna reflection coefficient requirement range can be converted into the corresponding components of the state function. The state function plays the role of describing the current state of the environment in reinforcement learning, and it contains all the important information for decision-making.

[0059] Once the state function is correctly set, it can be input into the target reinforcement deep learning model. At this point, the target reinforcement deep learning model will use its internal neural network structure, combined with the knowledge learned during the training process, to analyze and judge the input state. The main action network in the target reinforcement deep learning model will generate an action based on the current state, which is a suggestion to adjust the antenna size. The target reinforcement deep learning model will then evaluate the potential value of this suggested action. This is done through the main evaluation network, which predicts the future state that may result from the suggested action and calculates the corresponding value. If the value of the suggested action is high, it means that it has the potential to make the antenna performance closer to the design requirements, then the target reinforcement deep learning model will accept the suggestion and output the corresponding target antenna size.

[0060] Finally, we compared and verified the target antenna size output by the target reinforcement deep learning model with the original design requirements. This can be achieved by inputting the target antenna size into the electromagnetic simulation software for simulation. Through simulation, the actual performance parameters of the target antenna size can be obtained and compared with the design requirements. If the actual performance meets the design requirements, it means that the target reinforcement deep learning model has successfully completed the antenna design task; if not, the target reinforcement deep learning model can be fine-tuned or retrained to improve its performance.

[0061] In summary, step S400 realizes the automatic design of antenna size by inputting antenna design requirements into the target-enhanced deep learning model and using the model's intelligent decision-making ability. This step improves the efficiency and accuracy of antenna design.

[0062] In an antenna design method provided in an embodiment of the present application, firstly, through a trained target prediction network, it is possible to imitate the target antenna design idea and quickly output a preliminary antenna size that meets the design idea. Secondly, by combining multiple historical antenna sizes and their corresponding performance data, an efficient target reinforcement deep learning model is trained. This model can not only autonomously learn and optimize antenna design parameters, but also the weights of its main action network are synchronized with the target prediction network, ensuring the consistency and accuracy of the design idea and the target prediction network. Therefore, in practical applications, users only need to provide specific antenna design requirements, and the target reinforcement deep learning model can quickly output the target antenna size that meets the design requirements. This automated design process not only improves the efficiency of antenna design, but also reduces the dependence on the professional qualities of designers, making antenna design more efficient, accurate and flexible, thereby improving the efficiency of antenna design.

[0063] In one embodiment, the above step 100 may specifically perform the following steps:

[0064] S110, obtaining a first training sample set, where the first training sample set includes a plurality of first training samples, each of the first training samples includes an antenna design size and performance data corresponding to the antenna design size, where the performance data includes a second antenna gain and a second antenna reflection coefficient;

[0065] S120, determining the antenna design size as action data of the target prediction network, and determining the performance data as state data of the target prediction network;

[0066] S130, training an untrained target prediction network using the action data and the state data until a first preset training condition is met, thereby obtaining a trained target prediction network.

[0067] In one embodiment, the target prediction network includes a three-layer neural network, and the target prediction network adopts a rectified linear unit activation function.

[0068] Optionally, in a feasible implementation of the present application, as Figure 2 As shown, through imitation learning, the prediction network (i.e., target prediction network) is trained to imitate the design ideas of antenna experts. First, the data of previous antenna designs (i.e., antenna design dimensions) and debugging data (i.e., performance data corresponding to antenna design dimensions) are collated and summarized. If the actual antenna data is not enough, the antenna dimension data and antenna performance data need to be amplified through electromagnetic simulation software, so that the data set (i.e., the first training sample set) contains a series of attenuation rates from low to high and a series of bandwidth data from narrow to wide.

[0069] The training network uses a three-layer neural network, each hidden layer has 128 neurons, and uses a rectified linear unit activation function. Antenna size data, antenna reflection coefficient, and antenna gain are used as input, antenna size data is used as action data, antenna reflection coefficient and antenna gain are used as state data, and antenna size data is output. Finally, a training network with weights is obtained.

[0070] Optionally, in a specific implementation of the present application, first, a data set of a first training sample set is obtained. This set contains multiple first training samples, each of which contains two key parts: the antenna design size and the corresponding performance data. The antenna design size refers to the physical size parameters of the antenna, such as length, width, height, etc. These parameters directly determine the shape and structure of the antenna. The performance data describes the performance of the antenna under different design dimensions, specifically the second antenna gain and the second antenna reflection coefficient. The antenna gain reflects the ability of the antenna to radiate or receive signals in a specific direction, while the reflection coefficient describes the part of the antenna that is reflected back to the source when transmitting the signal.

[0071] Next, the antenna design size and performance data are used as the action data and state data of the target prediction network respectively. The "action data" here can be understood as the input data that affects the network state (i.e., output), while the "state data" is the result predicted or output by the network based on the input data. In this scenario, the action data is the antenna design size because it directly affects the performance of the antenna; and the state data is the performance data, i.e., the target value that the network needs to learn.

[0072] Finally, these action data and state data are used to train an untrained target prediction network. The training process usually involves multiple iterations. In each iteration, the network adjusts its internal parameters (such as weights and biases) according to the current input (action data) and the corresponding target output (state data) to reduce the difference between the predicted value and the actual value. This process continues until the first preset training condition is met. The first preset training condition may include the prediction error being lower than a certain threshold, the number of training iterations reaching a specified number, etc. When these conditions are met, the target prediction network is considered to have been fully trained and can be used in actual antenna performance prediction tasks.

[0073] In these optional embodiments, by obtaining a first training sample set containing antenna design dimensions and corresponding performance data, the target prediction network can learn a direct mapping relationship from design dimensions to performance, which lays the foundation for improving the accuracy and efficiency of antenna design. Secondly, using the antenna design dimensions as action data and the performance data as state data not only clarifies the corresponding relationship between the input and output of the target prediction network, but also helps the target prediction network to better understand the intrinsic relationship between design parameters and performance. Finally, through a three-layer neural network structure and the use of a rectified linear unit activation function, the target prediction network can capture complex nonlinear relationships in the data while maintaining efficient calculations. The rectified linear unit activation function helps solve the gradient vanishing problem, speeds up the training of the target prediction network, and improves the generalization ability of the target prediction network.

[0074] In one embodiment, the above step 300 may specifically perform the following steps:

[0075] S310, creating a state function and an action function of the target reinforcement deep learning model according to the multiple historical antenna sizes, and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size;

[0076] S320, creating a first loss function of the target reinforcement deep learning model according to the first antenna reflection coefficient corresponding to each historical antenna size;

[0077] S330, creating a second loss function of the target reinforcement deep learning model according to the first antenna gain corresponding to each historical antenna size;

[0078] S340, determining a total loss function of the target reinforcement deep learning model according to the first loss function and the second loss function;

[0079] S350, determining a reward function of the target reinforcement deep learning model according to the total loss function;

[0080] S360, using the state function, the action function, the total loss function and the reward function, to train the target reinforcement deep learning model until a second preset training condition is met, thereby obtaining a trained target reinforcement deep learning model.

[0081] Optionally, in a feasible implementation of the present application, first, based on multiple historical antenna sizes and the first antenna reflection coefficient and first antenna gain corresponding to each size, the state function and action function of the target reinforcement deep learning model are defined. The state function describes the current state in the antenna design process, that is, the current antenna size and known performance data; while the action function defines the possible actions based on the current state, that is, the possible antenna size adjustment.

[0082] Next, the first loss function of the target reinforcement deep learning model was created based on the first antenna reflection coefficient corresponding to each historical antenna size. The reflection coefficient is one of the important indicators to measure antenna performance. Therefore, by minimizing the loss function of the reflection coefficient, the target reinforcement deep learning model can learn how to adjust the antenna size to reduce the reflection coefficient.

[0083] Then, similarly, a second loss function of the target reinforcement deep learning model is created based on the first antenna gain corresponding to each historical antenna size. Antenna gain is another key performance indicator. By optimizing the loss function of gain, the target reinforcement deep learning model can learn how to adjust the antenna size to improve gain.

[0084] The first loss function and the second loss function are then combined to determine the total loss function of the target reinforcement deep learning model. This total loss function comprehensively considers the two performance indicators of reflection coefficient and gain, allowing the model to optimize these two aspects simultaneously during training.

[0085] Next, the reward function of the target reinforcement deep learning model is determined based on the total loss function. The reward function is a key concept in reinforcement learning that guides model learning. It measures the overall performance of the target reinforcement deep learning model after taking a specific action in a specific state. By defining a suitable reward function, it can be ensured that the target reinforcement deep learning model moves in the direction of optimizing the total loss function during the training process.

[0086] Finally, the target reinforcement deep learning model is trained using the state function, action function, total loss function, and reward function. During the training process, the target reinforcement deep learning model will continuously try different antenna size adjustments (i.e., actions) and observe the effects of these adjustments on the reflection coefficient and gain (i.e., state). By calculating the reward function and the total loss function, the target reinforcement deep learning model can gradually learn which actions can lead to better performance. When the target reinforcement deep learning model meets the second preset training condition during the training process (such as the performance reaches the preset threshold or the number of training rounds reaches the preset upper limit), the target reinforcement deep learning model is considered to have been trained and can be used for future antenna design optimization.

[0087] In summary, this embodiment builds a target-enhanced deep learning model that can automatically learn and optimize antenna design by defining state functions, action functions, loss functions, and reward functions. By training this model, the performance of different antenna sizes can be predicted based on historical data, and future antenna designs can be guided accordingly, thereby achieving performance optimization and cost reduction.

[0088] In these optional embodiments, by constructing state functions and action functions based on multiple historical antenna sizes and their corresponding performance data, the target reinforcement deep learning model can learn the complex relationship between antenna design size and performance. At the same time, by creating a loss function based on reflection coefficient and gain, and then combining them to determine the total loss function, the target reinforcement deep learning model can comprehensively consider different performance indicators during the training process to ensure that the gain is improved while reducing the reflection coefficient. This multi-index optimization method makes the target reinforcement deep learning model more practical and accurate in the field of antenna design. In addition, using the reward function in reinforcement learning, the target reinforcement deep learning model can self-adjust during the training process to find the optimal antenna design size, further improving the performance and generalization ability of the target reinforcement deep learning model.

[0089] In one embodiment, the target reinforcement deep learning model includes a main network and a target network, the main network includes the main action network and the main evaluation network, and the target network includes a target action network and a target evaluation network; the above step 360 can specifically perform the following steps:

[0090] S361, creating a second training sample set, wherein the second training sample set includes a plurality of second training samples, each of which includes the state function s corresponding to a historical antenna size. t , the reward function r t , and s t The corresponding next state function s t+1 ;

[0091] For each of the second training samples, the following steps are performed respectively:

[0092] Will s t+1 Input to the target action network to get s t+1 The corresponding target action function a t+1 ;

[0093] Will s t+1 and a t+1 Input into the target evaluation network to obtain the target value function;

[0094] Will s t Input into the main action network to obtain the action function a t ;

[0095] A t and t Input into the main evaluation network to obtain the main value function;

[0096] Constructing an error function according to the mean square error of the main value function and the target value function;

[0097] Calculating a first gradient of the error function with respect to a trainable weight of the primary evaluation network;

[0098] According to a gradient update algorithm, updating the parameters of the main evaluation network based on the first gradient to obtain an updated main evaluation network;

[0099] Calculating a second gradient of a negative value of a mean of the primary value function with respect to a trainable weight of the primary action network;

[0100] According to the gradient update algorithm, updating the parameters of the main action network based on the second gradient to obtain an updated main action network;

[0101] Using a sliding average assignment update method, based on the updated main evaluation network and the updated main action network, the target action and the target evaluation network are updated to obtain an updated target reinforcement deep learning model;

[0102] Determining whether the second preset training condition is met;

[0103] If the second preset training condition is not met, return to t+1 Input to the target action network to get s t+1 The corresponding target action function a t+1 Steps;

[0104] When the second preset training condition is met, the trained target reinforcement deep learning model is output.

[0105] In these optional embodiments, by constructing a second training sample set containing historical antenna sizes and their corresponding states, actions, and rewards, the target reinforcement deep learning model can learn the mapping relationship from the current state to the future state and the corresponding actions and rewards, thereby more accurately predicting and selecting the optimal antenna design action.

[0106] Secondly, by inputting the state into the main network and the target network respectively, the main value function and the target value function are obtained, and the error function is constructed based on the mean square error of the two. The target reinforcement deep learning model can continuously reduce the prediction error during the training process and improve the accuracy of the prediction.

[0107] Thirdly, by using the gradient descent algorithm to update the parameters of the main evaluation network and the main action network, the target reinforcement deep learning model can gradually optimize its internal representation and decision-making ability, making the target reinforcement deep learning model more accurate and efficient when selecting antenna design actions. Finally, by using the sliding average assignment update method to update the target network, the model can maintain a certain stability and avoid performance fluctuations caused by frequent updates. In summary, this training strategy can significantly improve the prediction accuracy and decision-making ability of the target reinforcement deep learning model in the field of antenna design, and provide strong technical support for designing antennas with better performance.

[0108] In one embodiment, the first loss function LS 11t for:

[0109] The second loss function LG t for:

[0110] The total loss function L t For: L t =ηLS 11t +(1-η)LG t ;

[0111] Among them, S min and S max is the upper and lower limits of the first antenna reflection coefficient, G min and G maxis the upper and lower limits of the first antenna gain, and N is the reflection coefficient S of the first antenna. 11t , the second antenna gain G t The number of samples, η is used to adjust the priority of the first antenna reflection coefficient and the first antenna gain in the target reinforcement deep learning model;

[0112] The reward function r t for:

[0113] Among them, L thre is the loss threshold, s t is the state function, a t is the action function.

[0114] Optionally, in a specific implementation of the present application, an action network (i.e., a main action network in a target reinforcement deep learning model) is trained for use in antenna design through reinforcement deep learning.

[0115] The Deep Deterministic Policy Gradient (DDPG) algorithm is selected as the reinforcement deep learning algorithm, which is particularly suitable for dealing with high-dimensional and continuous space problems. The real environment feedback is simulated by electromagnetic simulation software, and the antenna size data P is simulated to obtain the antenna reflection coefficient S 11 , antenna gain G.

[0116] The reinforcement deep learning (i.e., target reinforcement deep learning model) of this application has four networks: the main action network and the main evaluation network in the main network, and the target action network and the target evaluation network in the target network. The main action network uses the weighted training network (i.e., the target prediction network) obtained by pre-training, which can produce better parameter adjustment actions and improve learning efficiency.

[0117] Define the state function s t =(S 11t ,G t ,P t )

[0118] Where S 11t is the antenna reflection coefficient, G t is the antenna gain, S 11t and G t are all one-dimensional vectors after curve sampling, with the number of sampling points being N, P t Antenna size data.

[0119] Define action function a t =P t+1 =(w t+1 ,l t+1 ……,h t+1 )

[0120] Where P t+1 is the antenna size data, w t+1 ,l t+1 ……,h t+1 The specific dimensions of the antenna include length, width, height, etc.

[0121] Define the antenna reflection coefficient S 11t The loss function

[0122] Define the antenna gain G t The loss function

[0123] Define the total loss function L t =ηLS 11t +(1-η)LG t

[0124] Among them, S min and S max is the upper and lower limits of the target antenna reflection coefficient, G min and G max are the upper and lower limits of the target antenna gain, and N is the antenna reflection coefficient S 11t 、Antenna gain G t The number of sampling points is η, which is used to adjust the priority of antenna reflection coefficient and antenna gain in learning.

[0125] Define the reward function

[0126] Among them, L thre is the loss threshold, which is used to adjust the learning situation.

[0127] The state function s t Input into the main action network to get the action function a t =μ(s t )

[0128] like Figure 3 As shown, with action function a t As the mean value, construct a normal distribution function H with a standard deviation of V t , the action function is re-integrated in H t Randomly take values ​​in and get the action function a with normal distribution noise added t =H t-random

[0129] The action function a with normal distribution noise added t Put it into the simulation environment and get feedback. t and the next state st+1

[0130] Will (s t ,a t ,r t ,s t+1 ) (i.e., a second training sample) is stored in the cache space D, and when the cache space is full, learning begins.

[0131] Specifically, if Figure 4 As shown, BS group data is randomly sampled from the cache space as learning samples (i.e., the second training sample set), and the state function s t+1 Input the target action network to get the action function a t+1 , a t+1 and t+1 Input into the target evaluation network to obtain the target value function q t+1 , according to the reward function, we get the target value function y = r t +γq t+1 , a t and t Input into the main evaluation network to obtain the main value function q t .

[0132] According to the mean square error of the main value function and the target value function, the error function td_error=MSE(y,q t ), calculate the gradient of td_error with respect to the trainable weights of the main evaluation network (i.e., the first gradient), and update the parameters of the main evaluation network with the applied gradient and the first gradient.

[0133] The state function s t Input the main action network to get the action function a t , will s t and a t Input into the main evaluation network to obtain the main value function q t , ask for q t The mean and negative value a loss =-mean(q t ), calculate the gradient of a_loss relative to the trainable weights of the main action network (i.e., the second gradient), and update the parameters of the main action network with the applied gradient and the second gradient. Then use the sliding average assignment update method to update the target action and target evaluation network based on the main evaluation network and the main action network with updated parameters. After repeated iterations, if the second preset training condition is met, a main action network with weights is generated. Among them, the gradient update algorithm can be a gradient descent algorithm.

[0134] Specifically, first, based on the parameters of the current updated main evaluation network and main action network, the parameters of the target network are calculated using the sliding average method. Specifically, it involves maintaining a shadow variable for each target network, which can be updated by weighted averaging the current parameters and historical parameters of the target network. Specifically, in each iteration, the shadow variable of the target network is updated according to the decay rate, that is, the new shadow variable value is equal to the old shadow variable value multiplied by the decay rate plus the current parameter value multiplied by (1 minus the decay rate). As the iteration proceeds, the parameter update of the target network will be smoother, reducing the sensitivity to the latest parameter changes. Then, this iterative process is repeated until the preset training conditions are met (such as reaching a certain number of iterations, stable model performance, etc.). Finally, a weighted main action network is generated, the weights of which are the parameters of the main action network in the last iteration. These parameters have been fused with the historical parameters through the sliding average method, thereby improving the stability and generalization ability of the network.

[0135] Optionally, in a specific implementation of the present application, as Figure 5 As shown, through imitation learning, the existing antenna design data is trained into a weighted pre-trained model (i.e., the target prediction network) for use in reinforcement deep learning (i.e., the target reinforcement deep learning model). A loss function and a reward function suitable for almost all antenna designs are proposed. Reinforcement deep learning generates a reinforcement deep learning model. When designing the antenna, one only needs to input the antenna design requirements into the reinforcement deep learning model to obtain the specific size data of the antenna.

[0136] Optionally, in a specific embodiment of the present application, as Figure 6 and Figure 7 As shown in the figure, taking the slot filter antenna with a center frequency of 2.4G as an example, the design requirements of the 2.4G slot filter antenna are input into the trained target reinforcement deep learning model, and the antenna size data is obtained as P = (l,l1,l2,l3,d1,d2,d3,w1,w2,w3,h,h1); among them, Figure 6 As shown, l is the length of the entire square antenna, l1, l2, l3 are the lengths of different slot filter antennas in the forward direction antenna, d1, d2, d3 are the distances from the slot centers of different slot filter antennas in the forward direction antenna to the edge of the antenna board, w1, w2, w3 are the widths of different slot filter antennas in the forward direction antenna, h is the distance from the antenna to the ground, and h1 is the distance from the antenna to the antenna feeder.

[0137] like Figure 8 and Fig. 9As shown, it can be seen that when the design requirements are input into the trained target reinforcement deep learning model (curve designed based on deep learning), the reflection coefficient and antenna gain performance corresponding to the obtained antenna size data show better performance than when the same design requirements are input into the target prediction network that has not been optimized by deep learning (i.e., the curve without deep learning design).

[0138] In these optional embodiments, imitation learning (i.e., target prediction network) is used to strengthen deep reinforcement learning (i.e., target reinforcement deep learning model), and existing antenna design experience is used to generate a weighted training network to improve the learning efficiency of subsequent deep reinforcement learning and avoid the situation where learning cannot converge. A reward function and a loss function based on antenna design are defined, which are universal for all antenna designs and have a wide range of applications.

[0139] Optionally, in an embodiment of the present application, by creating a second training sample set, and based on the state function, reward function, and next state function of the historical antenna size, the target reinforcement deep learning model can learn the optimal mapping relationship between state and action in the antenna design process. In addition, by defining the first loss function and the second loss function, and constructing the total loss function accordingly, the model can comprehensively consider the two key performance indicators of antenna reflection coefficient and antenna gain during the training process to ensure that the designed antenna achieves the best performance.

[0140] During the training process, the gradient descent algorithm is used to update the parameters of the main evaluation network and the main action network, so that the model can gradually approach the optimal solution. At the same time, the sliding average assignment update method is used to update the target network to maintain the stability of the model. By setting the reward function, the model can adjust the learning direction according to the value of the loss function during the training process, further improving the training efficiency and model performance.

[0141] In summary, these steps not only optimize the antenna design process and improve design efficiency, but also achieve accurate prediction and optimization of antenna performance through deep learning methods, providing an efficient and accurate design tool for the field of antenna design.

[0142] It should be noted that in a real-time method of the present application, when using the target-enhanced deep learning model for antenna design, multiple types of antenna data can be used when training the target prediction network to ensure that the target-enhanced deep learning model can handle multiple types of antenna designs (i.e., multiple types corresponding to the training of the target prediction network). In this way, the generalization ability of the target-enhanced deep learning model can be enhanced through the diversity of training data, so that it can be applied to a wider range of antenna types.

[0143] Specifically, when training the target prediction network, if a data set covering multiple antenna types is used, the trained target-enhanced deep learning model will be able to learn the design rules and characteristics of these different types of antennas. Therefore, when the target-enhanced deep learning model is used for antenna design in the future, the target-enhanced deep learning model will be able to select the most suitable antenna type for design according to the design requirements and generate the corresponding antenna size data.

[0144] The advantage of this method is that it makes full use of the diversity of training data, so that the target-enhanced deep learning model can adapt to a wider range of antenna design requirements. At the same time, since the target-enhanced deep learning model has learned the design rules of multiple antenna types during the training process, it can more accurately predict the antenna size data that meets the design requirements in actual application, thereby improving the accuracy and efficiency of antenna design.

[0145] Fig.10 A schematic structural diagram of an antenna design device provided in another embodiment of the present application is shown. For ease of explanation, only the parts related to the embodiment of the present application are shown.

[0146] Reference Fig.10 , the antenna design device may include:

[0147] A first acquisition module 1001 is used to acquire a trained target prediction network, where the target prediction network is used to imitate a target antenna design idea to output a first antenna size that meets the target antenna design idea;

[0148] The second acquisition module 1002 is used to acquire multiple historical antenna sizes, and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size;

[0149] A first determination module 1003 is used to obtain a trained target reinforcement deep learning model according to the multiple historical antenna sizes and the first antenna reflection coefficient and the first antenna gain corresponding to each historical antenna size, wherein the main action network of the target reinforcement deep learning model is the target prediction network, and the weight of the main action network is synchronized with the weight of the target prediction network;

[0150] The second determination module 1004 is used to input the antenna design requirements into the target reinforcement deep learning model to obtain a target antenna size that meets the antenna design requirements, where the antenna design requirements include an antenna gain requirement range and an antenna reflection coefficient requirement range.

[0151] In one embodiment, the first acquisition module 1001 may include:

[0152] A first acquisition submodule is used to acquire a first training sample set, where the first training sample set includes a plurality of first training samples, each of the first training samples includes an antenna design size and performance data corresponding to the antenna design size, where the performance data includes a second antenna gain and a second antenna reflection coefficient;

[0153] A first determination submodule, configured to determine the antenna design size as action data of the target prediction network, and determine the performance data as state data of the target prediction network;

[0154] The first training submodule is used to train an untrained target prediction network using the action data and the state data until a first preset training condition is met to obtain a trained target prediction network.

[0155] In one embodiment, the target prediction network includes a three-layer neural network, and the target prediction network adopts a rectified linear unit activation function.

[0156] In one embodiment, the first determining module 1003 may include:

[0157] A first creation submodule, configured to create a state function and an action function of the target reinforcement deep learning model according to the multiple historical antenna sizes and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size;

[0158] A second creation submodule is used to create a first loss function of the target reinforcement deep learning model according to the first antenna reflection coefficient corresponding to each historical antenna size;

[0159] A third creation submodule is used to create a second loss function of the target reinforcement deep learning model according to the first antenna gain corresponding to each historical antenna size;

[0160] A second determination submodule, configured to determine a total loss function of the target reinforcement deep learning model according to the first loss function and the second loss function;

[0161] A third determination submodule is used to determine the reward function of the target reinforcement deep learning model according to the total loss function;

[0162] The second training submodule is used to train the target reinforcement deep learning model using the state function, the action function, the total loss function and the reward function until a second preset training condition is met to obtain a trained target reinforcement deep learning model.

[0163] In one embodiment, the target reinforcement deep learning model includes a main network and a target network, the main network includes the main action network and the main evaluation network, and the target network includes a target action network and a target evaluation network; the second training submodule may include:

[0164] The first creation unit is used to create a second training sample set, wherein the second training sample set includes a plurality of second training samples, each of which includes the state function s corresponding to a historical antenna size. t , the reward function r t , and s t The corresponding next state function s t+1 ;

[0165] The training unit is used to perform the following steps on each of the second training samples:

[0166] Will s t+1 Input to the target action network to get s t+1 The corresponding target action function a t+1 ;

[0167] Will s t+1 and a t+1 Input into the target evaluation network to obtain the target value function;

[0168] Will s t Input into the main action network to obtain the action function a t ;

[0169] A t and t Input into the main evaluation network to obtain the main value function;

[0170] Constructing an error function according to the mean square error of the main value function and the target value function;

[0171] Calculating a first gradient of the error function with respect to a trainable weight of the primary evaluation network;

[0172] According to a gradient update algorithm, updating the parameters of the main evaluation network based on the first gradient to obtain an updated main evaluation network;

[0173] Calculating a second gradient of a negative value of a mean of the primary value function with respect to a trainable weight of the primary action network;

[0174] According to the gradient update algorithm, updating the parameters of the main action network based on the second gradient to obtain an updated main action network;

[0175] Using a sliding average assignment update method, based on the updated main evaluation network and the updated main action network, the target action and the target evaluation network are updated to obtain an updated target reinforcement deep learning model;

[0176] Determining whether the second preset training condition is met;

[0177] If the second preset training condition is not met, return to t+1 Input to the target action network to get s t+1 The corresponding target action function a t+1 Steps;

[0178] When the second preset training condition is met, the trained target reinforcement deep learning model is output.

[0179] In one embodiment, the first loss function LS 11t for:

[0180] The second loss function LG t for:

[0181] The total loss function L t For: L t =ηLS 11t +(1-η)LG t ;

[0182] Among them, S min and S max is the upper and lower limits of the first antenna reflection coefficient, G min and G max is the upper and lower limits of the first antenna gain, and N is the reflection coefficient S of the first antenna. 11t , the second antenna gain G t The number of samples, η is used to adjust the priority of the first antenna reflection coefficient and the first antenna gain in the target reinforcement deep learning model;

[0183] The reward function r t for:

[0184] Among them, L thre is the loss threshold, s t is the state function, a t is the action function.

[0185] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application, and are devices corresponding to the above-mentioned battery thermal runaway warning method. All implementation methods in the above-mentioned method embodiment are applicable to the embodiments of the device. Its specific functions and technical effects can be found in the method embodiment part, which will not be repeated here.

[0186] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0187] Fig.11 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application is shown.

[0188] The device may include a processor 1101 and a memory 1102 storing program instructions.

[0189] When the processor 1101 executes the program, the steps in any of the above method embodiments are implemented.

[0190] Exemplarily, the program may be divided into one or more modules / units, one or more modules / units are stored in the memory 1102, and executed by the processor 1101 to complete the present application. One or more modules / units may be a series of program instruction segments capable of completing a specific function, and the instruction segments are used to describe the execution process of the program in the device.

[0191] Specifically, the processor 1101 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0192] The memory 1102 may include a large capacity memory for data or instructions. By way of example and not limitation, the memory 1102 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive or a combination of two or more of these. In appropriate cases, the memory 1102 may include a removable or non-removable (or fixed) medium. In appropriate cases, the memory 1102 may be inside or outside the integrated gateway disaster recovery device. In a specific embodiment, the memory 1102 is a non-volatile solid-state memory.

[0193] The memory may include read-only memory (ROM), random access memory (RAM), magnetic disk storage media devices, optical storage media devices, flash memory devices, electrical, optical or other physical / tangible memory storage devices. Thus, typically, the memory includes one or more tangible (non-transitory) readable storage media (e.g., memory devices) encoded with software including computer executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to an aspect of the present disclosure.

[0194] The processor 1101 implements any one of the methods in the above embodiments by reading and executing program instructions stored in the memory 1102 .

[0195] In one example, the electronic device may further include a communication interface 1103 and a bus 1110. The processor 1101, the memory 1102, and the communication interface 1103 are connected via the bus 1110 and communicate with each other.

[0196] The communication interface 1103 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiments of the present application.

[0197] Bus 1110 includes hardware, software or both, and the parts of online data flow billing equipment are coupled to each other. For example, but not limitation, bus may include accelerated graphics port (AGP) or other graphics bus, enhanced industrial standard architecture (EISA) bus, front-end bus (FSB), hypertransport (HT) interconnection, industrial standard architecture (ISA) bus, infinite bandwidth interconnection, low pin count (LPC) bus, memory bus, micro channel architecture (MCA) bus, peripheral component interconnection (PCI) bus, PCI-Express (PCI-X) bus, serial advanced technology attachment (SATA) bus, video electronics standard association local (VLB) bus or other suitable bus or two or more of these combinations. In appropriate cases, bus 1110 may include one or more buses. Although the present application embodiment describes and shows a specific bus, the present application considers any suitable bus or interconnection.

[0198] In addition, in combination with the method in the above embodiment, the embodiment of the present application can provide a storage medium for implementation. The storage medium stores program instructions; when the program instructions are executed by a processor, any one of the methods in the above embodiment is implemented.

[0199] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, wherein the communication interface and the processor are coupled, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0200] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0201] An embodiment of the present application provides a computer program product, which is stored in a storage medium. The program product is executed by at least one processor to implement the various processes of the above-mentioned method embodiment and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0202] It should be clear that the present application is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application.

[0203] The functional modules shown in the above block diagram can be implemented as hardware, software, firmware or their combination. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), suitable firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present application are programs or code segments that are used to perform the required tasks. Programs or code segments can be stored in machine-readable media, or transmitted on a transmission medium or a communication link by a data signal carried in a carrier wave. "Machine-readable media" can include any medium that can store or transmit information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, optical fiber media, radio frequency (RF) links, etc. Code segments can be downloaded via computer grids such as the Internet, intranets, etc.

[0204] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or devices. However, this application is not limited to the order of the above steps, that is, the steps can be performed in the order mentioned in the embodiment, or in a different order from the embodiment, or several steps can be performed simultaneously.

[0205] Aspects of the present disclosure are described above with reference to the flowchart and / or block diagram of the method, device (system) and program product according to the embodiment of the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer program instructions. These program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer or other programmable data processing device to produce a machine so that these instructions executed by the processor of the computer or other programmable data processing device enable the implementation of the function / action specified in one or more boxes of the flowchart and / or block diagram. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor or a field programmable logic circuit. It can also be understood that each box in the block diagram and / or flowchart and the combination of boxes in the block diagram and / or flowchart can also be implemented by dedicated hardware that performs a specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0206] The above are only specific implementation methods of the present application. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the protection scope of the present application is not limited to this. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in this application, and these modifications or replacements should be included in the protection scope of this application.

Claims

1. An antenna design method, characterized in that: The method comprises: Acquire a trained target prediction network, where the target prediction network is used to imitate a target antenna design idea to output a first antenna size that meets the target antenna design idea; Acquire multiple historical antenna sizes, and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size; According to the multiple historical antenna sizes, and the first antenna reflection coefficient and the first antenna gain corresponding to each historical antenna size, a trained target reinforcement deep learning model is obtained, wherein the main action network of the target reinforcement deep learning model is the target prediction network, and the weight of the main action network is synchronized with the weight of the target prediction network; The antenna design requirements are input into the target reinforcement deep learning model to obtain a target antenna size that meets the antenna design requirements. The antenna design requirements include an antenna gain requirement range and an antenna reflection coefficient requirement range.

2. The method according to claim 1, characterized in that The step of obtaining a trained target prediction network includes: Acquire a first training sample set, where the first training sample set includes a plurality of first training samples, each of the first training samples includes an antenna design size and performance data corresponding to the antenna design size, where the performance data includes a second antenna gain and a second antenna reflection coefficient; Determine the antenna design size as action data of the target prediction network, and determine the performance data as state data of the target prediction network; An untrained target prediction network is trained using the action data and the state data until a first preset training condition is met, thereby obtaining a trained target prediction network.

3. The method according to claim 2, characterized in that The target prediction network includes a three-layer neural network, and the target prediction network adopts a rectified linear unit activation function.

4. The method according to claim 1, characterized in that: The step of obtaining a trained target reinforcement deep learning model according to the multiple historical antenna sizes and the first antenna reflection coefficient and the first antenna gain corresponding to each historical antenna size includes: Creating a state function and an action function of the target reinforcement deep learning model according to the multiple historical antenna sizes and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size; Creating a first loss function of the target reinforcement deep learning model according to the first antenna reflection coefficient corresponding to each historical antenna size; Creating a second loss function of the target reinforcement deep learning model according to the first antenna gain corresponding to each historical antenna size; Determine a total loss function of the target reinforcement deep learning model according to the first loss function and the second loss function; Determining a reward function of the target reinforcement deep learning model according to the total loss function; The target reinforcement deep learning model is trained using the state function, the action function, the total loss function and the reward function until a second preset training condition is met to obtain a trained target reinforcement deep learning model.

5. The method according to claim 4, characterized in that The target reinforcement deep learning model includes a main network and a target network, the main network includes the main action network and the main evaluation network, and the target network includes a target action network and a target evaluation network; The method of training the target reinforcement deep learning model by using the state function, the action function, the total loss function and the reward function until a second preset training condition is met to obtain a trained target reinforcement deep learning model includes: Create a second training sample set, the second training sample set includes multiple second training samples, each of the second training samples includes the state function s corresponding to a historical antenna size t , the reward function r t , and s t The corresponding next state function s t+1 ; For each of the second training samples, the following steps are performed respectively: Will s t+1 Input to the target action network to get s t+1 The corresponding target action function a t+1 ; Will s t+1 and a t+1 Input into the target evaluation network to obtain the target value function; Will s t Input into the main action network to obtain the action function a t ; A t and t Input into the main evaluation network to obtain the main value function; Constructing an error function according to the mean square error of the main value function and the target value function; Calculating a first gradient of the error function with respect to a trainable weight of the primary evaluation network; According to a gradient update algorithm, updating the parameters of the main evaluation network based on the first gradient to obtain an updated main evaluation network; Calculating a second gradient of a negative value of a mean of the primary value function with respect to a trainable weight of the primary action network; According to the gradient update algorithm, updating the parameters of the main action network based on the second gradient to obtain an updated main action network; Using a sliding average assignment update method, based on the updated main evaluation network and the updated main action network, the target action and the target evaluation network are updated to obtain an updated target reinforcement deep learning model; Determining whether the second preset training condition is met; If the second preset training condition is not met, return to t+1 Input to the target action network to get s t+1 The corresponding target action function a t+1 Steps; When the second preset training condition is met, the trained target reinforcement deep learning model is output.

6. The method according to claim 4, characterized in that The first loss function LS 11t for: The second loss function LG t for: The total loss function L t For: L t =ηLS 11t +(1-η)LG t ; Among them, S min and S max is the upper and lower limits of the first antenna reflection coefficient, G min and G max is the upper and lower limits of the first antenna gain, and N is the reflection coefficient S of the first antenna. 11t , the second antenna gain G t The sampling number η is used to adjust the priority of the first antenna reflection coefficient and the first antenna gain in the target reinforcement deep learning model; The reward function r t for: Among them, L thre is the loss threshold, s t is the state function, a t is the action function.

7. An antenna design device, characterized in that: The device comprises: A first acquisition module is used to acquire a trained target prediction network, where the target prediction network is used to imitate a target antenna design idea to output a first antenna size that meets the target antenna design idea; A second acquisition module is used to acquire multiple historical antenna sizes, and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size; A first determination module is used to obtain a trained target reinforcement deep learning model according to the multiple historical antenna sizes and a first antenna reflection coefficient and a first antenna gain corresponding to each historical antenna size, wherein a main action network of the target reinforcement deep learning model is the target prediction network, and a weight of the main action network is synchronized with a weight of the target prediction network; The second determination module is used to input the antenna design requirements into the target reinforcement deep learning model to obtain a target antenna size that meets the antenna design requirements. The antenna design requirements include an antenna gain requirement range and an antenna reflection coefficient requirement range.

8. An electronic device, characterized in that: The device comprises: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, the antenna design method according to any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the antenna design method according to any one of claims 1 to 6 is implemented.

10. A computer program product, characterized in that When the instructions in the computer program product are executed by a processor of an electronic device, the electronic device executes the antenna design method as described in any one of claims 1 to 6.

Citation Information

Cited By

  • Design method of Fresnel zone plate antenna substitution model fused with physical mechanism

    CN120297165A