A method for antenna structure parameter optimization based on image translation

Abstract

The present invention discloses an antenna structural parameter optimization method based on image translation, comprising the following steps: S1. Based on the basic structure of the target antenna, the structural parameters to be optimized are selected and the target electromagnetic response performance parameters are determined; S2. Samples are generated for the selected structural parameter vectors using the LHS algorithm to obtain electromagnetic response data corresponding to the structural parameters; S3. The structural parameters and electromagnetic response data are converted into image data to form a data set, and an image translation network is trained using the data set; S4. The trained image translation network is used to perform image translation on the target electromagnetic response image to obtain a corresponding structural parameter image; S5. Parameters are extracted from the structural parameter image to obtain the structural parameters of the target antenna, thereby completing the antenna design. The present invention solves the problems of the prior art, such as the high requirement for prior knowledge of engineers, high computational and time costs, insufficient accuracy, and poor generalization to different antenna design tasks.

Description

Technical Field

[0001] The present invention relates to the field of mold technology, and in particular to an antenna structure parameter optimization method based on image translation. Background Art

[0002] Antennas are the physical foundation for transmitting and receiving wireless signals and play a central role in improving communication quality and efficiency. Advances in antenna technology have directly driven the development of wireless communication systems. Traditional antenna design relies on engineers' prior knowledge and extensive electromagnetic simulations, resulting in significant computational and time costs. These costs are particularly exacerbated for complex antennas such as multi-band antennas and phased array antennas. Therefore, designing a fast, accurate, and simulation-reliant antenna optimization method is crucial.

[0003] With the rise of deep learning, many deep learning-based antenna optimization methods have emerged in recent years. For example, Dhaliwal et al. proposed a hybrid algorithm based on particle swarm optimization (PSO) and artificial neural networks (ANNs) to optimize antenna size in the 5.8 GHz WLAN band. Zhang et al. proposed a deep Gaussian process (DGP) combined with a PSO algorithm to optimize the S11 curve of a multi-band antenna. Liu et al. proposed an ANN-GA hybrid algorithm to optimize the S21 parameters of a multiple-input multiple-output (MIMO) antenna to reduce coupling. Dong et al. used a radial basis function neural network (RBFNN) as a surrogate model to replace full-wave simulation and combined it with an improved PSO algorithm to optimize the impedance matching and size of the antenna. Farzad Mir et al. used a deep neural network (DNN) as a surrogate model and combined it with Thompson sampling efficient multi-objective optimization (TSEMO) to optimize the S11 curve and gain of a proposed microstrip antenna. The above methods use ANN, DGP, and other surrogate models to replace full-wave simulation. By defining optimization problems (such as the S11 curve and antenna size) and combining them with optimization algorithms such as PSO and GA, they achieve efficient iterative search for the optimal solution. Although this forward design method uses surrogate models to quickly obtain simulation results, it still requires multiple iterations to achieve the optimization goal. At the same time, the prediction results obtained by the surrogate model may have errors, which will be amplified during multiple iterations and are not conducive to finding the optimal solution. In addition, when there are multiple design tasks, the forward design method requires separate search iterations for different tasks, which will increase the time and computational cost exponentially.

[0004] Compared to forward design, inverse design directly obtains the corresponding antenna structural parameters using the target electromagnetic response as input, eliminating the iterative process of the optimization algorithm and being a more efficient approach. Sharma et al. sampled the S11 curve of a dual-band T-type antenna and expressed the sum of the absolute values ​​of the sampling points as the Form of Object (FOM). They used algorithms such as ANN to predict the FOM to optimize impedance matching and bandwidth. et al. compressed the S11 curve spectrum of a dual-band antenna into four point values ​​as data to train the ANN and optimize the four parameters of the proposed dual-band antenna. Singh et al. constructed a branched output convolutional neural network architecture as an inverse model, used 10,000 sets of data to train the model, and predicted the geometric parameters of the proposed polygonal patch antenna. Kabir et al. divided the data set into several groups and trained multiple networks, combining them into an inverse model for filter performance optimization. Although these studies have achieved the inverse design of antennas, there are generally problems such as inaccurate prediction of the target electromagnetic response, excessive amount of training data required for the model, and poor generalization. Therefore, it is urgent to propose an antenna inverse design method that is efficient, accurate, has low data dependency, and has strong generalization. Summary of the Invention

[0005] The purpose of the present invention is to provide an antenna structure parameter optimization method based on image translation, which solves the problems in the existing technology such as high requirements on engineers' prior knowledge, high computational and time costs, insufficient accuracy, and poor generalization for different antenna design tasks.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An antenna structure parameter optimization method based on image translation comprises the following steps:

[0008] S1. Based on the basic structure of the target antenna, select the structural parameters that need to be optimized and determine the target electromagnetic response performance parameters and the number of training samples.

[0009] S2. Samples of the selected structural parameter vectors are generated through the Latin hypercube sampling algorithm LHS, and the electromagnetic response data corresponding to the structural parameters are obtained through the Python-HFSS joint simulation script.

[0010] S3. Convert the structural parameters and electromagnetic response data into image data to form a data set, and use the data set to train the image translation network.

[0011] S4, using the trained image translation network to perform image translation on the target electromagnetic response image to obtain a corresponding structural parameter image;

[0012] S5. Extract parameters from the structural parameter image to obtain the structural parameters of the target antenna, and complete the antenna design.

[0013] Preferably, the structural parameters in S1 include the length, width and distance to the edge of the substrate of each component in the antenna; the electromagnetic response performance parameters include the S11 curve and the radiation pattern.

[0014] Preferably, the sample generation of the structure parameter vector includes the following steps:

[0015] First, the sample size and parameter dimension are determined, and the range of each parameter dimension is divided into N equal probability intervals.

[0016] Secondly, a sample point is randomly selected in each equal probability interval of each parameter dimension and randomized.

[0017] Finally, the sample points of each parameter dimension are combined to form N samples.

[0018] Preferably, in S3, the image translation network is an improved pixpix image translation network, including a generator and a discriminator.

[0019] The generator uses U-Net as the backbone network. In order to enhance the feature extraction and processing of multiple image channel superpositions, the SE attention mechanism is added between the upsampling layer, downsampling layer and skip layer connections in U-Net. At the same time, the CondConv dynamic convolution enhanced translation network is used to replace the ordinary convolution in the upsampling layer and downsampling layer to further improve the feature extraction capability.

[0020] The SE attention mechanism is also added between the downsampling layers of the discriminator, and the CondConv dynamic convolution enhanced translation network is used to replace the ordinary convolution.

[0021] Preferably, the CondConv dynamic convolution enhanced translation network calculates the weights of different convolution kernels according to the input features before ordinary convolution, and dynamically calculates the convolution kernels for different inputs, so that the network can adjust its convolution operation according to the characteristics of each data sample, thereby better processing the data characteristics of the superimposed channel. Finally, the weighted sum of each convolution kernel is obtained to obtain the dynamic convolution kernel.

[0022] The calculation expression is:

[0023] Output(x)=σ((α1W1+…+α n W n )*x);

[0024] Where σ is the activation function; W1,...,W n is the convolution kernel; α1,...,α n is the weight corresponding to each convolution kernel; x is the input feature map.

[0025] Preferably, the loss function of the generator is composed of a generative adversarial loss function and a pixel loss function. The generative adversarial loss is used to guide the model to converge in the direction of accurately fitting the real electromagnetic response. The pixel loss is used to adjust the consistency of the generator output image and the real value in surface semantics (such as image color and basic shape). The adversarial loss function is:

[0026]

[0027] The pixel loss function is:

[0028]

[0029] In the formula, Image GT and Image G Real images and images generated by the generator respectively; and are the generative adversarial losses for the conditional information path and the noise path, respectively; Pixel loss for applying smoothL1; is the final generator loss; G and D are the generator and discriminator respectively; Image and Noise are the input conditional information image and random noise image respectively; is the final generator loss; is the expected value of the image under the distribution of training set data; is the expected value of the random noise distribution.

[0030] Preferably, S5 further includes the following steps:

[0031] The performance of the trained image translation network is measured by calculating the relative error between the electromagnetic response corresponding to the structural parameters output by the image translation network and the actual electromagnetic response:

[0032]

[0033] Where misfit1 and misfit2 are the relative error values ​​of the S11 curve and the radiation pattern, respectively. The lower the error value, the higher the prediction accuracy of the image translation network. n is the number of test samples. They are the S11 curve and radiation pattern predicted and fitted by the image translation network; These are the label values ​​of the S11 curve and radiation pattern, respectively.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] This invention utilizes an improved image translation network to translate an input electromagnetic response image into a corresponding structural parameter image within seconds, eliminating the need for an iterative search process. Furthermore, compared to existing deep learning algorithms, the amount of training data required is significantly reduced while maintaining high prediction accuracy. The time and computational costs of the entire optimization process are limited to the data collection and model training phases. Multiple design tasks for the same antenna do not require an iterative search process. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of the structure of a multi-frequency dipole antenna according to an embodiment of the present invention; (a) is a schematic diagram of the front structure; (b) is a schematic diagram of the back structure;

[0037] Figure 2 Schematic diagram of the optimization method of the present invention;

[0038] Figure 3 A schematic diagram of converting one-dimensional vector data into image data according to an embodiment of the present invention;

[0039] Figure 4 Schematic diagram of the structure of the image translation network according to an embodiment of the present invention; (a) is a schematic diagram of the generator structure; (b) is a schematic diagram of the discriminator structure; (c) is a schematic diagram of the detailed structure of the upsampling layer and downsampling layer in the generator and discriminator;

[0040] Figure 5 This is a schematic diagram of the SE attention mechanism principle of an embodiment of the present invention;

[0041] Figure 6 This is a schematic diagram of the CondConv dynamic convolution principle in an embodiment of the present invention;

[0042] Figure 7 Schematic diagram of the training process of the image translation network according to an embodiment of the present invention;

[0043] Figure 8 Schematic diagram of some test results in the test set of an embodiment of the present invention; (a) is a comparison of the results of test example 1 in Table 3; (b) is a comparison of the results of test example 2 in Table 3; and (c) is a comparison of the results of test example 3 in Table 3. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0045] See also Figures 1-8A method for optimizing antenna structural parameters based on image translation was developed. This method generates random structural parameter samples and automatically collects data using a co-simulation script. The raw data is then converted to construct a dataset. The image translation network is trained, and the trained model is used to rapidly implement various antenna design tasks. Finally, the model is tested and evaluated.

[0046] The specific steps include:

[0047] S1. Based on the basic structure of the target antenna, select the structural parameters that need to be optimized and determine the target electromagnetic response performance parameters; optimize the impedance matching and radiation pattern to obtain good resonance and omnidirectional radiation at the target frequency.

[0048] Specifically, the initial structural model of the target antenna is preliminarily constructed in the simulation software HFSS, such as Figure 1 As shown, the target antenna uses Rogers RT Duroid 5880 with a dielectric constant Er=2.2 and a loss tangent tanδ=0.0009 as a substrate material, and the size of the substrate is 100 mm*55 mm*0.787 mm.

[0049] The design targets are: return loss below -10dB at 0.9GHz, 1.8GHz, and 2.4GHz to achieve good resonance. Furthermore, the XoZ plane radiation pattern non-circularity (the difference between the maximum and minimum values ​​of the pattern curve) at these three frequencies is less than 3dB to achieve omnidirectional radiation characteristics. The 12 main parameters that determine the antenna performance in this embodiment are listed in Table 1.

[0050] Table 1 Target antenna performance parameters

[0051]

[0052]

[0053] S2. Samples of the selected structural parameter vectors are generated through the Latin hypercube sampling algorithm LHS, and the electromagnetic response data corresponding to the structural parameters are obtained through the Python-HFSS joint simulation script.

[0054] Use vector P to represent the structural parameters, that is, P = [L1, L2, L3, L4, L5, L6, L7, W1, W2, W3, W4] T. The Latin Hypercube Sampling algorithm (LHS) was used to generate 800 sets of structural parameters. Then, Python was used to automatically call the HFSS interface. The structural parameters generated by LHS were input into HFSS for simulation to obtain the corresponding electromagnetic response data. P, S, and R are used to represent the structural parameters, S11 curve, and radiation pattern, respectively. Each structural parameter P and its corresponding parameter S and radiation pattern R are used as a data sample to form a data set. Among them, 600 groups are used for training the model, 150 groups are used for verification, and 50 groups are used to test the model performance.

[0055] The Latin Hypercube Sampling (LHS) algorithm comprises the following steps:

[0056] Determine the sample size N and the parameter dimension D, which is the number of elements in the structural parameter vector P. For each parameter dimension D, divide its range into N equally probable intervals, ensuring that each equally probable interval is evenly considered during the sampling process. Randomly select a sample point within each equally probable interval for each parameter dimension D. Then, randomize the sample points within each parameter dimension D. Finally, combine the sample points for each parameter dimension D to form N samples.

[0057] S3, converting the structural parameters and electromagnetic response data into image data to form a data set, and training the image translation network through the data set;

[0058] In order to convert the one-dimensional vector structure parameters P, S11 curve S and radiation pattern R into corresponding image data while keeping the semantics of the data unchanged, the following three transformations are performed respectively:

[0059] For the structure parameter P, a blank image of size 256 pixels * 256 pixels is divided into 12 rectangular areas of equal size at equal intervals in the vertical direction, and filled with black pixels according to the proportional coefficient η. The relationship between the proportional coefficient η and the corresponding parameters in the structure parameter P can be expressed linearly:

[0060]

[0061] Therefore, the proportion of black pixels in the 12 rectangular areas can directly correspond to the 12 parameters in the structural parameter P. The complete image after filling is the structural parameter image P′:

[0062] P′={p′1,p′2…p′ 12};

[0063] in,

[0064] p′ i =η i ×s,i=1,2…12;

[0065] Where ηi is the proportional coefficient of the i-th parameter; p i is the i-th parameter in the structural parameter P, p imin is p in Table 1 i The left edge of the range, p imax is p in Table 1 i The right boundary of the range; s is the area of ​​12 rectangular regions of equal size.

[0066] For the S11 curve S: First, draw its curve on a blank image of size 256 pixels by 256 pixels, and fix the horizontal and vertical coordinate ranges of the image so that each pixel value can directly correspond to the data characteristics of the electromagnetic response curve. To increase the pixel ratio in the image after the one-dimensional vector conversion, the S11 curve is symmetrically folded, and the area enclosed by the two curves is filled with black pixels. Then, the image is rotated 90° in the opposite direction to obtain S′, and the image size is adjusted to maintain 256 pixels by 256 pixels. This image transformation enriches the image features while maintaining the semantic information.

[0067] For the radiation pattern R: A blank image of 256 pixels by 256 pixels is divided vertically into three equal regions. Radiation pattern curves for 0.9 GHz, 1.8 GHz, and 2.4 GHz are plotted. Unlike the S11 curve, the parameter values ​​for the radiation pattern R can be both positive and negative. For ease of representation, the region enclosed by the radiation pattern curve and the horizontal line at 0 dB is filled to obtain R′. The three regions are filled with pixel values ​​of RGB(255, 0, 0), RGB(0, 255, 0), and RGB(0, 0, 255), respectively. This ensures that the feature values ​​within the three regions of the RGB channels of R′ do not interfere with each other, achieving optimal feature representation.

[0068] The image translation network adopts an improved pixpix image translation network, including a generator and a discriminator;

[0069] like Figure 4 As shown in the figure, the generator uses U-Net as the backbone network, and adds the SE attention mechanism between the upsampling layer, downsampling layer and skip layer connections in U-Net. At the same time, the CondConv dynamic convolution enhanced translation network is used to replace the ordinary convolution in the upsampling layer and downsampling layer to further improve the feature extraction capability.

[0070] The SE attention mechanism is also added between the downsampling layers of the discriminator, and the CondConv dynamic convolution enhanced translation network is used to replace the ordinary convolution.

[0071] like Figure 5As shown in Figure 2, the SE attention mechanism includes the Squeeze stage and the Excitation stage. The process of generating an attention feature map through the SE attention mechanism is as follows:

[0072] In the Squeeze stage, after performing a global average pooling operation on the input feature map U of size H×W×C, the spatial information (width and height) of each channel is compressed into a single global feature, thereby obtaining a feature map of size 1×1×C.

[0073] In the Excitation stage, the 1×1×C feature map obtained in the above steps is first reduced in dimension (down-sampled) and then increased in dimension (upsampling). After the dimensionality reduction and upsampling processes, the obtained feature map is normalized to generate the final channel attention weight.

[0074] The attention weights are element-wise multiplied with the original feature map to generate the feature map X with attention. The SE attention mechanism distinguishes the importance of different features for model learning through attention, helping the model learn important features.

[0075] like Figure 6 As shown in the figure, the CondConv dynamic convolution enhanced translation network calculates the weights of different convolution kernels according to the input features before the ordinary convolution, and finally obtains the dynamic convolution kernel by weighted summation of each convolution kernel; the calculation expression is:

[0076] Output(x)=σ((α1W1+…+α n W n )*x);

[0077] Where σ is the activation function; W1,...,W n is the convolution kernel; α1,...,α n is the weight corresponding to each convolution kernel; x is the input feature map.

[0078] By dynamically calculating the convolution kernel for different inputs, the network can adjust its convolution operation according to the characteristics of each data sample, thereby better processing the data characteristics of the superimposed channels.

[0079] The generator's loss function consists of a generative adversarial loss function and a pixel loss function. The generative adversarial loss function is used to guide the model to converge in the direction of accurately fitting the real electromagnetic response. The pixel loss function is used to adjust the consistency of the generator's output image and the real value in terms of surface semantics (such as image color and basic shape). The adversarial loss function is:

[0080]

[0081] The pixel loss function is:

[0082]

[0083] In the formula, Image GT and Image G Real images and images generated by the generator respectively; and are the generative adversarial losses for the conditional information path and the noise path, respectively; Pixel loss for applying smoothL1; is the final generator loss; G and D are the generator and discriminator respectively; Image and Noise are the input conditional information image and random noise image respectively; is the final generator loss; is the expected value of the image under the distribution of training set data; is the expected value of the random noise distribution.

[0084] like Figure 7 Figure 1 shows a schematic diagram of the training process for an image translation model. In this embodiment, conditional information (i.e., electromagnetic response image) and random noise are fed into the network via two separate paths. The generator then generates two predicted outputs (conditional output and noise output). The discriminator then determines the truth or falsehood of data pairs consisting of the two outputs and the label input, along with their corresponding label data pairs. The discriminator then outputs two discriminant values, which are used to calculate the generative adversarial loss and backpropagate to update the network parameters of the generator and discriminator.

[0085] In the conditional information path, the generator is able to learn the intrinsic mapping relationship between the antenna's electromagnetic response and structural parameters. In the noise path, a noise bottleneck module is added, consisting of a downsampling layer and a SE attention layer. By limiting the flow of noise information through the noise bottleneck module, the generator in the improved pixpix image translation network does not directly extract data features from the noise, but instead learns how to map the noise to the target distribution. This noise-driven process helps the model learn the entire distribution of the target domain. The dual-input path structure allows the model to simultaneously learn features from real images and the randomness provided by noise, thereby better learning the data distribution of the output domain under a limited data range.

[0086] S4. The trained image translation network is used to translate the target electromagnetic response image into the corresponding structural parameter image. The dual-input path architecture enhances the network's ability to learn the randomness of noise, helping it achieve better training results with smaller datasets and further reducing simulation costs. The SE attention mechanism and CondConv dynamic convolution enhance the translation network's ability to extract features from multi-channel input data, ensuring the model accurately learns important information about the antenna's electromagnetic response.

[0087] By inputting the target electromagnetic response image into a well-trained image translation network, the corresponding structural parameter image can be obtained within a few seconds, and the optimization can be completed directly without repeated optimization search.

[0088] S5. Extract parameters from the structural parameter image to obtain the structural parameters of the target antenna, and complete the antenna design.

[0089] After the target electromagnetic response image is input into the image translation network to obtain the output image, the output structural parameter image is subjected to parameter extraction to obtain the structural parameters. This parameter extraction process is the inverse of the process of converting the structural parameter vector P to the image P′ in the S3 data conversion.

[0090] After extracting the structural parameters of the target antenna, error calculation can be performed to verify whether the parameters meet the requirements. The performance of the trained image translation network is measured by calculating the relative error between the electromagnetic response corresponding to the structural parameters output by the image translation network and the actual electromagnetic response:

[0091]

[0092] Where, misfit1 and misfit2 are the relative error values ​​of the S11 curve and the radiation pattern respectively; n is the number of test samples; They are the S11 curve and radiation pattern predicted and fitted by the image translation network; These are the label values ​​of the S11 curve and radiation pattern, respectively.

[0093] In this embodiment, the 50 groups of test data are subjected to image translation to extract the structural parameters predicted by the model, and the corresponding electromagnetic responses are input into the simulation to calculate the relative errors with the true values.

[0094] As shown in Table 2, Er1, Er2, and Er3 represent the relative errors between the model-predicted radiation patterns of the antenna structure at 0.9 GHz, 1.8 GHz, and 2.4 GHz, respectively, and the true values. Es represents the relative error between the S11 curves corresponding to the structural parameters output by the model and the true values. The relative errors are all less than 0.1, indicating that the image translation network can well fit the input target electromagnetic response.

[0095] Table 2 Relative error table of test set

[0096] Es Er1 Er2 Er3 0.096 0.051 0.082 0.079

[0097] In order to show the results more intuitively, some comparison results are selected in the test set as follows Figure 8As shown, the S11 curve and radiation pattern curve corresponding to the antenna structure predicted by the image translation network of this embodiment have almost no deviation from the S11 curve and radiation pattern curve corresponding to the actual antenna structure, indicating that the proposed image translation network has excellent electromagnetic response fitting performance. The following Table 3 compares the structural parameters with the actual values:

[0098] Table 3 Comparison of the actual values ​​of structural parameters of some examples in the test set and the predicted structural parameter values ​​of the present invention

[0099] <![CDATA[L1]]> <![CDATA[L2]]> <![CDATA[L3]]> <![CDATA[L4]]> <![CDATA[L5]]> <![CDATA[L6 <!-- 8 -->]]> Test Example 1 38.92 / 36.86 32.36 / 32.82 26.66 / 26.09 46.26 / 44.12 11.38 / 11.44 46.99 / 45.69 Test Example 2 36.49 / 35.75 31.52 / 31.74 26.09 / 25.75 44.89 / 43.07 12.24 / 11.83 44.54 / 46.49 Test Example 3 36.67 / 36.78 32.98 / 33.23 24.09 / 24.15 43.04 / 41.96 12.26 / 12.04 47.77 / 47.43 <![CDATA[L7]]> <![CDATA[L G ]]> <![CDATA[W1]]> <![CDATA[W2]]> <![CDATA[W3]]> <![CDATA[W4]]> Test Example 1 78.47 / 77.80 34.54 / 33.62 2.71 / 2.54 5.74 / 5.92 2.05 / 2.09 0.83 / 0.96 Test Example 2 78.43 / 78.02 39.41 / 38.07 2.06 / 2.14 5.29 / 5.68 2.08 / 1.79 0.98 / 1.01 Test Example 3 76.70 / 76.54 34.72 / 34.81 2.45 / 2.30 5.81 / 5.59 2.29 / 2.06 0.99 / 0.91

[0100] As shown in Table 3, the structural parameters predicted by the model are not exactly the same as the true values, but the electromagnetic responses are very similar. The image translation network has well learned the underlying distribution of the data, rather than just the surface relationship of the training data, demonstrating good data generalization.

Claims

1. A method for optimizing antenna structure parameters based on image translation, characterized in that: The following steps are involved: S1. Based on the basic structure of the target antenna, select the structural parameters that need to be optimized and determine the target electromagnetic response performance parameters; S2. Generate samples of the selected structural parameter vectors using the Latin hypercube sampling algorithm (LHS), and obtain the electromagnetic response data corresponding to the structural parameters using the Python-HFSS joint simulation script; S3, converting the structural parameters and electromagnetic response data into image data to form a data set, and training the image translation network using the data set; comprising the following steps: For the structure parameter P, the size is 256 pixels The 256-pixel blank image is divided into 12 rectangular areas of equal size at equal distances in the vertical direction and the proportional coefficients are used to calculate the image area. Fill with black pixels; For S11 Curve S: First, set the size to 256 pixels The curve is drawn on a blank image of 256 pixels, and the horizontal and vertical coordinate ranges of the image are fixed so that each pixel value can directly correspond to the data characteristics of the electromagnetic response curve; the S11 curve is symmetrically folded, and the area surrounded by the two curves is filled with black pixels. Then the image is rotated 90° in the opposite direction to obtain ; For the radiation pattern R: set the size to 256 pixels The 256-pixel blank image is divided into three equal-sized areas vertically, and the radiation pattern curves of 0.9GHz, 1.8GHz, and 2.4GHz are drawn respectively; the area surrounded by the radiation pattern curve and the horizontal line at the 0dB value is filled to obtain ; Fill three areas with pixel values ​​of RGB (255,0,0), RGB (0,255,0), and RGB (0,0,255) respectively; The image translation network is an improved pixpix image translation network, including a generator and a discriminator; The generator uses U-Net as the backbone network and adds the SE attention mechanism between the upsampling layer, downsampling layer and skip layer connection in U-Net. At the same time, the CondConv dynamic convolution enhanced translation network is used to replace the ordinary convolution in the upsampling layer and downsampling layer. The SE attention mechanism is also added between the downsampling layers of the discriminator, and the CondConv dynamic convolution enhanced translation network is used to replace the ordinary convolution; S4, using the trained image translation network to perform image translation on the target electromagnetic response image to obtain a corresponding structural parameter image; S5. Extract parameters from the structural parameter image to obtain the structural parameters of the target antenna, and complete the antenna design.

2. The antenna structure parameter optimization method based on image translation according to claim 1, characterized in that: The structural parameters in S1 include the length, width and distance to the edge of the substrate of each component in the antenna; the electromagnetic response performance parameters include the S11 curve and the radiation pattern.

3. The antenna structure parameter optimization method based on image translation according to claim 1, characterized in that: The sample generation of the structural parameter vector includes the following steps: First, determine the sample size and parameter dimension, and divide the range of each parameter dimension into N equal probability intervals; Secondly, a sample point is randomly selected in each equal probability interval of each parameter dimension and randomized; Finally, the sample points of each parameter dimension are combined to form N samples.

4. The antenna structure parameter optimization method based on image translation according to claim 1, characterized in that: The CondConv dynamic convolution enhanced translation network first calculates the weights of different convolution kernels according to the input features before the ordinary convolution, and finally obtains the dynamic convolution kernel by weighted summation of each convolution kernel. The calculation expression is: ; Where, is the activation function; is the convolution kernel; The weight corresponding to each convolution kernel; is the input feature map.

5. The method for optimizing antenna structure parameters based on image translation according to claim 4, characterized in that: The loss function of the generator is composed of a generation adversarial loss function and a pixel loss function. The adversarial loss function is: ; ; The pixel loss function is: ; ; Where, and Real images and images generated by the generator respectively; and are the generative adversarial losses for the conditional information path and the noise path, respectively; For application Pixel loss; is the final generator loss; and They are the generator and the discriminator respectively; , are the input conditional information image and random noise image respectively; is the final generator loss; is the expected value of the image under the distribution of training set data; is the expected value of the random noise distribution.

6. The method for optimizing antenna structure parameters based on image translation according to claim 1, characterized in that: S5 also includes the following steps: The performance of the trained image translation network is measured by calculating the relative error between the electromagnetic response corresponding to the structural parameters output by the image translation network and the actual electromagnetic response: ; ; Where, , are the relative error values ​​of the S11 curve and radiation pattern respectively; is the number of test samples; , They are the S11 curve and radiation pattern predicted and fitted by the image translation network; , These are the label values ​​of the S11 curve and radiation pattern, respectively.

Images