A patch antenna self-decoupling design method based on deep learning and patch antenna

Abstract

The application discloses a patch antenna self-decoupling design method based on deep learning and a patch antenna. The design method comprises multi-section point electromagnetic coupling input parameter planning and data generation, model training based on deep learning, and prediction and verification based on multi-target screening. The input parameter planning method based on multi-section point electromagnetic coupling is used to drive the deep learning training to optimize the neural network model, and a multi-target screening mode combining port decoupling and pattern improvement is used to realize the patch antenna self-decoupling design based on deep learning. The designed patch antenna has the advantages of port decoupling and pattern synchronous improvement, low profile and simple structure.

Description

Technical Field

[0001] The present invention relates to a patch antenna design method and a patch antenna, and in particular to a patch antenna self-decoupling design method and a patch antenna based on deep learning. Background Art

[0002] Closely spaced antenna elements can induce electromagnetic mutual coupling, leading to radiation distortion, port impedance mismatch, and system performance degradation. Therefore, effective decoupling technology is required to reduce the coupling effect and thus improve the antenna's radiation efficiency, pattern stability, and port matching characteristics.

[0003] Patch antenna self-decoupling technology achieves decoupling without adding additional parasitic structures, reducing structural complexity, profile height, overall size, and cost. Introducing deep learning into the self-decoupling design of closely spaced patch antennas can reduce manual intervention while improving the optimization efficiency and performance ceiling of the decoupling scheme. Therefore, researching and developing a deep learning-based self-decoupling design method for patch antennas has important engineering applications and research significance.

[0004] Most of the reported decoupling methods for closely spaced patch antennas lack artificial intelligence (AI) technology and primarily focus on port decoupling, failing to effectively address the pattern distortion caused by mutual coupling. Designs that simultaneously achieve port decoupling and pattern recovery, due to the dual optimization objectives involved, often rely heavily on manual experience, resulting in long research cycles and low optimization efficiency. Existing AI-based antenna decoupling research is relatively limited, with only one metasurface decoupling design method for patch antennas based on reinforcement learning and agent modeling. While this method improves port decoupling performance, it is not suitable for self-decoupling of patch antennas, primarily due to differences in prior knowledge, prior knowledge-based input parameter planning, and the dual-objective optimization of port decoupling and pattern improvement. The resulting results suffer from increased profile height and structural complexity caused by the additional loading of the metasurface structure, as well as limited port decoupling and an unknown pattern improvement. Summary of the Invention

[0005] Purpose of the invention: In response to the above-mentioned existing technologies, a deep learning-based patch antenna self-decoupling design method and patch antenna are proposed, which can quickly achieve significant port decoupling and improvement of the radiation pattern, and have a lower profile and simple structure.

[0006] Technical solution: A deep learning-based self-decoupling design method for patch antennas, including:

[0007] Step 1: Based on the prior knowledge of regional electromagnetic distribution characteristics and coupling effects, set the multi-point input structural parameters, and combine the metal patch length and feeding position to form the input structural parameters; obtain the value of each set of input structural parameters through full-wave simulation. S Parameters, maximum gain and positive zenith gain, and extract the difference between the maximum gain and the positive zenith gain to form a data set; wherein, the S The parameter is used to measure the matching and port decoupling effects, and the gain difference is used to measure the degree of pattern distortion;

[0008] The specific method for setting the input structure parameters includes: based on the fact that the patch antenna has vertical and horizontal symmetry when working in linear polarization, the original rectangular metal patch is divided into four uniform areas through the cross axis; then, four points with equal longitudinal spacing are set in one of the quarter areas, which are respectively marked as A, B, C and D, and a vertical line segment is marked as E; the distances between each point and the line segment E and the edge of the original rectangular metal patch are respectively marked as d 1. d 2. d 3. d 4 and d 5. The vertical distance between point A and point D is recorded as l ; The straight line between the two adjacent points is connected together with the straight line segment E to form the side of the designed metal patch in the quarter area; the longitudinal length of the metal patch is set to l p , the distance between the feeding position and the upper and lower symmetry lines of the metal patch is recorded as k , thus forming d 1. d 2. d 3. d 4. d 5. l 、 l p 、 k There are eight input structure parameters in total;

[0009] Step 2: Use the data set to perform deep learning-based model training, and after pre-verification, obtain a neural network model with convergence and qualified quality;

[0010] Step 3: Use the neural network model obtained in step 2 to perform prediction and verification based on multi-objective screening.

[0011] Furthermore, the center distance between two adjacent rectangular metal patches is 0.44 l 0-0.45 l 0, with a margin of 0.003 l 0-0.004 l 0, the preset parameter range of the data set is as follows:d 1. d 2. d 3 and d All 4 values ​​are within 0.018 l 0-0.054 l 0, d 5 value at 0.018 l 0-0.036 l 0, l The value is 0.12 l 0-0.19 l 0, l p The value is 0.41 l 0-0.44 l 0, k The value is 0.088 l 0-0.124 l 0, l 0 is the wavelength of air corresponding to the center frequency.

[0012] Furthermore, the step 2 specifically includes: firstly dividing the data set into a training set and a test set, then inputting the data in the training set into a neural network based on deep learning, and the neural network outputs a predicted value ,in W is the weight in the neural network, b is the bias, X For input data, X =( d 1, d 2, d 3, d 4, d 5, l , l p , k ); During the training process, the parameters of the neural network are adjusted by the Adam optimizer W and b , so that the loss function gradually decreases until convergence; after passing the test set data test, a neural network model with convergence training and qualified quality is obtained.

[0013] Furthermore, the step 3 specifically includes: for the neural network model that has converged and qualified, input multiple sets of parameter groups to predict the corresponding SParameters, the difference between the maximum gain and the positive zenith gain, and then the target result is obtained through multi-objective screening. Finally, the qualified results obtained by screening are simulated in the full-wave simulation software. If the simulation results are consistent with the predicted results or the deviation is within the preset range, the verification is completed.

[0014] Furthermore, the conditions for the multi-target screening include: S 11 The minimum matching value is less than -15 dB, the mutual coupling level within the working band is less than -20 dB, the mutual coupling zero point value is less than -40 dB, and the mutual coupling zero point is within the working band. The difference between the maximum gain and the positive zenith gain is less than 0.8 dB.

[0015] A patch antenna obtained by a deep learning-based self-decoupling design method for patch antennas.

[0016] Beneficial Effects: Most existing decoupling design methods for closely spaced patch antennas do not employ AI technology, and decoupling methods based on AI technology are not suitable for patch antenna self-decoupling. They also suffer from problems such as high profile, complex structure, limited port decoupling, and lack of attention to pattern improvement. This invention proposes an input parameter planning method based on multi-segment point electromagnetic coupling to drive deep learning training and optimize a neural network model. This method, combined with a multi-objective screening method for port decoupling and pattern improvement, implements a deep learning-based self-decoupling design for patch antennas. The designed patch antenna has the advantages of simultaneous port decoupling and pattern improvement, a low profile, and a simple structure.

[0017] Specifically, based on the symmetry of linear polarization, the original rectangular metal patch is divided into four symmetrical areas, and the left or right side of each area is marked with four points with equal spacing in the Y direction from far to near, as well as a vertical line. The areas between the points and the vertical line can finely adjust the strength of electric coupling and magnetic coupling as the structural parameters change, realizing multi-segment electromagnetic coupling input parameter planning, which is conducive to improving the training quality based on deep learning neural network models and the possibility of simultaneously achieving port decoupling and directional pattern improvement.

[0018] The difference between the maximum gain and the positive zenith gain is defined as the screening condition for pattern improvement. Combined with the screening of port mutual coupling level and matching within the working frequency band, a three-objective screening mechanism is formed. By combining it with a large number of predictions from a trained and converged neural network model, it is further ensured that the output parameter corresponding model can simultaneously achieve port decoupling and pattern improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 2 is a cross-sectional view of the antenna of the present invention.

[0020] Figure 2 Schematic diagram for setting the input structural parameters of the antenna.

[0021] Figure 3 Design a flow chart for the overall method.

[0022] Figure 4 This is a schematic diagram of the antenna structure designed by the embodiment method;

[0023] Figure 5 The antenna of the embodiment before and after decoupling S Parameter comparison, where (a) corresponds to before decoupling and (b) corresponds to after decoupling.

[0024] Figure 6 1 is a comparison of the radiation patterns of the antenna of the embodiment before and after decoupling, where (a) corresponds to before decoupling and (b) corresponds to after decoupling. DETAILED DESCRIPTION

[0025] The present invention will be further explained below with reference to the accompanying drawings.

[0026] A decoupling design method for densely packed patch antennas based on deep learning, such as Figure 1 As shown in FIG, the antenna is mainly composed of a top metal layer 1, a dielectric substrate 1 2, a dielectric substrate 2 3, a bottom metal ground 4 and a metal probe 5. Figure 2 As shown, the top metal layer 1 is mainly composed of two symmetrical and closely arranged metal patches 101. Two metal probes 5 pass through the dielectric substrate 2 3 and connect to the corresponding metal patches 101, and are both located below the center line of the metal patches 101.

[0027] Input structure parameter settings: Figure 2 The two dotted boxes 102 are the original rectangular metal patches before the shape of the metal patch 101 is designed. The center distance between the two rectangular metal patches is 0.44. l 0-0.45 l 0, l 0 is the air wavelength corresponding to the center frequency, and the margin is 0.003 l 0-0.004 l 0. Based on the fact that the patch antenna has symmetry in the vertical and horizontal directions when working in linear polarization, the original rectangular metal patch is divided into four uniform regions through the cross axis. The structures of the four regions are consistent, so as to reduce the input structural parameters, reduce the data complexity, and improve the design efficiency. According to the characteristics of the electromagnetic field distribution, four points with equal spacing in the Y direction are set in one of the quarter regions, which are marked as A, B, C and D respectively, and a vertical line segment is marked as E; the distance from each point to the edge of the original rectangular metal patch, that is, the distance from the edge of the dotted box 102 from top to bottom is recorded as d 1. d 2. d 3 and d 4. The distance between the vertical line segment E and the edge of the original rectangular metal patch is recorded asd 5. The distance between point A and point D in the Y direction is recorded as l The straight line between the two adjacent points is connected and together with the straight line segment E constitutes the side of the metal patch 101 in the quarter area.

[0028] By adjusting and setting the above input structural parameters, the side of the patch presents four areas with different electromagnetic field distribution characteristics: the area between point A and point B presents strong electric field characteristics, and the coupling is strong electric field coupling, which can be achieved through d 1 and d 2 to control the coupling strength; the area between point B and point C presents the characteristics of strong electric field and weak magnetic field. When coupling, it is a combination of strong electric coupling and weak magnetic coupling, which can be controlled by d 2 and d 3 is controlled; the area between point C and point D presents the characteristics of strong magnetic field and weak electric field. When coupled, it is a strong magnetic coupling combined with a weak electric coupling, which can be controlled by d 3 and d 4 for regulation; the line segment E area presents the characteristics of strong magnetic field, and the coupling is strong magnetic field coupling, which can be controlled by d 5. In addition, in order to adapt the resonant frequency and matching of the antenna during the design process, the longitudinal length of the metal patch 101 is set to l p The distance between the feeding position and the upper and lower symmetry lines (Y=0) of the metal patch 101 is recorded as k , so there are eight input structure parameters involved in dataset generation and training. To ensure the quality of subsequent deep learning-based training, the preset parameter ranges of the dataset are as follows: d 1. d 2. d 3 and d All 4 values ​​are within 0.018 l 0-0.054 l 0, with a step size of 0.018 l 0; d 5 value at 0.018 l 0-0.036 l 0, with a step size of 0.018 l 0; l The value is 0.12 l 0-0.19 l 0, the step size is 0.035 l 0; l p The value is 0.41 l 0-0.44 l 0, with a step size of 0.018 l 0; k The value is 0.088 l 0-0.124 l 0, with a step size of 0.018 l 0.

[0029] like Figure 3 As shown, the overall design process is divided into three steps.

[0030] The first step is the planning and data generation of multi-segment electromagnetic coupling input parameters.

[0031] By using the prior knowledge of regional electromagnetic distribution characteristics and coupling effects, multiple segment input structural parameters are set, and eight input structural parameters are formed in combination with the metal patch length and feeding position. The full-wave simulation software is then called to obtain the value of each set of parameters. S Parameters, maximum gain and positive zenith gain, and obtain the difference between the maximum gain and the positive zenith gain through data extraction, and then generate a data set, which should include several groups of parameters, the corresponding parameters of each group of parameters S Parameters, the difference between the maximum gain and the positive zenith gain corresponding to each set of parameters. S The parameter is used to measure the matching and port decoupling effects, and the gain difference is used to measure the degree of pattern distortion.

[0032] The second step is model training based on deep learning.

[0033] In this embodiment, the deep neural network specifically uses a variational autoencoder (VAE). First, the prepared data set is divided into a training set and a test set. Then, the data in the training set is fed into the neural network based on deep learning. The neurons in each layer calculate the output of the next layer through weighted summation and activation function until the final output layer, the predicted value ŷ It can be obtained by the following formula:

[0034]

[0035] in W is the weight in the neural network, b is the bias, X To input data, the present invention inputs data X =( d 1, d 2, d 3, d 4, d 5, l , l p , k ). During this process, the parameters of the neural network are adjusted through the Adam optimizer Wand b , gradually decreasing the loss function and continuously improving the model's predictive capabilities until convergence results in a neural network model that can be used for prediction. To preliminarily verify the trained neural network model, the test set is input into the network model. If the output matches the simulation results, the third step is entered. If not, the neural network model is readjusted until a neural network model with convergence and acceptable quality is obtained.

[0036] The third step is prediction and verification based on multi-target screening.

[0037] For a neural network model that has converged and has good quality, a large number of parameter groups can be input, for example, no less than 100,000 groups of parameter groups, to predict the corresponding S The difference between the maximum gain and the positive zenith gain is then used to determine the target result through multi-objective screening, where the multi-objective screening conditions are as follows: S 11 The minimum matching value is less than -15 dB, the mutual coupling level within the operating band is less than -20 dB, the mutual coupling zero value is less than -40 dB, and the mutual coupling zero point is within the operating band. The difference between the maximum gain and the positive zenith gain is less than 0.8 dB. Finally, the results that meet the requirements are simulated in full-wave simulation software. If the simulation results are basically consistent with the predicted results or the deviation is within the preset range, the verification is completed.

[0038] The dielectric substrate used in this embodiment is RO4003C, and the electrical size of the antenna is 0.91 l 0×0.41 l 0, the cross section is 0.06 l 0. This embodiment obtains a neural network model by training 2916 sets of data sets, and then uses 179200 sets of parameters to participate in the prediction, and then obtains the following through the multi-objective screening command: Figure 4 The schematic diagram of the structure of the metal patch 101 shown in FIG. S Parameter comparison Figure 5 This embodiment covers the operating frequency band of 5.14 GHz to 5.48 GHz, with a relative bandwidth of 6.4%. The mutual coupling level across the entire operating frequency band is below -20 dB, and the depth of the mutual coupling zero reaches -44 dB. Compared with the initial complete square patch antennas that were closely arranged before decoupling, this example significantly improves the decoupling capability and achieves a good port decoupling effect.

[0039] Figure 6 The comparison of the antenna pattern before and after decoupling is given. Figure 6As shown in (a), the antenna pattern is severely distorted before decoupling, with a 3dB beamwidth of only 53.6°, a difference of 1.07 dB between the maximum gain and the positive zenith gain, and a maximum cross-polarization level of -10 dB within 3dB. Figure 3 dB beamwidth can reach 69.7°, such as Figure 6 As shown in (b), the difference between the maximum gain and the positive zenith gain is only 0.73 dB, and the maximum cross-polarization level within 3dB is improved to -14.02 dB. It can be seen that the radiation pattern has been greatly improved.

[0040] Compared with the existing technology, the deep learning-based decoupling design method proposed in the present invention can be applied to the self-decoupling of patch antennas. The designed patch antenna has the advantages of port decoupling and synchronous improvement of the radiation pattern, low profile and simple structure.

[0041] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A deep learning-based patch antenna self-decoupling design method, characterized in that: include: Step 1: Based on the prior knowledge of regional electromagnetic distribution characteristics and coupling effects, set the multi-point input structural parameters, and combine the metal patch length and feeding position to form the input structural parameters; The full-wave simulation is used to obtain the S Parameters, maximum gain and positive zenith gain, and extract the difference between the maximum gain and the positive zenith gain to form a data set; wherein, the S The parameter is used to measure the matching and port decoupling effects, and the gain difference is used to measure the degree of pattern distortion; The specific method for setting the input structure parameters includes: based on the fact that the patch antenna has vertical and horizontal symmetry when working in linear polarization, the original rectangular metal patch is divided into four uniform areas through the cross axis; then, four points with equal longitudinal spacing are set in one of the quarter areas, which are respectively marked as A, B, C and D, and a vertical line segment is marked as E; the distances between each point and the line segment E and the edge of the original rectangular metal patch are respectively marked as d 1. d 2. d 3. d 4 and d 5. The vertical distance between point A and point D is recorded as l ; The straight line between the two adjacent points is connected together with the straight line segment E to form the side of the designed metal patch in the quarter area; the longitudinal length of the metal patch is set to l p , the distance between the feeding position and the upper and lower symmetry lines of the metal patch is recorded as k , thus forming d 1. d 2. d 3. d 4. d 5. l 、 l p 、 k There are eight input structure parameters in total; Step 2: Use the data set to perform deep learning-based model training, and after pre-verification, obtain a neural network model with convergence and qualified quality; Step 3: Use the neural network model obtained in step 2 to perform prediction and verification based on multi-objective screening.

2. The deep learning-based patch antenna self-decoupling design method according to claim 1, characterized in that: The center distance between two adjacent rectangular metal patches is 0.44 λ 0-0.45 λ 0, with a margin of 0.003 λ 0-0.004 λ 0, the preset parameter range of the data set is as follows: d 1. d 2. d 3 and d All 4 values ​​are within 0.018 λ 0-0.054 λ 0, d 5 value at 0.018 λ 0-0.036 λ 0, l The value is 0.12 λ 0-0.19 λ 0, l p The value is 0.41 λ 0-0.44 λ 0, k The value is 0.088 λ 0-0.124 λ 0, λ 0 is the wavelength of air corresponding to the center frequency.

3. The deep learning-based patch antenna self-decoupling design method according to claim 1 or 2, characterized in that: The step 2 specifically includes: firstly, dividing the data set into a training set and a test set, then inputting the data in the training set into a neural network based on deep learning, and the neural network outputs a predicted value ,in W is the weight in the neural network, b is the bias, X For input data, X =( d 1, d 2, d 3, d 4, d 5, l , l p , k ); During the training process, the parameters of the neural network are adjusted by the Adam optimizer W and b , so that the loss function gradually decreases until convergence; after passing the test set data test, a neural network model with convergence training and qualified quality is obtained.

4. The deep learning-based patch antenna self-decoupling design method according to claim 3, characterized in that: The step 3 specifically includes: for a neural network model with convergence and qualified training quality, input multiple sets of parameter groups to predict the corresponding S Parameters, the difference between the maximum gain and the positive zenith gain, and then the target result is obtained through multi-objective screening. Finally, the qualified results obtained by screening are simulated in the full-wave simulation software. If the simulation results are consistent with the predicted results or the deviation is within the preset range, the verification is completed.

5. The deep learning-based patch antenna self-decoupling design method according to claim 4, characterized in that: The conditions for the multi-target screening include: S 11 The minimum matching value is less than -15 dB, the mutual coupling level within the working band is less than -20 dB, the mutual coupling zero point value is less than -40 dB, and the mutual coupling zero point is within the working band. The difference between the maximum gain and the positive zenith gain is less than 0.8 dB.

6. A patch antenna obtained according to any one of claims 1-5, wherein the patch antenna self-decoupling design method based on deep learning is used.

Images