Millimeter wave filtering antenna and intelligent aided design method

Abstract

The invention discloses a millimeter wave filtering antenna and an intelligent aided design method, the antenna adopts a differential aperture coupling feed mode to realize relatively wide impedance matching bandwidth and relatively high gain, low-frequency and high-frequency radiation zero points are respectively generated by adding a short circuit patch and slotting a radiation patch, and out-of-band gain suppression is realized. The out-of-band rejection level of a low-frequency band and the out-of-band rejection level of a high-frequency band are improved by adopting the planar inverted-F antenna and adding the parasitic Deep learning is adopted on a machine learning algorithm, a proxy model of selected antenna structure parameters and corresponding performance parameters is established, rapid prediction of antenna performance of a specific parameter structure is realized, a genetic algorithm is adopted to optimize the proxy model, an antenna structure corresponding to a target performance parameter is rapidly designed, and the method has the advantages of high efficiency, flexibility and expandability. The design efficiency and performance are improved, the design cost is reduced, and the antenna structure has the characteristics of low loss, high gain and high reliability.

Description

Technical Field

[0001] The present invention relates to communication antenna technology, and in particular to a millimeter wave filtering antenna and an intelligent auxiliary design method. Background Art

[0002] Millimeter wave technology is an important branch of modern wireless communication. It refers to the technology of using electromagnetic waves with a wavelength of 1 to 10 mm (i.e., a frequency of 30 GHz to 300 GHz) for communication. This frequency band has abundant spectrum resources and can provide a larger communication bandwidth than traditional low-frequency bands, which is essential to meet the needs of 5G and future 6G mobile communication networks for high data transmission rates. The main millimeter wave frequency bands currently receiving attention include 28-30 GHz, 38 GHz, 45 GHz, 57-71 GHz, 71-76 GHz, 81-86 GHz, 100 GHz, etc. The core of the 5G mobile communication module is its RF front end, and filters and antennas are key components in the RF front end. In the design of the 5G millimeter wave RF front end, the role of filters is particularly important. They are responsible for suppressing out-of-band noise that may interfere with signal reception and transmission. With the continuous advancement of wireless communication technology, RF front-end systems are moving towards higher integration, smaller size, and higher performance. In the millimeter wave frequency band, the current trend is to compactly integrate antennas and filters to achieve functional integration. This integrated design not only optimizes the performance of the device, but also reduces the size of the device, meeting the strict space and efficiency requirements of modern mobile communication devices.

[0003] The most commonly used design methods now are fusion methods based on bandpass filters. The component to be designed by the fusion method is the filter antenna. The filter antenna is a new type of component that has both filtering and radiating functions. One of the most important characteristics of the filter antenna is that the filter and the antenna can no longer be clearly distinguished (that is, the filter antenna plays the role of filtering and radiating at the same time). The filter antenna can replace the cascade connection between the filter and the antenna (another filter antenna design method), avoid additional insertion loss from their interface, and achieve a more compact size. Since the resonant structure is designed to resonate outside the passband, they have little effect on the antenna performance within the passband. These radiation zeros can be purposefully designed to control the impedance bandwidth, passband bandwidth and frequency selectivity of the antenna. The main design challenge is to generate and control the radiation zeros. The advantages of this fusion method are low insertion loss and high radiation efficiency, and it can have a smaller size, but this modification of the antenna structure currently lacks a complete and rigorous calculation formula support, and more requires the researcher's experience and electromagnetic simulation software to obtain results.

[0004] Based on our own experience, we continuously simulate and debug the antenna's shape, structure, material and other factors. Although reliable, it is time-consuming and inefficient, making it difficult to achieve the global optimal solution and meet today's growing needs. Moreover, the existing millimeter-wave filter antenna structure not only has high loss, but also low gain, and is not reliable enough. Summary of the invention

[0005] The purpose of the present invention is to overcome the above-mentioned problems and provide a millimeter wave filter antenna, which has the advantages of low loss, high gain, and greater reliability.

[0006] Another object of the present invention is to provide an intelligent assisted design method for a millimeter wave filter antenna, which has the advantages of being efficient, flexible and scalable, and helps to improve design efficiency and performance and reduce design costs.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A millimeter wave filter antenna comprises a patch, a dielectric substrate, a planar inverted F antenna metal layer, a metal floor, a rectangular feed slot, a microstrip feed line and a short-circuit through hole;

[0009] The patches include radiation patches, driving patches, parasitic patches and short-circuit patches;

[0010] The dielectric substrates include three dielectric substrates stacked in sequence from bottom to top, and the three dielectric substrates are respectively a first dielectric substrate, a second dielectric substrate and a third dielectric substrate;

[0011] The radiation patch, parasitic patch, and planar inverted F antenna metal layer are printed on the upper surface of the third dielectric substrate; the radiation patch is provided with a plurality of slots; the driving patch and the short-circuit patch are printed on the upper surface of the second dielectric substrate; the metal floor is printed on the upper surface of the first dielectric substrate; a pair of rectangular feed slots are provided and symmetrically etched on the metal floor; the microstrip feed line is printed on the lower surface of the first dielectric substrate;

[0012] The short-circuit via comprises a first short-circuit via and a second short-circuit via. The first short-circuit via is arranged through the second dielectric substrate, and the first short-circuit via is connected between the short-circuit patch and the metal floor. The second short-circuit via is arranged through the second dielectric substrate and the third dielectric substrate, and the second short-circuit via is connected between the metal layer of the planar inverted-F antenna and the metal floor.

[0013] In a preferred embodiment of the present invention, the dielectric substrate is made of FerroA6M-E material; and the thickness of the dielectric substrate is a multiple of 0.094 mm.

[0014] Furthermore, the thickness of the first dielectric substrate is 0.564 mm, the thickness of the second dielectric substrate is 0.47 mm, and the thickness of the third dielectric substrate is 0.094 mm.

[0015] In a preferred embodiment of the present invention, the size of the driving patch is 0.14λc×0.14λc, where λc is the wavelength at the center frequency of the passband.

[0016] In a preferred embodiment of the present invention, the two rectangular feeding slots are completely identical and have a length of 0.18λc, where λc is the wavelength at the center frequency of the passband.

[0017] In a preferred embodiment of the present invention, the size of the metal layer of the planar inverted-F antenna is 0.05λc×0.03λc.

[0018] In a preferred embodiment of the present invention, the diameters of the first short-circuit through hole and the second short-circuit through hole are both 0.1 mm.

[0019] An intelligent auxiliary design method for a millimeter wave filtering antenna comprises the following steps:

[0020] The structural parameters that affect the radiation zero point and out-of-band suppression level of the antenna are extracted, which are the length of the short-circuit patch, the length of the metal layer of the planar inverted-F antenna, the width of the slot of the radiation patch, and the length of the parasitic patch.

[0021] The above four parameters are sampled exponentially;

[0022] The sampled parameters are combined in the simulation tool for full-wave simulation, and the S parameters and gains of each group from 17GHZ to 37GHZ are obtained by simulation. The above data are used to form a data set; the data set is imported into MATLAB to build a deep learning model;

[0023] After the model is established, it is optimized for the target;

[0024] By using the established proxy model, the genetic algorithm is used to optimize the antenna size parameters to obtain the optimal prediction value and its corresponding antenna parameter combination.

[0025] In a preferred embodiment of the present invention, the sampling ranges of the four parameters are:

[0026] {0.05-0.15}, {1-1.75}, {0.1-0.5}, {1-1.6}, 5 samples for each parameter, a total of 625 parameter combinations.

[0027] In a preferred embodiment of the present invention, constructing a deep learning model comprises the following steps:

[0028] For data preprocessing, the first five columns of the data exported from the three-dimensional electromagnetic field simulation tool HFSS are four structural parameters and one frequency parameter, which are output as S parameters;

[0029] x = _data(:,1:5);

[0030] y = _data(:,6);

[0031] Set the first five columns in the data list as input and put them into the above x variable, and put the sixth column in the data list as output into the above y variable;

[0032] X_norm=(X-mean(X)). / std(X);

[0033] y_norm=(y-mean(y)). / std(y);

[0034] Normalize the input and output data and put them into new variables X_norm and y_norm respectively. Normalize the data so that the eigenvalues ​​have the same scale.

[0035] cv=cvpartition(size(X_norm,1),'HoldOut',0.20);

[0036] idxTrain = training(cv);

[0037] idxTest = test(cv);

[0038] XTrain=X_norm(idxTrain,:);

[0039] yTrain=y_norm(idxTrain,:);

[0040] XTest = X_norm(idxTest,:);

[0041] yTest = y_norm(idxTest,:);

[0042] Use the cvpartition function to divide the dataset into training and test sets. 20% of the data is reserved as the test set and the remaining 80% is used for training.

[0043] Furthermore, after data preprocessing and division, MATLAB is used to build the model:

[0044] layers=[

[0045] featureInputLayer(5,'Normalization','none','Name','input')

[0046] fullyConnectedLayer(64,'Name','fc1')

[0047] reluLayer('Name','relu1')

[0048] fullyConnectedLayer(32,'Name','fc2')

[0049] reluLayer('Name','relu2')

[0050] fullyConnectedLayer(1,'Name','output')

[0051] regressionLayer('Name','regression')];

[0052] First, define and construct the input layer featureInputLayer. The input data is four structure parameters and one frequency parameter. The input feature is 5. The layer is named input.

[0053] A fully connected layer is added after the input layer. The fully connected layer is used to learn the complex mapping of the input data, perform linear transformation, and map the input data to a new space. Each neuron is connected to all neurons in the previous layer and is calculated through weights and biases. The number of neurons needs to be set based on experience and the prediction effect of the model. The number of neurons in this layer is set to 64.

[0054] Add an activation layer reluLayer and set the activation function to introduce nonlinearity. The ReLU activation function provides the neural network with the ability of nonlinear mapping. The ReLU function is used to perform nonlinear transformation on it and pass the result to the subsequent layers.

[0055] Add a fully connected layer fullyConnectedLayer and an activation layer reluLayer to gradually reduce the dimension of the features and extract more abstract representations;

[0056] Define the output layer of the network, which is used to generate prediction results.

[0057] Add a regression layer regressionLayer to specify that the output layer of the network is suitable for regression tasks, that is, predicting continuous values;

[0058] After the framework is established, train the settings of various options.

[0059] Further, the training algorithm is:

[0060] option=trainingOptions('adam',...

[0061] 'MaxEpochs',10000,...

[0062] 'MiniBatchSize',32,...

[0063] 'InitialLearnRate',1e-3,...

[0064] 'Plots','training-progress',...

[0065] 'Verbose',false);

[0066] Set the maximum number of iterations MaxEpochs during the training process. Each Epoch means that the entire training data set is traversed once. Set the number of samples in each mini-batch. Set the initial learning rate.

[0067] The model is trained to obtain a proxy model. After the proxy model is completed, the performance of different antenna structure parameters at different frequencies is predicted.

[0068] In a preferred embodiment of the present invention, the optimization of antenna size parameters using a genetic algorithm comprises the following steps:

[0069] functionfitness=proxyModelFunction(x,y,z,c);

[0070] params = [x, y, z, c];

[0071] S11 = predict(net, params);

[0072] s24=S11(:,1);

[0073] s30=S11(:,2);

[0074] fitness=s24^2+s30^2;

[0075] end;

[0076] Define a fitness function. The S11 variable is a proxy model that accepts four structural parameters and returns a 24.2 GHZ and 30 GHZ S parameter vector group, and then splits the vector group into two frequency point scalars.

[0077] options=optimoptions('ga',...

[0078] 'MaxGenerations',100,...

[0079] 'PopulationSize',20,...

[0080] 'EliteCount',2,...

[0081] 'CrossoverFraction',0.8,...

[0082] 'PlotFcn',@gaplotbestf,...

[0083] 'Display','iter');

[0084] ga is a genetic algorithm; the MaxGenerations option sets the maximum number of generations executed by the algorithm, and each generation corresponds to a complete population evaluation; the PopulationSize option sets the number of individuals in each generation; the EliteCount option specifies the optimal number of individuals in each generation retained to the next generation; the CrossoverFraction option sets the proportion of crossover operations; use the ga function to run the genetic algorithm.

[0085] [optimalParams,optimalSParam]=ga(fitness,4,...

[0086] [], [],[],[],...

[0087] L_norm_bounds,U_norm_bounds,...

[0088] [], options);

[0089] The fitness function is passed in, the number of parameters is 4, the parameters in the empty array are set to various constraints including linear and nonlinear, L_norm_bounds and U_norm_bounds are the upper and lower bounds of the optimization parameters respectively, and finally run and wait for the optimization results.

[0090] Compared with the prior art, the present invention has the following beneficial effects:

[0091] 1. The millimeter-wave filter antenna of the present invention adopts a differential aperture coupling feeding method to achieve a relatively wide impedance matching bandwidth and a higher gain. The low-frequency and high-frequency radiation zero points are generated respectively by adding short-circuit patches and slots of radiation patches, thereby achieving out-of-band gain suppression. A planar inverted F antenna and the addition of parasitic patches are used to improve the out-of-band suppression levels of the low-frequency band and the high-frequency band respectively.

[0092] 2. The present invention adopts deep learning to establish a proxy model of selected antenna structure parameters and corresponding performance parameters, realizes rapid prediction of the performance of antennas with specific parameter structures, adopts genetic algorithms to optimize the proxy model, and rapidly designs the antenna structure corresponding to the target performance parameters. It has the advantages of high efficiency, flexibility and scalability, helps to improve design efficiency and performance, and reduce design costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 This is a simplified structural diagram of the millimeter wave filtering antenna of the present invention.

[0094] Figure 2 is the reflection coefficient of the initial antenna design structure of the present invention.

[0095] Figure 3 This is the gain diagram of the initial antenna design structure of the present invention.

[0096] Figure 4 It is the reflection coefficient of the optimized antenna structure of the present invention.

[0097] Figure 5 It is the gain diagram of the optimized antenna structure of the present invention.

[0098] Figure 6 Part of the dataset of the present invention. DETAILED DESCRIPTION

[0099] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0100] See also Figure 1The millimeter wave filtering antenna of this embodiment includes a patch, a dielectric substrate, a planar inverted F antenna metal layer 1, a metal floor 2, a rectangular feeding slot 3, a microstrip feed line 4 and a short-circuit through hole; the patch includes a radiation patch 5, a driving patch 6, a parasitic patch 7 and a short-circuit patch 8; the dielectric substrate includes three and is stacked together from bottom to top, and the three dielectric substrates are respectively a first dielectric substrate 9, a second dielectric substrate 10 and a third dielectric substrate 11; the radiation patch 5, the parasitic patch 7, and the planar inverted F antenna metal layer 1 are printed on the upper surface of the third dielectric substrate 11; the radiation patch 5 is provided with a plurality of slots; the driving patch 6 and the short-circuit patch 8 are printed on the second ... The upper surface of the dielectric substrate 10; the metal floor 2 is printed on the upper surface of the first dielectric substrate 9; the rectangular feeding slot 3 is provided with a pair and symmetrically etched on the metal floor 2; the microstrip feed line 4 is printed on the lower surface of the first dielectric substrate 9; the short-circuit through hole includes a first short-circuit through hole 12 and a second short-circuit through hole 13, the first short-circuit through hole 12 is arranged on the second dielectric substrate 10, and the first short-circuit through hole 12 is connected between the short-circuit patch 8 and the metal floor 2; the second short-circuit through hole 13 is arranged on the second dielectric substrate 10 and the third dielectric substrate 11, and the second short-circuit through hole 13 is connected between the planar inverted-F antenna metal layer 1 and the metal floor 2.

[0101] See also Figure 1 The working principle of the millimeter wave filter antenna of this embodiment is:

[0102] At the frequency point where the low-frequency radiation zero point is located, the current with high current density is mainly concentrated on the surface of the short-circuit patch 8, and the current density on the radiation patch 5 is relatively small. In addition, the current on the short-circuit patch 8 is opposite to the current on the radiation patch 5, generating radiation in the opposite direction to the radiation patch 5. Therefore, the radiation from the reverse current produces a cancellation effect, forming a radiation zero point in the low-frequency band, and the current of the planar inverted F antenna is in the same direction as the current of the radiation patch 5, which can better cancel the reverse current generated by the radiation patch 5 to the short-circuit patch 8, bringing better low-frequency band suppression effect. For the slots and parasitic patches 7 of the radiation patch 5, at the frequency point where the low-frequency radiation zero point is located, the current direction of the parasitic patch 7 is opposite to that of the radiation patch 5, and a radiation zero point is generated in the high-frequency band.

[0103] Furthermore, the dielectric substrate is made of Ferro A6M-E material; the thickness of the dielectric substrate is a multiple of 0.094 mm.

[0104] Furthermore, the thickness of the first dielectric substrate 9 is 0.564 mm, the thickness of the second dielectric substrate 10 is 0.47 mm, and the thickness of the third dielectric substrate 11 is 0.094 mm.

[0105] Furthermore, the size of the driving patch 6 is 0.14λc×0.14λc, where λc is the wavelength at the center frequency of the passband.

[0106] Furthermore, the two rectangular feeding slots 3 are completely identical, with a length of 0.18λc, where λc is the wavelength at the center frequency of the passband.

[0107] Furthermore, the size of the planar inverted-F antenna metal layer 1 is 0.05λc×0.03λc.

[0108] Furthermore, the diameters of the first short-circuit through hole 12 and the second short-circuit through hole 13 are both 0.1 mm.

[0109] Specifically, the manufacturing process of the millimeter wave filter antenna of this embodiment is as follows:

[0110] First, the basic modeling of the antenna is carried out in the three-dimensional electromagnetic field simulation tool HFSS. The basic model consists of a bottom material layer made of ferroA6M, on which is a metal floor GND made of silver (sliver). There are two microstrip feed lines under the material layer, which are symmetrical about the X-axis. Specifically, the structure of the microstrip feed line is designed on a dielectric substrate, and one end is fed through slot coupling. Compared with direct contact feeding, it has a wider bandwidth and is not easy to generate surface waves; the other end is connected to an external signal source or circuit, that is, the ports set at the left and right ends; the microstrip feed line is made of metal material silver, and the shape selected is rectangular; when current passes through the microstrip feed line, the energy is coupled and fed to the driving patch located above through the feed line, which will generate an electromagnetic field between the driving patch and the ground plate, and then radiate electromagnetic waves. This radiation is usually achieved through the gap between the patch and the surrounding ground plane, so two corresponding grooves need to be dug on the GND to complete the energy radiation. Next, add a dielectric layer between the ground and the driver chip. The material is ferroA6M. Next, add a driver patch on the dielectric layer. The material used is silver. Then lay another dielectric layer on top. The material is still ferroA6M. The last step is to add a radiation chip to radiate energy. Next is the optimization part. Connect a short-circuit patch on each side of the lower end of the driver patch. The short-circuit patch is composed of a short-circuit column and a patch on the short-circuit column. Both materials are metallic silver. Connect the short-circuit column to GND to achieve a short-circuit effect. Then slot the radiation patch. It should be emphasized that the slotting here is divided into two parts. One part is the slotting in the middle of the patch, and the other part is the slotting at the edge of the patch. There are two slots in one part, a total of four. Slots can change the internal current distribution of the antenna and introduce new resonance points, which helps to expand the working bandwidth of the antenna and introduce radiation zero points. The number of slots can be tested. A pair of slots can introduce a radiation zero point. Then, according to the smoothness of the S parameters, the number of slots is continuously optimized until it approaches the target parameters. The inverted plane F antenna is composed of a short-circuit column and a patch. The metal layer of the inverted plane F antenna is added on the third dielectric substrate. The last part of the optimization is to add parasitic patches (Parasitic Patch Antenna), that is, to add patches around the three sides of the radiation patch. The parasitic patch antenna is a design technology that uses electromagnetic coupling effects to enhance antenna performance. It adds one or more radiation patches next to the antenna radiation element (main radiation patch). These radiation patches are not directly connected to the feed line, but interact with the main radiation patch through electromagnetic coupling. Adding three parasitic patches changes the current distribution at the high-frequency gain stopband frequency, changes the radiation pattern of the high-frequency stopband, and reduces the gain of the antenna in the Z-axis direction, thereby enhancing the stopband suppression level and achieving better high-frequency filtering effects.

[0111] The intelligent auxiliary design method of the millimeter wave filter antenna of this embodiment includes the following steps:

[0112] The structural parameters that affect the radiation zero point and out-of-band suppression level of the antenna are extracted, which are the length of the short-circuit patch, the length of the planar inverted F antenna, the width of the slot for the radiation patch, and the length of the parasitic patch. The parameter scanning function in the three-dimensional electromagnetic field simulation tool HFSS is used to perform exponential sampling. This sampling helps to capture key feature changes in a wide frequency range, because physical phenomena (such as the propagation of electromagnetic waves) often follow exponential laws. The sampling ranges of the above four parameters are:

[0113] {0.05-0.15}{1-1.75}{0.1-0.5}{1-1.6}, 5 samples for each parameter, so there are 625 parameter combinations in total, and then full-wave simulation is performed in the simulation tool to obtain each group of S parameters and gains from 17GHZ to 37GHZ, and the above data are used to form a data set. The data set is imported into MATLAB, and then the code is written to build a deep learning model. The first step is to preprocess the data. In the data structure exported from the three-dimensional electromagnetic field simulation tool HFSS, the first five columns are four structural parameters and one frequency parameter, and the output is S parameters, such as Figure 6 The following is a partial explanation of the code.

[0114] x = _data(:,1:5);

[0115] y = _data(:,6);

[0116] The first five columns in the data list are set as input and placed in the x variable, and the sixth column of the data list is placed as output in the y variable.

[0117] X_norm=(X-mean(X)). / std(X);

[0118] y_norm=(y-mean(y)). / std(y);

[0119] The input and output data are normalized and put into new variables X_norm and y_norm respectively. The data is normalized so that the eigenvalues ​​have the same scale, which helps the convergence of the deep learning model.

[0120] cv=cvpartition(size(X_norm,1),'HoldOut',0.20);

[0121] idxTrain = training(cv);

[0122] idxTest = test(cv);

[0123] XTrain=X_norm(idxTrain,:);

[0124] yTrain=y_norm(idxTrain,:);

[0125] XTest = X_norm(idxTest,:);

[0126] yTest = y_norm(idxTest,:);

[0127] Use the cvpartition function to partition the dataset into training and test sets. 20% of the data is reserved as a test set, and the remaining 80% is used for training. cvpartition creates a segmentation object, and the training, validation, and test methods return the indices of the training, validation, and test sets, respectively.

[0128] After data preprocessing and segmentation, the model is built using MATLAB's deep learning toolbox.

[0129] layers=[

[0130] featureInputLayer(5,'Normalization','none','Name','input')

[0131] fullyConnectedLayer(64,'Name','fc1')

[0132] reluLayer('Name','relu1')

[0133] fullyConnectedLayer(32,'Name','fc2')

[0134] reluLayer('Name','relu2')

[0135] fullyConnectedLayer(1,'Name','output')

[0136] regressionLayer('Name','regression')];

[0137] First, define and construct the input layer featureInputLayer. Since the input data is four structural parameters and one frequency parameter, the input feature is 5. Since the data has been normalized before, Normalization here means no normalization. Finally, the layer is named input. Add a fully connected layer fullyConnectedLayer after the input layer. The fully connected layer is used to learn the complex mapping of the input data. It performs linear transformation and maps the input data to a new space. Each neuron is connected to all neurons in the previous layer. The calculation is performed through weights and biases. The number of neurons needs to be set based on experience and the prediction effect of the model. The number of neurons in this layer is set to 64. Then add an activation layer reluLayer. Set the activation function to introduce nonlinearity so that the model can learn complex patterns. The ReLU activation function provides the neural network with the ability of nonlinear mapping, enabling the neural network to learn and simulate complex functional relationships. This layer will be applied to the output of the previous layer, and it will be nonlinearly transformed through the ReLU function, and then the result will be passed to the subsequent layers. Add a fully connected layer fullyConnectedLayer and an activation layer reluLayer again to gradually reduce the dimension of the feature and extract more abstract representations. Next, define the output layer of the network to generate prediction results. Finally, add the regression layer regressionLayer to specify that the output layer of the network is suitable for regression tasks, that is, predicting continuous values. After the framework is established, set various training options, such as optimizer, number of iterations, learning rate, etc.

[0138] option=trainingOptions('adam',...

[0139] 'MaxEpochs',10000,...

[0140] 'MiniBatchSize',32,...

[0141] 'InitialLearnRate',1e-3,...

[0142] 'Plots','training-progress',...

[0143] 'Verbose',false);

[0144] 'adam' is the specified optimizer, which is an adaptive learning rate optimization algorithm that combines the advantages of AdaGrad and RMSProp, and generally provides good performance and fast convergence speed. Set the maximum number of iterations (MaxEpochs) during training. Each Epoch means that the entire training data set is traversed once. Here it is set to 100, which means that the model will be trained at most 100 times. Set the number of samples in each mini-batch. Here it is set to 32, which means that 32 samples will be used each time the model parameters are updated. The mini-batch size affects the training speed and memory usage of the model, as well as the generalization ability of the model. Set the initial learning rate. The learning rate determines the step size of the parameter update during the optimization process. Here it is set to 0.001, which is a commonly used initial learning rate that helps the model converge quickly in the early stage of training. Then start training the model to obtain the proxy model. After the proxy model is completed, the performance of different antenna structure parameters at different frequencies can be quickly predicted.

[0145] After the model is established, the next step is to optimize its target. Since the goal is to find structural parameters with a bandwidth of less than -10dB from 24.2GHz to 30GHz, and in order to improve accuracy and training speed, the proxy model needs to be adjusted. The input of the model is changed to four structural parameters, and the output is S parameters at 24.2GHZ and 30GHZ frequency points. The data division and the settings of some neural network layers are adjusted, and the input features of the input layer featureInputLayer are changed to 4, and the output features are changed to 2. Due to the reduction of frequency input features, the data is greatly reduced. In order to make the model more accurate, it is necessary to adjust the maximum number of iterations in the training options. After testing multiple different numbers of iterations, it is more appropriate to set it to 5000. Too many iterations will lead to overfitting of the model, learning rate decay, and decreased generalization ability. After adjusting the code, retrain the model.

[0146] Using the established proxy model, the genetic algorithm is used to optimize the size parameters of the antenna to obtain the optimal predicted value and its corresponding antenna parameter combination. In the genetic algorithm, the first thing to determine is the fitness function, which plays a core role in the algorithm and defines the goal of the problem. In optimization problems, the fitness function is usually designed to minimize or maximize a specific objective function. The higher the fitness value of an individual, the closer it is to the optimal solution of the problem.

[0147] functionfitness=proxyModelFunction(x,y,z,c)

[0148] params = [x, y, z, c];

[0149] S11 = predict(net, params);

[0150] s24=S11(:,1);

[0151] s30=S11(:,2);

[0152] fitness=s24^2+s30^2;

[0153] end;

[0154] First, define a fitness function. The S11 variable is a proxy model that accepts four structural parameters and returns an S parameter vector group of 24.2GHZ and 30GHZ. The vector group is then split into scalars of two frequency points. The genetic algorithm is a minimization problem, so we hope that the smaller the fitness value, the better.

[0155] Since our goal is to make the S parameters of both frequency bands zero, we can calculate the sum of the squares of the S parameters as the fitness function. The sum of squares is used here because it can consider the S parameters of both frequency bands at the same time, and the square can amplify the error, making the optimization process more sensitive. After the fitness function is written, it is time to configure the genetic algorithm options.

[0156] options=optimoptions('ga',...

[0157] 'MaxGenerations',100,...

[0158] 'PopulationSize',20,...

[0159] 'EliteCount',2,...

[0160] 'CrossoverFraction',0.8,...

[0161] 'PlotFcn',@gaplotbestf,...

[0162] 'Display','iter');

[0163] The first parameter 'ga' specifies the type of optimization algorithm, which is a genetic algorithm. The MaxGenerations option sets the maximum number of generations performed by the algorithm. Each generation corresponds to a complete population evaluation. The PopulationSize option sets the number of individuals (candidate solutions) in each generation. A larger population can provide more diversity, but it increases the computational cost. The EliteCount option specifies the number of optimal individuals in each generation that are retained to the next generation. These individuals have the highest fitness in the current population. The CrossoverFraction option sets the proportion of crossover operations, that is, what proportion of individuals in the new generation are generated by crossover operations. Then the genetic algorithm is run using the ga function.

[0164] [optimalParams,optimalSParam]=ga(fitness,4,...

[0165] [],[],[],[],...

[0166] L_norm_bounds,U_norm_bounds,...

[0167] [],options);

[0168] The fitness function is passed in, the number of parameters is 4, the parameters in the empty array are set to various constraints including linear and nonlinear, etc. The optimization target has no constraints so it is set to an empty array, L_norm_bounds and U_norm_bounds are the upper and lower bounds of the optimization parameters. Finally, run and wait for the optimization results.

[0169] Furthermore, the reflection coefficient of the initial antenna design structure is as follows Figure 2 As shown in the figure, the impedance matching bandwidth with S11 response less than -10dB is 23.060GHz-29.710GHz, and the frequency range basically covers the N257, N258, and N261 bands in the 5G millimeter wave communication band, achieving good impedance matching. Figure 3 As shown in the figure, the 5dB bandwidth covers 23.607GHz-29.086GHz, the maximum gain in the passband is 5.5dBi, there is a radiation zero at 19.3GHz and 34.3GHz respectively, and a good out-of-band gain suppression level of more than 12.5dB is achieved in the frequency range of 17-37GHz. The overall size of the antenna is 0.75λc×0.54λc×0.1λc, achieving a compact size.

[0170] The proxy model established by deep learning can directly input the structural parameters and predict the corresponding performance, and then use the structural parameters to perform full-wave simulation. Then, the genetic algorithm is used for optimization to find the structural parameters with a bandwidth of less than -10dB from 24.2GHz to 30GHz. The optimized antenna reflection coefficient is as follows Figure 4 The impedance matching bandwidth with S11 response less than -10dB is 24.00-31.34, and the frequency range is wider than the initial result bandwidth and close to the target bandwidth. Figure 5 As shown, the 5dB bandwidth covers 24.519GHz-31.150GHz. The frequency range is wider than the initial result bandwidth and closer to the target bandwidth, and the out-of-band suppression at high frequencies is more obvious.

[0171] The above is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited to the above content. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A millimeter wave filtering antenna, characterized in that: It includes a patch, a dielectric substrate, a planar inverted F antenna metal layer, a metal floor, a rectangular feed slot, a microstrip feed line and a short-circuit through hole; The patches include radiation patches, driving patches, parasitic patches and short-circuit patches; The dielectric substrates include three dielectric substrates stacked in sequence from bottom to top, and the three dielectric substrates are respectively a first dielectric substrate, a second dielectric substrate and a third dielectric substrate; The radiation patch, the parasitic patch, and the planar inverted-F antenna metal layer are printed on the upper surface of the third dielectric substrate; The radiation patch is provided with a plurality of slots; the driving patch and the short-circuit patch are printed on the upper surface of the second dielectric substrate; The metal floor is printed on the upper surface of the first dielectric substrate; a pair of rectangular feed slots are provided and symmetrically etched on the metal floor; the microstrip feed line is printed on the lower surface of the first dielectric substrate; The short-circuit via comprises a first short-circuit via and a second short-circuit via. The first short-circuit via is arranged through the second dielectric substrate, and the first short-circuit via is connected between the short-circuit patch and the metal floor. The second short-circuit via is arranged through the second dielectric substrate and the third dielectric substrate, and the second short-circuit via is connected between the metal layer of the planar inverted-F antenna and the metal floor.

2. The millimeter wave filtering antenna according to claim 1, characterized in that: The dielectric substrate is made of FerroA6M-E material; the thickness of the dielectric substrate is a multiple of 0.094 mm; the thickness of the first dielectric substrate is 0.564 mm, the thickness of the second dielectric substrate is 0.47 mm, and the thickness of the third dielectric substrate is 0.094 mm.

3. The millimeter wave filtering antenna according to claim 1, characterized in that: The size of the driving patch is 0.14λc×0.14λc, and the size of the metal layer of the planar inverted-F antenna is 0.05λc×0.03λc, where λc is the wavelength at the center frequency of the passband.

4. The millimeter wave filtering antenna according to claim 1, characterized in that: The two rectangular feed slots are exactly the same, with a length of 0.18λc, where λc is the wavelength at the center frequency of the passband.

5. An intelligent auxiliary design method applied to the millimeter wave filter antenna according to claims 1-4, characterized in that: The following steps are involved: The structural parameters that affect the radiation zero point and out-of-band suppression level of the antenna are extracted, which are the length of the short-circuit patch, the length of the metal layer of the planar inverted-F antenna, the width of the slot of the radiation patch, and the length of the parasitic patch. The above four parameters are sampled exponentially; The sampled parameters are combined in the simulation tool for full-wave simulation, and the S parameters and gains of each group from 17GHZ to 37GHZ are obtained by simulation. The above data are used to form a data set; the data set is imported into MATLAB to build a deep learning model; After the model is established, it is optimized for the target; By using the established proxy model, the genetic algorithm is used to optimize the antenna size parameters to obtain the optimal prediction value and its corresponding antenna parameter combination.

6. The intelligent assisted design method according to claim 5, characterized in that: The sampling ranges of the four parameters are: {0.05-0.15}, {1-1.75}, {0.1-0.5}, {1-1.6}, 5 samples for each parameter, a total of 625 parameter combinations.

7. The intelligent assisted design method according to claim 5, characterized in that: Building a deep learning model involves the following steps: For data preprocessing, the first five columns of the data exported from the three-dimensional electromagnetic field simulation tool HFSS are four structural parameters and one frequency parameter, which are output as S parameters; x = _data(:,1:5); y = _data(:,6); Set the first five columns in the data list as input and put them into the above x variable, and put the sixth column in the data list as output into the above y variable; X_norm=(X-mean(X)). / std(X); y_norm=(y-mean(y)). / std(y); Normalize the input and output data and put them into new variables X_norm and y_norm respectively. Normalize the data so that the eigenvalues ​​have the same scale. cv=cvpartition(size(X_norm,1),'HoldOut',0.20); idxTrain = training(cv); idxTest = test(cv); XTrain=X_norm(idxTrain,:); yTrain=y_norm(idxTrain,:); XTest = X_norm(idxTest,:); yTest = y_norm(idxTest,:); Use the cvpartition function to divide the dataset into training and test sets. 20% of the data is reserved as the test set and the remaining 80% is used for training.

8. The intelligent assisted design method according to claim 7, characterized in that: After data preprocessing and division, the model is constructed using MATLAB: layers=[ featureInputLayer(5,'Normalization','none','Name','input') fullyConnectedLayer(64,'Name','fc1') reluLayer('Name','relu1') fullyConnectedLayer(32,'Name','fc2') reluLayer('Name','relu2') fullyConnectedLayer(1,'Name','output') regressionLayer('Name','regression')]; First, define and construct the input layer featureInputLayer. The input data is four structure parameters and one frequency parameter. The input feature is 5. The layer is named input. Add a fully connected layer after the input layer. The fully connected layer is used to learn the complex mapping of input data, perform linear transformation, and map the input data to a new space. Each neuron is connected to all neurons in the previous layer and is calculated through weights and biases. The number of neurons needs to be set based on experience and the prediction effect of the model. The number of neurons in this layer is set to 64. Add an activation layer reluLayer and set the activation function to introduce nonlinearity. The ReLU activation function provides the neural network with the ability of nonlinear mapping. The ReLU function is used to perform nonlinear transformation on it and pass the result to the subsequent layers. Add a fully connected layer fullyConnectedLayer and an activation layer reluLayer to gradually reduce the dimension of the features and extract more abstract representations; Define the output layer of the network to generate prediction results; Add a regression layer regressionLayer to specify that the output layer of the network is suitable for regression tasks, that is, predicting continuous values; After the framework is established, train the settings of various options.

9. The intelligent assisted design method according to claim 8, characterized in that: The training algorithm is: option=trainingOptions('adam',... 'MaxEpochs',10000,... 'MiniBatchSize',32,... 'InitialLearnRate',1e-3,... 'Plots','training-progress',... 'Verbose',false); Set the maximum number of iterations MaxEpochs during the training process. Each Epoch means that the entire training data set is traversed once. Set the number of samples in each mini-batch. Set the initial learning rate. The model is trained to obtain a proxy model. After the proxy model is completed, the performance of different antenna structure parameters at different frequencies is predicted.

10. The intelligent assisted design method according to claim 1, characterized in that: The optimization of antenna size parameters using genetic algorithm includes the following steps: functionfitness=proxyModelFunction(x,y,z,c); params = [x, y, z, c]; S11 = predict(net, params); s24=S11(:,1); s30=S11(:,2); fitness=s24^2+s30^2; end; Define a fitness function. The S11 variable is a proxy model that accepts four structural parameters and returns a 24.2 GHZ and 30 GHZ S parameter vector group, and then splits the vector group into two frequency point scalars. options=optimoptions('ga',... 'MaxGenerations',100,... 'PopulationSize',20,... 'EliteCount',2,... 'CrossoverFraction',0.8,... 'PlotFcn',@gaplotbestf,... 'Display','iter'); ga is a genetic algorithm; the MaxGenerations option sets the maximum number of generations executed by the algorithm, and each generation corresponds to a complete population evaluation; the PopulationSize option sets the number of individuals in each generation; the EliteCount option specifies the optimal number of individuals in each generation to be retained to the next generation; the CrossoverFraction option sets the proportion of crossover operations; use the ga function to run the genetic algorithm; [optimalParams,optimalSParam]=ga(fitness,4,... [],[],[],[],... L_norm_bounds,U_norm_bounds,... [],options); The fitness function is passed in, the number of parameters is 4, the parameters in the empty array are set to various constraints including linear and nonlinear, L_norm_bounds and U_norm_bounds are the upper and lower bounds of the optimization parameters respectively, and finally run and wait for the optimization results.

Images