Machine learning assisted antenna parameter and topological structure hybrid optimization method

Abstract

The invention belongs to the technical field of antenna optimization, and discloses a machine learning assisted antenna parameter and topological structure hybrid optimization method, which comprises the following steps of: 1, carrying out binary coding on an antenna parameter to be optimized, and combining the coded antenna parameter with a binary vector corresponding to a topological structure to obtain a binary vector; combining into a new mixed binary vector; step 2, generating a data set based on the mixed binary vector set in the step 1; 3, introducing a principal component analysis simplified model training process, and obtaining an optimized mixed binary vector in combination with an optimization algorithm; 4, electromagnetic simulation verification is carried out; 5, judging whether the optimization process meets a termination condition or not, and if yes, outputting an optimization result at the moment; and if not, updating the data set, and returning to the step 3. According to the method, the antenna parameters and the antenna topological structure can be optimized at the same time, machine learning is still adopted for auxiliary optimization at the same time, the optimization efficiency is improved to a certain degree, and a better optimization result is obtained.

Description

Technical Field

[0001] This application belongs to the technical field of antenna optimization, and specifically relates to a method for hybrid optimization of antenna parameters and topology structure assisted by machine learning. Background Art

[0002] As a carrier for transmitting and receiving electromagnetic waves, the optimal design of antennas has always been a core issue in antenna research. Traditional antennas often have relatively regular shapes designed and corresponding parameters optimized. With the popularization of computer technology, the automated design of antenna topology has also become a new popular direction in antenna optimization design. Usually, different functions of antennas can be achieved by the flexible arrangement of the topology. Currently, the optimal design of antennas mainly focuses on the above two types of directions. One is the optimal design of antenna parameters, which usually takes values within a certain range and is often a continuous variable. The other is the optimal design of antenna topology, which usually uses 0 or 1 to represent the presence or absence of metal units in the topological position of the antenna and is often a discrete variable. As the difficulty of antenna design increases, manual optimization design becomes more and more complex. Machine learning, as a tool with high prediction accuracy and fast prediction speed, has been widely used in electronic design fields such as antennas and passive devices and achieved good results.

[0003] Currently, a relatively good optimization method is to use machine learning with online update to assist in the optimal design of antennas. For the first type of optimal design of antenna parameters, technologies such as Gaussian process regression and K-nearest neighbor have been used for auxiliary optimization. For the second type of optimal design of antenna topology, convolutional neural networks have also been used for auxiliary optimization. However, currently, the machine learning-assisted antenna optimization design often only selects one of these types of directions, that is, either optimizing antenna parameters or optimizing antenna topology. At this time, the optimization space is small, and the ideal optimization result may not necessarily be obtained, which also poses higher requirements for engineers. Summary of the Invention

[0004] To solve the above technical problems, this application provides a method for hybrid optimization of antenna parameters and topology structure assisted by machine learning. Compared with only optimizing one of these types of variables, this application significantly expands the space of optimization variables and has a certain improvement in the optimization effect.

[0005] To achieve the above object, this application is implemented through the following technical solutions:

[0006] This application is a method for hybrid optimization of antenna parameters and topology structure assisted by machine learning, specifically including the following steps:

[0007] Step 1: Construct a hybrid binary variable: Binary-encode the antenna parameters to be optimized, and combine the binary vector corresponding to the binary-encoded antenna parameters with the binary vector corresponding to the antenna topology to form a new hybrid binary vector. The new hybrid binary vector includes information on the antenna parameters and the antenna topology;

[0008] Step 2: Construct a dataset: Based on the hybrid binary vector set in Step 1, randomly generate an initial hybrid binary vector consisting of element 0 or element 1 through random sampling, and use simulation software such as HFSS to obtain the electromagnetic response of the performance indicators corresponding to the antenna parameters and topology constituted by the initial hybrid binary vector, thereby constructing a dataset for the training of a multi-layer perceptron (MLP) surrogate model;

[0009] Step 3: Optimize the antenna parameters and topology based on the multi-layer perceptron surrogate model: Introduce principal component analysis (PCA) to simplify the training process of the multi-layer perceptron (MLP) surrogate model, and at the same time obtain an optimized hybrid binary vector based on the multi-layer perceptron (MLP) surrogate model and the binary particle swarm optimization algorithm (BPSO);

[0010] Step 4: Electromagnetic simulation verification: Convert the optimized hybrid binary vector obtained in Step 3 into antenna parameters and topology respectively, and use simulation software such as HFSS for calculation to obtain the verified true performance indicators of the antenna and record them as the target optimal solution;

[0011] Step 5: Determine whether the termination condition is satisfied: Determine whether the optimization process satisfies the optimization termination condition. If not, add the verification result to the dataset, update the dataset, and then return to Step 3. If satisfied, output the optimization result at this time.

[0012] A further improvement of this application is that: Step 1 specifically includes the following steps:

[0013] Step 1.1: Antenna parameter encoding: Take equally spaced values for the antenna parameters to be optimized and represent them with a binary vector of length p, where each group of binary vectors corresponds to a value of the antenna parameter;

[0014] Step 1.2, Topology Encoding: For the antenna topology, determine the antenna topology region to be optimized, and divide the antenna topology region to be optimized into p×q square regions. Each square region corresponds to an element in an m×n - dimensional binary matrix. Map the element 0 in the binary matrix to air and 1 to metal, and transform the obtained binary matrix into a binary vector. In this way, the antenna topology can be represented by constructing a binary vector with a length of r. Among them, the total number of square regions p×q is equal to the length r of the transformed binary vector, and the total dimension m×n of the binary matrix is equal to the length r of the transformed binary vector;

[0015] Step 1.3, Variable Combination: Combine the binary vector with a length of r corresponding to the antenna topology and the binary vector with a length of p corresponding to the antenna parameters to finally obtain a new hybrid binary vector with a length of p + r. This vector contains the relevant information of the antenna parameters and the antenna topology;

[0016] A further improvement of this application is that: The specific steps of constructing the initial data set in Step 2 are as follows:

[0017] Step 2.1, Based on the set new hybrid binary vector, through random sampling, obtain k hybrid binary vectors with a length of p + r, defined as the input parameter X of the data set;

[0018] Step 2.2, Convert the initial hybrid binary vector into specific antenna parameters and topologies respectively, and use simulation software such as HFSS for calculation to obtain the electromagnetic response of the performance indicators corresponding to the antenna parameters and topologies, defined as the output parameter Y of the data set;

[0019] Step 2.3, The input parameter X and the output parameter Y construct a data set for subsequent model training.

[0020] A further improvement of this application is that: Step 3 specifically includes the following steps:

[0021] Step 3.1, Introduce the principal component analysis method to reduce the dimension of the electromagnetic response of the performance indicators corresponding to the antenna parameters and topologies, thereby simplifying the complexity of the multi - layer perceptron (MLP) surrogate model;

[0022] Step 3.2, Then use the multi - layer perceptron (MLP) surrogate model for training to establish a non - linear relationship between the hybrid binary variables and the electromagnetic response of the performance indicators corresponding to the antenna parameters and topologies after dimension reduction in Step 3.1;

[0023] Step 3.3: Combine the multi-layer perceptron (MLP) surrogate model established in Step 3.2 with the binary particle swarm optimization algorithm (BPSO), and optimize according to the set fitness function to finally obtain the optimized hybrid binary vector.

[0024] A further improvement of this application is that in Step 5, the termination condition can be set to reach the maximum iteration limit or the final optimization result of the antenna has met the set optimization goal; if the termination condition is met, the optimization process ends, and if not, the hybrid binary variables obtained from this optimization iteration and their corresponding electromagnetic responses of the true antenna performance indicators are added to the dataset, and the dataset is updated, and then return to Step 3.

[0025] The beneficial effects of this application are:

[0026] This application can optimize both antenna parameters and antenna topology, further increasing the optimization space of variables, thus having the opportunity to obtain better optimization results.

[0027] This application still uses machine learning for auxiliary optimization at the same time. Compared with ordinary evolutionary algorithms, the optimization quality has been improved to a certain extent, and the introduction of principal component analysis technology can further improve the optimization speed.

[0028] This application is convenient in design and simple in structure, and is expected to be extended to the optimization design of other electromagnetic components. Description of the Drawings

[0029] Figure 1 is the basic flowchart of this application.

[0030] Figure 2 is the initial structure diagram of the pixel monopole antenna used in this application.

[0031] Figure 3 is the comparison chart of the optimization mean curves between the optimization method of this application and the optimization method using only BPSO.

[0032] Figure 4 is the optimization curve diagram of a certain result in the optimization method of this application.

[0033] Figure 5 is the topological structure diagram of the pixel monopole antenna after optimization in this application.

[0034] Figure 6 is the S-parameter diagram of the pixel monopole antenna after optimization in this application. Detailed Implementation Manner

[0035] The embodiments of the present application will be disclosed below with reference to the drawings. For the sake of clarity, many practical details will be described together in the following narrative. However, it should be understood that these practical details are not used to limit the present application. That is to say, in some embodiments of the present application, these practical details are not necessary. In addition, for the sake of simplifying the drawings, some conventional structures and components will be shown in a simple schematic manner in the drawings.

[0036] As Figure 1 shown, the present application is a method for hybrid optimization of antenna parameters and topological structures assisted by machine learning. First, according to the antenna task to be optimized, the hyperparameters of the discrete binary particle swarm optimization algorithm (BPSO) and the multi-layer perceptron (MLP) used are set. Then, a specific operation process for hybrid optimization is designed, which mainly includes the following steps:

[0037] Step 1: Construct a hybrid binary variable: Binary encode the antenna parameters to be optimized, and combine the binary vector corresponding to the binary-encoded antenna parameters with the binary vector corresponding to the antenna topology to form a new hybrid binary vector. The new hybrid binary vector includes information on antenna parameters and antenna topology.

[0038] Step 2: Construct a data set: In the machine learning-assisted optimization module, based on the hybrid binary vector set in Step 1, randomly generate an initial hybrid binary vector composed of element 0 or element 1 by random sampling, and use simulation software such as HFSS to obtain the electromagnetic response of the performance indicators corresponding to the antenna parameters and topology composed of the initial hybrid binary vector, thereby constructing a data set for training the multi-layer perceptron (MLP) model;

[0039] Step 3: Optimize antenna parameters and topology based on the multi-layer perceptron surrogate model: Introduce principal component analysis (PCA) to simplify the training process of the multi-layer perceptron (MLP) model, and at the same time obtain an optimized hybrid binary vector based on the multi-layer perceptron (MLP) surrogate model and the binary particle swarm optimization algorithm (BPSO);

[0040] Step 4: Electromagnetic simulation verification: Convert the optimized hybrid binary vector obtained in Step 3 into antenna parameters and topology respectively, and use simulation software such as HFSS for calculation to obtain the verified true performance indicators of the antenna, and record them as the target optimal solution;

[0041] Step 5: Determine whether the termination condition is met: Determine whether the optimization process meets the optimization termination condition. If not, add the verification result to the data set, update the data set, and then return to Step 3. If so, output the optimization result at this time.

[0042] This embodiment optimizes a pixel monopole antenna operating in the Wi-Fi band. First, the relevant hyperparameters of BPSO and MLP are determined. The initial structure of the MLP used in this embodiment has two hidden layers, with the corresponding number of neurons being 48 and 32 respectively, and the rest of the initial parameters can use common default parameters. Then, the proposed optimization method is used to design the antenna. The specific steps are as follows:

[0043] Step 1 requires determining the corresponding mixed binary variables.

[0044] First, the initial structure diagram of the monopole antenna before optimization is given, as Figure 2 shown, where the dielectric constant of the dielectric plate is 2.65, the height h = 0.8 mm, the width w1 = 24 mm, the length l1 = 40 mm, the length of the metal floor l2 = 10 mm, the distance from the port to the area to be optimized l3 = 12 mm, and the width of the microstrip feeder w f = 2 mm.

[0045] Then, the side length of the square pixel unit is selected as the antenna parameter to be optimized. The value range of the side length of the square unit is set to 2 - 3.5 mm, and then it is equally divided at intervals of 0.1 mm, resulting in a total of 16 values, which are successively represented using four different binary encodings. In addition, to avoid the situation of corner-to-corner connection between units, the square unit is extended by 0.2 mm along the feeder direction.

[0046] Secondly, the area to be optimized is set as Figure 2 the dashed part in, and its structure is symmetric about the microstrip feeder. Then, this area is divided into 6×8 square units, where it is equally divided into six parts perpendicular to the microstrip feeder direction and eight parts parallel to the microstrip feeder direction, and then randomly filled with metal or air in sequence from left to right and from bottom to top.

[0047] Finally, the four-bit binary encoding corresponding to the antenna parameters is combined with the 48-bit binary encoding corresponding to the topology structure, and finally 52 binary variables are obtained.

[0048] The design goal of the antenna in the embodiment of this application is set to minimize the value in the frequency bands of 2.4 - 2.5 GHz and 5.15 - 5.85 GHz. The specific objective function is:

[0049]

[0050] where f is the frequency band to be optimized, is the value of the S parameter of the antenna predicted by the MLP at all frequency points under the mixed binary variable X of the given antenna parameters and topology structure.

[0051] In step 2, a dataset needs to be constructed. In this example, the initial samples of the dataset can be obtained by random sampling, and the number of initial samples of the dataset is set to 50.

[0052] Since step 3 needs to be called repeatedly, the multi-layer perceptron (MLP) model needs to be trained multiple times during the optimization process, which significantly occupies a certain amount of optimization time. Therefore, the principal component analysis method is introduced to reduce the dimensionality of the electromagnetic response of the performance indicators corresponding to the antenna parameters and topological structure, thereby simplifying the complexity of the multi-layer perceptron (MLP) model. Taking 500 samples as an example in this instance, the original training time of the MLP was 94s. After introducing PCA, the training time decreased to 66s. On the premise that the information of the antenna performance indicators is basically complete, the training efficiency of the model is significantly improved.

[0053] In step 5, the iteration convergence condition can be set as the total number of electromagnetic simulations. In this example, the number of electromagnetic simulations is set to 1000 times.

[0054] Analyze the final optimization results. Figure 3 The optimization mean curves of the optimization method of this application and the method only using BPSO for optimization under different initial samples are given, where 5 groups of different initial samples are obtained for each optimization method. From Figure 3 it can be seen that after 1000 electromagnetic simulations, the optimization mean result of the optimization method of this application is -14.2555dB, and the optimization mean result of the BPSO method is -12.5653dB, indicating that the introduction of machine learning in this application can improve the quality of the optimization results, thereby obtaining better optimization results. The specific optimization results of the two optimization methods are shown in Table 1.

[0055] Table 1 Final optimization results of two optimization methods (dB)

[0056]

[0057] Figure 4 and Figure 5 respectively give the optimization process curve corresponding to the first optimization result under the optimization of this application and the pixel monopole antenna structure diagram. The final optimization result after 1000 simulations is approximately -14.62dB, and the 48 binary variables corresponding to the antenna topological structure are as Figure 5 shown. In addition, the binary encoding corresponding to the side length of the square unit is [1, 1, 1, 0], indicating that the side length of the square unit is 3.4mm at this time, and the side length along the feeder direction is 3.6mm.

[0058] Figure 6The S-parameter indicators of the optimized antenna are shown. It can be found that the -10 dB bandwidth is in two frequency bands of 2.29 - 2.87 GHz and 4.85 - 6.62 GHz, and among them, within the entire Wi-Fi frequency band the values are all less than -14.62 dB, indicating that the optimized antenna can work well in the Wi-Fi frequency band.

[0059] The above are only the embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various changes and modifications can be made to the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A method for hybrid optimization of antenna parameters and topology assisted by machine learning, characterized in that: The method for hybrid optimization of antenna parameters and topology specifically includes the following steps: Step 1, construct a hybrid binary variable: perform binary encoding on the antenna parameters to be optimized, and combine the binary vector corresponding to the binary-encoded antenna parameters with the binary vector corresponding to the antenna topology to form a new hybrid binary vector, where the new hybrid binary vector includes information on antenna parameters and antenna topology; Step 2, construct a data set: based on the new hybrid binary vector set in Step 1, randomly generate initial hybrid binary vectors through random sampling, and obtain the electromagnetic response of the performance indicators corresponding to the antenna parameters and topology composed of the initial hybrid binary vectors, thereby constructing a data set; Step 3, optimize antenna parameters and topology based on a multi-layer perceptron surrogate model: introduce principal component analysis to simplify the training process of the multi-layer perceptron surrogate model, and at the same time obtain the optimized hybrid binary vector based on the multi-layer perceptron surrogate model and binary optimization algorithm; Step 4, electromagnetic simulation verification: convert the optimized hybrid binary vector obtained in Step 3 into antenna parameters and topology respectively, and use simulation software for calculation to obtain the verified true performance indicators of the antenna, and record them as the target optimal solution; Step 5, determine whether the termination condition is satisfied: determine whether the optimization process satisfies the optimization termination condition. If not, add the verification result to the data set, update the data set, and then return to Step 3. If satisfied, output the optimization result at this time.

2. The method for hybrid optimization of antenna parameters and topological structure assisted by machine learning according to claim 1, characterized in that: The specific steps of Step 1 are as follows: Step 1.1, antenna parameter encoding: perform equally spaced value taking on the antenna parameters to be optimized, and represent them with binary vectors of length p, where each group of binary vectors corresponds to a value of the antenna parameters; Step 1.2, topology encoding: for the antenna topology, determine the antenna topology area to be optimized, divide the antenna topology area to be optimized into p×q square areas, each square area corresponds to an element in an m×n-dimensional binary matrix, map the element 0 in the binary matrix to air, 1 to metal, and transform the obtained binary matrix into a binary vector, and represent the antenna topology by constructing a binary vector of length r, where the total number of square areas p×q is equal to the length r of the transformed binary vector, and the total dimension m×n of the binary matrix is equal to the length r of the transformed binary vector; Step 1.3, variable combination: combine the binary vector of length r corresponding to the antenna topology with the binary vector of length p corresponding to the antenna parameters to finally obtain a new hybrid binary vector of length p+r.

3. A method for hybrid optimization of antenna parameters and topology assisted by machine learning according to claim 1, characterized in that: The specific steps for constructing the initial data set in Step 2 are as follows: Step 2.1, based on the set new hybrid binary vector, obtain k initial hybrid binary vectors of length p+r through random sampling, defined as the input parameter X of the data set; Step 2.2: Convert the initial mixed binary vectors into specific antenna parameters and topologies respectively, and use simulation software for calculation to obtain the electromagnetic response of the performance indicators corresponding to the antenna parameters and topologies, which is defined as the output parameter Y of the data set; Step 2.3: The input parameter X and the output parameter Y construct the data set.

4. A method for hybrid optimization of antenna parameters and topology assisted by machine learning according to claim 3, characterized in that: The said Step 3 specifically includes the following steps: Step 3.1: Introduce the principal component analysis method to reduce the dimension of the electromagnetic response of the performance indicators corresponding to the antenna parameters and topologies, and simplify the complexity of the multi-layer perceptron surrogate model; Step 3.2: Then use the multi-layer perceptron surrogate model for training to establish a non-linear relationship between the mixed binary variables and the electromagnetic response of the performance indicators corresponding to the antenna parameters and topologies after dimension reduction in Step 3.1; Step 3.3: Combine the multi-layer perceptron surrogate model established in Step 3.2 with the binary optimization algorithm to obtain the optimized mixed binary vectors.

5. A method for hybrid optimization of antenna parameters and topology assisted by machine learning according to claim 1, characterized in that: In Step 5, the termination condition is to reach the maximum iteration limit or the final optimization result of the antenna has met the set optimization goal; if the termination condition is met, the optimization process ends, if not, add the mixed binary variables obtained in this optimization iteration and their corresponding electromagnetic responses of the true antenna performance indicators to the data set, update the data set, and then return to Step 3.

Images