Parallel antenna topology optimization design method and system adopting multiple transfer functions

Abstract

The invention discloses a parallel antenna topology optimization design method and system adopting multiple transfer functions. According to the method, on the basis of a traditional machine learning assisted antenna topological pixel optimization design algorithm architecture, in the optimization stage of each iteration, a parallel BPSO algorithm based on multiple transfer functions is adopted, a plurality of mutually independent BPSO algorithms are called in parallel, and on the basis of the same machine learning agent model, the same objective function is optimized; different algorithms are distinguished as different transfer functions and whether a known optimal solution is added into an initial population of the algorithm. According to the method, a plurality of high-quality candidate solutions for parallel simulation verification can be obtained in each iteration, and compared with a traditional machine learning assisted optimization antenna topological pixel optimization design method, the time required for optimization is greatly shortened, and the optimization result is greatly improved within the same time. The method can be applied to the field of topological optimization of different types of antennas.

Description

Technical Field

[0001] The present invention belongs to the technical field of antenna design, and relates to a parallel antenna topology optimization design method and system using multiple transfer functions. Background Art

[0002] As a research issue that has been the focus of specialties such as electromagnetic fields and microwave technology, communication and information systems, the topological optimization design of antennas has always been a hot and difficult point in academic research. Traditional antenna topological optimization methods can be divided into antenna structure topological optimization design based on evolutionary algorithms and antenna structure topological optimization design based on gradient algorithms. The former is easy to integrate with commercial software, has strong scalability, does not require sensitivity information, and can search for the globally optimal topological structure as much as possible, but it takes a long time and has high requirements for computing resources; the latter can effectively improve the solution efficiency of the antenna topological structure optimization problem, but it is easy to fall into local optimality, and it is difficult to integrate with commercial software because it requires sensitivity information.

[0003] In the past decade or so, machine learning methods have been widely introduced into the design fields of electronic devices such as antennas, passive devices, and circuits, and have achieved good results. Currently, the vast majority of machine learning-assisted antenna designs only consider the parameter design of antennas after fixing the antenna topology, and cannot be applied to the topological optimization design of antennas. In recent years, machine learning-assisted antenna topological optimization design methods have been used to achieve a faster convergence speed compared to traditional evolutionary algorithm-based antenna topological optimization methods, but it has an obvious defect: it belongs to a serial algorithm, that is, only one candidate solution is obtained for each iteration for full-wave simulation software to verify. With the development of computer technology, personal computers used for antenna design usually can provide 3-10 parallel processor cores, and the traditional machine learning-assisted antenna topological optimization design method is difficult to make full use of the advantages of multi-core resources, so it is difficult to achieve a further breakthrough in the convergence speed. Summary of the Invention

[0004] Object of the Invention: Aiming at the above problems, the present invention proposes a parallel antenna topology optimization design method and system using multiple transfer functions, which achieves a significant acceleration of convergence compared to the traditional machine learning-assisted antenna topological optimization design method.

[0005] Technical Solution: To achieve the above object of the invention, the present invention adopts the following technical solutions:

[0006] In a first aspect, the present invention provides a parallel antenna topology optimization design method using multiple transfer functions. Based on an iterative machine learning-assisted antenna topology pixel optimization design algorithm architecture, in the optimization stage of each iteration, a parallel binary particle swarm BPSO algorithm based on multiple transfer functions is adopted. Multiple mutually independent BPSO algorithms are called in parallel. Based on the same machine learning surrogate model, the same objective function is optimized. The difference between different algorithms lies in the different transfer functions used and whether the known optimal solution is added to the initial population of the algorithm.

[0007] Further, the iterative machine learning-assisted antenna topology pixel optimization design algorithm architecture includes: initial sampling; full-wave simulation to construct a data set; convolutional neural network training; binary particle swarm algorithm optimization; parallel full-wave simulation verification; determining whether the termination condition is met. If the termination condition is met, the algorithm ends. If not, the antenna topology obtained by iteration and the true performance of the corresponding antenna are added to the data set, and the steps of convolutional neural network training are returned to retrain the convolutional neural network.

[0008] Further, the transfer function represents the process of mapping the velocity of particles to binary values in the binary particle swarm algorithm. Different forms of transfer functions affect the exploratory and convergence properties of the binary particle swarm algorithm.

[0009] Further, in the optimization stage of each iteration, the binary particle swarm algorithm will perform two optimizations in parallel. The number of parallel processes for each optimization is N, and N transfer functions are used. In the first optimization, the initial population is randomly generated. In the second optimization, the existing optimal solution is added to the initial population. After removing duplicates and sorting the obtained 2N groups of solutions, the N groups of solutions with the best fitness function are selected, and the full-wave simulation software is started for parallel simulation verification, where 3 ≤ N ≤ 10.

[0010] Further, the number of parallel processes for each optimization is 5, and the transfer functions are respectively:

[0011]

[0012] T3(v) = |v(t) 1.5 |

[0013]

[0014] where α(t) is a value that linearly increases as the iteration number t increases, and v(t) represents the velocity value of the particle at the iteration number t.

[0015] Second aspect, the present invention provides a parallel antenna topology optimization design system using multiple transfer functions. Based on an iterative machine learning-assisted antenna topology pixel optimization design algorithm architecture, in the optimization stage of each iteration, a parallel binary particle swarm optimization algorithm based on multiple transfer functions is adopted, and multiple independent BPSO algorithms are called in parallel. Based on the same machine learning surrogate model, the same objective function is optimized. The difference between different algorithms lies in the different transfer functions used and whether the known optimal solution is added to the initial population of the algorithm.

[0016] Third aspect, the present invention provides a computer system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the steps of a parallel antenna topology optimization design using multiple transfer functions are implemented.

[0017] Fourth aspect, the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by the processor, the steps of a parallel antenna topology optimization design using multiple transfer functions are implemented.

[0018] Fifth aspect, the present invention provides a computer program product including a computer program. When the computer program is executed by the processor, the steps of a parallel antenna topology optimization design using multiple transfer functions are implemented.

[0019] Beneficial effects: Compared with the traditional machine learning-assisted antenna topology design method, the method provided by the present invention can affect the exploration and convergence of the binary particle swarm optimization algorithm by using different forms of transfer functions, so as to explore different antenna topologies in the design space, generate multiple high-quality candidate solutions each time, and then drive the simulation software to perform parallel verification, which can greatly improve the convergence time and the performance of the finally optimized antenna structure. This method can be used in the field of topology optimization design of different types of antenna structures. Description of the Drawings

[0020] Figure 1 is the algorithm flowchart of the parallel antenna topology optimization design method using multiple transfer functions described in the present invention;

[0021] Figure 2 is the structural block diagram of a typical convolutional neural network used in the present invention;

[0022] Figure 3 is the structural schematic diagram of a side-fed microstrip patch antenna optimized by the implementation example of the present invention;

[0023] Figure 4It is a comparison graph of the convergence time between the optimization method described in the present invention and the traditional machine learning-assisted antenna topology design method in the implementation example;

[0024] Figure 5 It is a schematic comparison diagram of the antenna performance optimized by the optimization method of the present invention and the initial value in the implementation example. Detailed implementation manners

[0025] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] A parallel antenna topology optimization design method using multiple transfer functions disclosed in an embodiment of the present invention is based on an iterative machine learning-assisted antenna topology pixel optimization design algorithm architecture. In the optimization stage of each iteration, a parallel binary particle swarm algorithm based on multiple transfer functions is used, which can generate multiple high-quality candidate solutions each time, and then drive the simulation software to perform parallel verification, realizing an improvement in the convergence speed compared with the traditional algorithm architecture. Among them, the parallel binary particle swarm algorithm based on multiple transfer functions parallelly calls multiple independent BPSO algorithms, optimizes the same objective function based on the same machine learning surrogate model, and the difference between different algorithms lies in the different transfer functions used and whether the known optimal solution is added to the initial population of the algorithm.

[0027] Specifically, the parallel antenna topology optimization design method using multiple transfer functions is as Figure 1 shown, and includes the following steps:

[0028] (1) Initial sampling: First, determine the topological region range for antenna design and divide it into m×n sub-blocks, which respectively correspond to the elements in a binary m×n matrix. The following mapping relationship is adopted: Map the elements in the matrix with a value of 0 to air, and map 1 to metal. Thus, the topological structure of the planar antenna can be represented by constructing a binary m×n matrix. Using random sampling, uniform sampling, or hybrid sampling methods, k binary m×n matrices are obtained, which are defined as the input parameters X of the data set.

[0029] (2) Full-wave simulation to construct the data set: Convert the sampling X in step (1) into an antenna topology structure model, and use full-wave simulation software such as HFSS for calculation to obtain the corresponding set of antenna performance responses, which is defined as the output parameter Y of the data set. Thus, a data set M composed of the input parameter X and the output parameter Y is constructed.

[0030] (3) Convolutional neural network training: Introduce convolutional neural network and its learning method, train the data set M, learn the relationship between input parameters and output parameters, and obtain the proxy model R. For common antenna topology optimization design problems, a convolutional neural network structure with fewer layers should be selected to reduce the consumption of training time. A typical convolutional neural network structure is as follows: Figure 2 As shown, it includes two convolutional layers, two activation function layers, two pooling layers and one fully connected layer.

[0031] (4) Parallel binary particle swarm optimization with multiple transfer functions: The binary particle swarm optimization (BPSO) is introduced to optimize the elements in the binary m×n matrix. The fitness function can be set to the antenna's reflection, radiation performance, or a combination of multiple performances. The antenna is optimized at different inputs X. i Performance under Y i It is predicted by the proxy model R trained in step (3). Since the proxy model is used to optimize the antenna topology, this step greatly reduces the computing time required for the traditional optimization process based on evolutionary algorithms. Here, taking advantage of the parallel computing capabilities of multi-core computers, multiple independent BPSO algorithms are called in parallel to optimize the same objective function based on the same machine learning proxy model R. The difference between different algorithms is the different transfer functions used and whether the known optimal solution is added to the initial population of the algorithm. The optimized matrix is ​​defined as X1, and the corresponding proxy model prediction performance is Y1.

[0032] (5) Parallel full-wave simulation verification: The optimized matrix X1 is converted into an antenna topology model, and full-wave simulation is used for parallel calculation to obtain the actual performance Y2 of the verified antenna.

[0033] (6) Determine whether the termination condition is met: The termination condition can be set to the maximum number of iterations being reached or the actual performance Y2 of the antenna meeting the optimization goal; if the termination condition is met, the algorithm ends; if not, the antenna topology X1 obtained in this iteration and its corresponding actual performance Y2 of the antenna are added to the data set M, that is, the data set is updated, and then return to step (3) to retrain the convolutional neural network.

[0034] For example, this embodiment discloses a method for optimizing antenna topology pixel using machine learning-assisted optimization. Consider a typical quad-band side-fed microstrip patch antenna design, whose structure is as follows: Figure 3As shown, it includes a dielectric plate with a height h = 1.2 mm, a width w1 = 18 mm, a length l1 = 34 mm, and a dielectric constant of 4.4. Its bottom is a metal floor with a width w1 and a length l5 = 10 mm, and its top is composed of a microstrip feeder and a rectangular topology optimization design area. The width of the microstrip feeder is w3 = 2 mm, and the length l3 = 14 mm. The width of the area to be optimized is w2 = 14 mm, and the length l2 = 18 mm. It is divided into m × n sub-blocks, where m = 10, n = 8, the length of the sub-block l4 = 1.9 mm, the width w4 = 1.9 mm, and the length of the overlapping part between two adjacent patches is d = 0.1 mm. Figure 3 An example of mapping the antenna topology into a matrix X is given, where the gray squares represent metal and are mapped to 1 in the matrix, and the white squares represent air and are mapped to 0 in the matrix. The design goal of the antenna is set to minimize the |S 11 | value in the frequency bands of 2.35 - 2.45, 3.45 - 3.55, 5.15 - 5.25, and 5.75 - 5.85 GHz. The algorithm runs on a computer with a maximum parallel number of 5.

[0035] In this example, the convolutional neural network structure adopted in step (3) is as Figure 2 shown, which are, from front to back, a convolutional layer, an activation function layer, a pooling layer, a convolutional layer, an activation function layer, a pooling layer, and a fully connected layer. The parameters of the first convolutional layer include a square convolutional kernel of size 3, the number of convolutional kernels is 16, and the zero-padding range is set to 2 in both the up, down, left, and right directions. The parameters of the second convolutional layer include a square convolutional kernel of size 3, the number of convolutional kernels is 64, and the zero-padding range remains unchanged; the activation function layer all adopts the ReLU activation function; the pooling layer all adopts the max pooling method, and the pooling kernel size is 2; a batch normalization layer is used for processing between the convolutional layer and the activation function layer.

[0036] In this example, the parallel BPSO algorithm based on multiple transfer functions used in step (4) has its optimization goal set as:

[0037]

[0038] where f c is the frequency point to be considered, represents the amplitude of the reflection coefficient of the antenna predicted by the convolutional neural network at all frequency points under the matrix X corresponding to the given antenna topology. The optimization task is to find the matrix X corresponding to the appropriate antenna topology such that is minimized, that is, to minimize its worst value. In BPSO, the role of the transfer function is to convert the continuous velocity value of the particle into a probability value to determine the update of the particle's position (0 or 1), and the update criterion is as follows:

[0039]

[0040] Five transfer functions are used here, and the transfer functions are respectively:

[0041]

[0042] T3(v) = |v(t) 1.5 |

[0043]

[0044]

[0045] Among them, α(t) in T1(v) is a value that linearly increases with the increase of the iteration number t, used to adjust the exploratory nature of this transfer function. v(t) represents the velocity value of the particle at the iteration number t. The algorithm will run two optimizations in parallel. The number of parallel processes for each optimization is 5. The BPSO based on the above 5 transfer functions is used. The difference between the two optimizations is that the initial population is randomly generated in the first optimization, while in the second optimization, the existing optimal solution is added to the initial population. Randomly generating the initial population in the first optimization will have some randomness and is more inclined to explore different regions. Adding the current optimal solution to the initial population in the second optimization to explore near the current optimal solution is more inclined to local convergence. By performing two optimizations, the exploratory nature and convergence are balanced. The 10 groups of solutions obtained from the two optimizations are first screened to remove the solutions that duplicate the existing antenna topologies in the dataset, and then the remaining solutions are sorted. The 5 groups of solutions with the best fitness function are selected, and the full-wave simulation software is started for parallel simulation verification.

[0046] Figure 4 A comparison between the parallel antenna topology pixel optimization design method of the present invention and the traditional antenna topology pixel optimization design method is given. Both algorithms are respectively performed 5 times, and the average value of the descent curves of their fitness functions is taken. It can be seen that compared with the traditional antenna topology pixel optimization design method, the parallel antenna topology pixel optimization design method given by the present invention can greatly reduce the calculation time. Taking the |S 11 | value being better than -10 dB in the full frequency band, that is, the fitness function value being lower than 0.3162 as the design goal, it can be seen that the proposed parallel antenna topology pixel optimization design method can reach the design goal within 3.17 hours, while the traditional method requires 13.53 hours to reach the design goal.

[0047] Figure 5 A reflection coefficient diagram of a typical optimization result is given. It can be seen that compared with the initial value (i.e., the all-metal patch with all 1s in the matrix), the optimization result meets the reflection coefficient requirement of -10 dB in four frequency bands.

[0048] An embodiment of the present invention also discloses a parallel antenna topology optimization design system using multiple transfer functions. This system is based on an iterative machine learning-assisted antenna topology pixel optimization design algorithm architecture. In the optimization stage of each iteration, a parallel binary particle swarm optimization algorithm based on multiple transfer functions is used to parallelly call multiple independent BPSO algorithms. Based on the same machine learning surrogate model, the same objective function is optimized. The difference between different algorithms lies in the different transfer functions used and whether the known optimal solution is added to the initial population of the algorithm.

[0049] An embodiment of the present invention also discloses a computer system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the computer program is executed by the processor, the steps of a parallel antenna topology optimization design using multiple transfer functions are implemented.

[0050] An embodiment of the present invention also discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program. When the computer program is executed by the processor, the steps of a parallel antenna topology optimization design using multiple transfer functions are implemented.

[0051] An embodiment of the present invention also discloses a computer program product, including a computer program. When the computer program is executed by the processor, the steps of a parallel antenna topology optimization design using multiple transfer functions are implemented.

[0052] As mentioned above, the above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A parallel antenna topology optimization design method using multiple transfer functions, characterized in that An iterative machine learning-assisted antenna topology pixel optimization design algorithm architecture. In the optimization stage of each iteration, a parallel binary particle swarm BPSO algorithm based on multiple transfer functions is adopted. Multiple independent BPSO algorithms are called in parallel. Based on the same machine learning surrogate model, the same objective function is optimized. The difference between different algorithms lies in the different transfer functions used and whether the known optimal solution is added to the initial population of the algorithm.

2. The parallel antenna topology optimization design method using multiple transfer functions according to claim 1, characterized in that, The iterative machine learning-assisted antenna topology pixel optimization design algorithm architecture includes: initial sampling; full-wave simulation to construct a data set; convolutional neural network training; binary particle swarm algorithm optimization; parallel full-wave simulation verification; determining whether the termination condition is met. If the termination condition is met, the algorithm ends. If not, the antenna topology obtained by iteration and the true performance of the corresponding antenna are added to the data set, and the steps of convolutional neural network training are returned to retrain the convolutional neural network.

3. A parallel antenna topology optimization design method using multiple transfer functions according to claim 1, characterized in that, The transfer function represents the process of mapping the velocity of particles to binary values in the binary particle swarm algorithm. Different forms of transfer functions affect the exploration and convergence of the binary particle swarm algorithm.

4. A parallel antenna topology optimization design method using multiple transfer functions according to claim 1, characterized in that, In the optimization stage of each iteration, the binary particle swarm algorithm will run two optimizations in parallel. The number of parallel processes for each optimization is N, and N transfer functions are used. In the first optimization, the initial population is randomly generated. In the second optimization, the existing optimal solution is added to the initial population. The 2N groups of solutions obtained are de-duplicated and sorted, and the N groups of solutions with the best fitness function are selected, and the full-wave simulation software is started for parallel simulation verification, where 3 ≤ N ≤ 10.

5. A parallel antenna topology optimization design method using multiple transfer functions according to claim 1, characterized in that, The number of parallel processes for each optimization is 5, and the transfer functions are respectively: T3(v) = |v(t) 1.5 | Where α(t) is a value that linearly increases as the iteration number t increases, and v(t) represents the velocity value of the particle at the iteration number t.

6. A parallel antenna topology optimization design system using multiple transfer functions, characterized in that, An iterative machine learning-assisted antenna topology pixel optimization design algorithm architecture. In the optimization stage of each iteration, a parallel binary particle swarm algorithm based on multiple transfer functions is adopted. Multiple independent BPSO algorithms are called in parallel. Based on the same machine learning surrogate model, the same objective function is optimized. The difference between different algorithms lies in the different transfer functions used and whether the known optimal solution is added to the initial population of the algorithm.

7. A parallel antenna topology optimization design system using multiple transfer functions according to claim 6, characterized in that, In the optimization stage of each iteration, the binary particle swarm algorithm will run two optimizations in parallel. The number of parallel processes for each optimization is N, and N transfer functions are used. In the first optimization, the initial population is randomly generated. In the second optimization, the existing optimal solution is added to the initial population. The 2N groups of solutions obtained are de-duplicated and sorted, and the N groups of solutions with the best fitness function are selected, and the full-wave simulation software is started for parallel simulation verification.

8. A computer system, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the computer program is executed by a processor, it implements the steps of a parallel antenna topology optimization design using multiple transfer functions according to any one of claims 1-5.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of a parallel antenna topology optimization design using multiple transfer functions according to any one of claims 1-5.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of a parallel antenna topology optimization design using multiple transfer functions according to any one of claims 1-5.

Images