CN108920841B - Antenna design method - Google Patents

Abstract

The invention discloses an antenna design method, which includes constructing an initial model of the antenna; selecting an input sample to input the initial model of the antenna to obtain the response of the antenna model; simulating the structure of a neural network to obtain an antenna proxy model; constructing antenna design parameter variables and objective functions; The parameter variables are input into the antenna proxy model to get the response and calculate the objective function value; the final antenna design parameters are obtained by judging the objective function value. The present invention uses a hybrid optimization algorithm to jointly optimize the network structure and initial structural parameters of the neural network, simplifies the neural network structure, reduces the calculation cost of the neural network, improves the prediction accuracy of the neural network, and then uses the neural network as a proxy model to fit antenna design parameter samples The electromagnetic simulation data, instead of the time-consuming electromagnetic simulation, realizes the instantaneous approximate calculation from the antenna structure parameters to the electromagnetic response, reduces the number of electromagnetic simulations, greatly reduces the calculation cost, and significantly improves the antenna design efficiency.

Description

Antenna Design Method

technical field

The invention specifically relates to an antenna design method.

Background technique

With the development of economy and technology and the improvement of people's living standards, communication has become an indispensable link in people's production and life.

As an energy conversion device between guided electromagnetic waves and free space electromagnetic waves, antennas are widely used in mobile communications, radar, satellite communications and other fields. The development of modern wireless communication systems not only requires antennas to be light in weight, low in cost, easy to manufacture and easy to integrate, but also puts forward unprecedented requirements for miniaturization, broadband, multiband, conformal and integrated design of antennas.

The design of conventional antennas is generally based on regular structures, using existing empirical formulas, combined with the design experience of antenna engineers and physical measurement and debugging. With such antenna design methods, the antenna design cycle is often longer; more importantly, these conventional antenna design methods are powerless for antenna designs with irregular structures, new structures, and high-performance requirements. Moreover, when optimally designing a multi-parameter and multi-objective antenna structure, the design process is lengthy, and the optimization ability and efficiency become very poor.

Intelligent optimization algorithms can be considered as simple and general objective optimization strategies, usually imitating various biological or social phenomena (such as swarm intelligence, genetic processes, etc.). These algorithms can realize automatic adjustment of antenna structure parameters and simultaneous optimization of multiple design objectives. Nevertheless, along with the benefits of intelligent population-based optimization algorithms, an obvious drawback of such algorithms is that the optimization process requires a huge number of model evaluations. A single evaluation of a realistic antenna model usually takes several minutes to tens of minutes, and there are often more than one evaluation model in practical applications, so the calculation cost is extremely high, which also hinders the direct application of intelligent optimization algorithms in the design process. It has also indirectly led to the development of various strategies aimed at reducing computational costs. On the other hand, the problem of high computational cost may be partially solved by exploiting large-scale computing resources in the form of supercomputers with multiple CPU or GPU units and multiple licenses of auxiliary computational design software (especially EM solvers). However, such hardware configurations are not widely available, they provide very low speed-up-cost ratios, and are therefore not realistic.

Contents of the invention

The purpose of the present invention is to provide an antenna design method that can greatly reduce the calculation cost of antenna design, thereby ensuring efficient and reliable antenna design.

This antenna design method provided by the present invention comprises the following steps:

S1. According to the design requirements of the antenna, construct the initial model of the antenna;

S2. Select several groups of antenna design parameter values in the antenna design space as input samples, and input the samples into the initial antenna model obtained in step S1, so as to obtain the antenna model responses corresponding to each input sample;

S3. Using the neural network to simulate the mapping relationship between the input samples obtained in step S2 and the antenna model response, so as to obtain the corresponding antenna proxy model;

S4. Construct several groups of antenna design parameter variables, and construct several antenna design objective functions according to antenna design requirements;

S5. Inputting the antenna design parameter variable obtained in step S4 into the antenna proxy model obtained in step S3, thereby obtaining the corresponding responses of each group of antenna design parameters, and calculating the corresponding objective function value of each antenna design parameter according to the obtained responses;

S6. Judging the objective function value obtained in step S5:

If any objective function value meets the pre-set requirements, the antenna design parameter variable corresponding to the objective function value is determined to be the final antenna design parameter;

If all the objective function values do not meet the pre-set requirements, generate new sets of antenna design parameter variables, and repeat steps S5-S6 until the objective function values meet the pre-set requirements, so as to obtain the final antenna design parameter.

Selecting several groups of antenna design parameter values in the antenna design space described in step S2 is specifically selecting several groups of antenna design parameter values in the antenna design space by using the Latin hypercube sampling method.

As described in step S2, input the samples into the initial antenna model to obtain the antenna model response corresponding to each input sample. Specifically, the electromagnetic simulation tool is used to simulate and solve the initial antenna model with the input samples, so as to obtain the response of each input sample. The corresponding antenna model response.

In step S3, the neural network is used to simulate the mapping relationship between the obtained input samples and the antenna model responses, specifically, according to the input samples and the corresponding antenna model responses, a hybrid optimization algorithm is used to optimize the neural network structure and parameters, so that A neural network structure capable of simulating input samples and corresponding antenna model responses is obtained, and the neural network structure is used as the final antenna proxy model.

The described adopting hybrid optimization algorithm to optimize the neural network structure and parameters is specifically optimized by adopting the following steps:

A. According to the antenna design parameter variables and their corresponding response vectors, respectively determine the number of input neurons n _i and the number of output neurons n _o of the neural network;

B. Determine the number n _h of neurons in the hidden layer of the neural network;

C. Encode the neural network structure and initial structure parameters respectively, and initialize the mixed particle swarm;

D. Construct a fitness function f, which is used to represent the prediction error between the predicted response of the neural network structure to the input sample and the real response;

E. Calculate the fitness function value of each mixed particle;

G. After optimizing the neural network corresponding to the neuron data of different hidden layers, the prediction error of the neural network is obtained, and the neural network with the smallest prediction error is selected as the final antenna proxy model.

In step C, the neural network structure and initial structural parameters are respectively encoded and the mixed particle swarm is initialized. Specifically, the following steps are used for encoding and initialization:

(1) Perform binary encoding on the neural network structure, perform real-number encoding on the initial structural parameters of the neural network, and use the following formula to calculate the dimension d of binary and real-numbered particles:

d＝n _i ×n _h +n _h +n _h ×n _o +n _o

In the formula, n _i is the number of input neurons of the neural network, n _o is the number of output neurons, and n _h is the number of hidden layer neurons of the neural network;

(2) Generate d-dimensional binary particles, which represent the link switch between the input layer and the hidden layer of the neural network, the link switch between the hidden layer and the output layer, and the link threshold between the hidden layer and the output layer, where 1 means the link exists, 0 means the link does not exist;

(3) Generate real number particles in the d dimension (0,1), representing the weight between the input layer and the hidden layer of the neural network, the weight between the hidden layer and the output layer, and the weight between the hidden layer and the output layer threshold;

(4) Arrange real number particles and binary particles in order to form 2d-dimensional mixed particles representing the neural network structure and initial structure parameters, and initialize the mixed particle swarm.

The construction of the fitness function f described in step D is specifically to use the following formula to construct the fitness function f:

式中q为输入样本数目，k表示1到n₀之间的变量索引，Y_k(t)为各组输入样本的响应值，y_k(t)为神经网络对于各组输入样本的预测响应值。where err is the absolute mean error and

In the formula, q is the number of input samples, k represents the variable index between 1 and n ₀ , Y _k (t) is the response value of each group of input samples, and y _k (t) is the predicted response of the neural network to each group of input samples value.

The calculation of the fitness function value of each mixed particle described in step E is specifically calculated by the following steps:

1) Multiply the d-dimensional real part of each mixed particle with the binary part to obtain the structural parameters of the neural network, and then use the structural parameters to construct the neural network;

2) Input each group of input samples into the neural network as input data to obtain the predicted response vector corresponding to the input samples, and solve the prediction error of the input samples by the neural network, and the prediction error is the fitness function value.

The optimization of the neural network corresponding to the neuron data of different hidden layers described in step G is optimized by using HPSO (Hybrid real-binary particle swarm optimization, hybrid real-binary particle swarm optimization algorithm).

The neural network is a BP neural network, a perceptron neural network or a linear neural network; the hybrid optimization algorithm is a hybrid particle swarm optimization algorithm or a hybrid Taguchi genetic algorithm; the generation of new groups of antenna design parameter variables described in step S6 , specifically using a multi-objective intelligent algorithm to generate new sets of antenna design parameter variables, and the multi-objective intelligent algorithm is a multi-objective evolutionary algorithm based on decomposition, a non-dominated sorting evolutionary algorithm, a multi-objective genetic algorithm or a multi-objective particle swarm algorithm.

The antenna design method provided by the present invention uses a hybrid optimization algorithm to jointly optimize the network structure and initial structural parameters of the neural network, simplify the neural network structure, reduce the calculation cost of the neural network, improve the prediction accuracy of the neural network, and then use the neural network as The surrogate model fits the electromagnetic simulation data of the antenna design parameter sample, replaces the time-consuming electromagnetic simulation to realize the instantaneous approximate calculation from the antenna structure parameters to the electromagnetic response, reduces the number of electromagnetic simulations, and greatly reduces the calculation cost; and the method of the present invention effectively combines The multi-objective intelligent algorithm, surrogate model and antenna design can significantly improve the efficiency of antenna design, especially in solving complex high-dimensional multi-objective antenna design problems, and its advantages are more obvious.

Description of drawings

Fig. 1 is the method flowchart of the method of the present invention.

Fig. 2 is a schematic diagram of an initial antenna model constructed in an embodiment of the present invention.

FIG. 3 is a graph of return loss curves of 6 antennas meeting the design objectives obtained through the design of the embodiment of the present invention.

Detailed ways

As shown in Figure 1, it is a method flowchart of the method of the present invention: this antenna design method provided by the present invention comprises the following steps:

S1. According to the design requirements of the antenna, construct the initial model of the antenna;

S2. Select several groups of antenna design parameters in the antenna design space as input samples (the Latin hypercube sampling method can be used to select), and input the samples into the initial model of the antenna obtained in step S1, and use electromagnetic simulation tools to analyze the input samples The initial model of the antenna is simulated and solved, so as to obtain the corresponding antenna model response of each input sample (the response is each performance index of the antenna, including antenna return loss value, gain or standing wave ratio, etc.);

S3. Use the neural network to simulate the mapping relationship between the input samples obtained in step S2 and the antenna model response, so as to obtain the corresponding antenna proxy model; specifically, according to the input samples and their corresponding antenna model responses, use a hybrid optimization algorithm to optimize Neural network structure and parameters, so as to obtain a neural network structure capable of simulating the input sample and its corresponding antenna model response, and use the neural network structure as the final antenna proxy model; specifically, the following steps are used for optimization:

A. According to the antenna design parameter variables and their corresponding response vectors, respectively determine the number of input neurons n _i and the number of output neurons n _o of the neural network;

B. Determine the number n _h of neurons in the hidden layer of the neural network;

C. Encode the neural network structure and initial structural parameters, and initialize the mixed particle swarm; specifically, the following steps are used for encoding and initialization:

(1) Perform binary encoding on the neural network structure, perform real-number encoding on the initial structural parameters of the neural network, and use the following formula to calculate the dimension d of binary and real-numbered particles:

d＝n _i ×n _h +n _h +n _h ×n _o +n _o

In the formula, n _i is the number of input neurons of the neural network, n _o is the number of output neurons, and n _h is the number of hidden layer neurons of the neural network;

(2) Generate d-dimensional binary particles, which represent the link switch between the input layer and the hidden layer of the neural network, the link switch between the hidden layer and the output layer, and the link threshold between the hidden layer and the output layer, where 1 means the link exists, 0 means the link does not exist;

(3) Generate real number particles in the d dimension (0,1), representing the weight between the input layer and the hidden layer of the neural network, the weight between the hidden layer and the output layer, and the weight between the hidden layer and the output layer threshold;

(4) Arranging real number particles and binary particles in order to form 2d-dimensional mixed particles representing the neural network structure and initial structural parameters, and initializing the mixed particle swarm;

D. Construct a fitness function f, which is used to represent the prediction error between the predicted response of the neural network structure to the input sample and the real response; specifically, the following formula is used to construct the fitness function f:

式中q为输入样本数目，k为1到n₀之间的变量索引，Y_k(t)为各组输入样本的响应值，y_k(t)为神经网络对于各组输入样本的预测响应值；where err is the absolute mean error and

In the formula, q is the number of input samples, k is the variable index between 1 and n ₀ , Y _k (t) is the response value of each group of input samples, and y _k (t) is the predicted response of the neural network to each group of input samples value;

E. Calculate the fitness function value of each mixed particle; specifically, the following steps are used for calculation:

1) Multiply the d-dimensional real part of each mixed particle with the binary part to obtain the structural parameters of the neural network, and then use the structural parameters to construct the neural network;

2) Input each group of input samples into the neural network as input data, obtain the corresponding predicted response vector of the input samples, and solve the prediction error of the input samples by the neural network, and the prediction error is the fitness function value;

G. Optimize the neural network corresponding to the neuron data of different hidden layers (such as using HPSO (Hybridreal-binary particle swarm optimization, hybrid real-binary particle swarm optimization algorithm) to optimize) to obtain the prediction error of the neural network, and select The neural network with the smallest prediction error is used as the final antenna proxy model;

In specific implementation, the neural network can use BP neural network, perceptron neural network or linear neural network, etc.; while the hybrid optimization algorithm can use hybrid particle swarm optimization algorithm or hybrid Taguchi genetic algorithm, etc.;

S4. Construct several groups of antenna design parameter variables, and construct several antenna design objective functions according to antenna design requirements;

S5. Inputting the antenna design parameter variable obtained in step S4 into the antenna proxy model obtained in step S3, thereby obtaining the corresponding responses of each group of antenna design parameters, and calculating the corresponding objective function value of each antenna design parameter according to the obtained responses;

S6. Judging the objective function value obtained in step S5:

If any objective function value meets the pre-set requirements, the antenna design parameter variable corresponding to the objective function value is determined to be the final antenna design parameter;

If all the objective function values do not meet the pre-set requirements, generate new sets of antenna design parameter variables, and repeat steps S5-S6 until the objective function values meet the pre-set requirements, so as to obtain the final antenna design Parameters; wherein, a multi-objective intelligent algorithm can be used to generate new sets of antenna design parameter variables, and the multi-objective intelligent algorithm is a multi-objective evolutionary algorithm based on decomposition, a non-dominated sorting evolutionary algorithm, a multi-objective genetic algorithm or a multi-objective Particle Swarm Algorithm.

The method of the present invention is further described below in conjunction with a specific embodiment:

The goal is to design a two-objective planar multi-band antenna; the neural network adopts BP neural network model, the hybrid optimization algorithm adopts hybrid particle swarm optimization algorithm (HPSO), and the multi-objective intelligent algorithm adopts decomposition-based multi-objective evolutionary algorithm (MOEA/D) , the electromagnetic simulation tool adopts HFSS.

The specific design process is as follows:

The antenna model is constructed as shown in Figure 2. The design space of the antenna model, that is, its constraint condition is the size limit of 10 antenna parameters, as shown in Table 1 below:

Table 1 Antenna modeling constraints (unit: mm)

parameter L L1 L2 L3 L4 scope [36.4,40] [16,19] [10,12.5] [8.5,10.5] [2.8,3.9] parameter L5 W W1 W2 g scope [9.5,11.5] [19,24] [6.5,8.3] [8.7,11.2] [1.8,2.1]

Use the Latin hypercube sampling method to select 200 groups of antenna design parameter variables in the antenna design space as input samples, and use electromagnetic simulation tools to solve the response vectors of each group of antenna design parameter variables, that is, the return loss values of 15 frequency sampling points as output samples , 200 groups of antenna design parameter variables and the return loss values corresponding to each frequency sampling point constitute the sample set for constructing the proxy model.

Determine the number of input neurons and output neurons n _i =10, n _o =15 of the BP neural network proxy model according to the antenna design parameter variable and the return loss value corresponding to each frequency sampling point;

Determine the range of neurons in the hidden layer of the BP neural network proxy model [10,20];

Calculate binary and real number particle dimensions d=25×n _h +15, respectively initialize d-dimensional binary and real number particles, combine them into mixed particles, and initialize particle swarm;

Construct the HPSO algorithm to optimize the network structure of BP neural network and the fitness function of the initial structural parameters:

Y _k (t) is the return loss value of calling the electromagnetic simulation tool to simulate and solve each group of antenna design parameter variables, and y _k (t) is the return loss value of each group of antenna design parameter variables predicted by using the BP neural network proxy model;

Multiply the real number part of each mixed particle with the binary part to obtain the structural parameters of the non-fully connected neural network proxy model, use these structural parameters and sample sets to construct the non-fully connected BP neural network proxy model, and combine the design parameters of each group of antennas The variable is input as the input data into the non-fully connected BP neural network proxy model, and the response vector of the antenna design parameter variable is predicted, and the prediction error of the proxy model to the antenna design parameter variable is solved;

Use HPSO to optimize the non-fully connected BP neural network proxy model corresponding to the number of neurons in different hidden layers 1000 times and output the prediction error, and select the one with the smallest error as the proxy model of the planar multi-band antenna;

Randomly initialize 40 groups of antenna design parameter variables x ₁ , x ₂ ,...,x ₄₀ in the antenna design space as the initial population of the MOEA/D algorithm, and construct 2 antenna design targets according to the antenna design requirements;

Objective function 1: Return loss value S ₁₁ <-10dB in the three frequency bands of 2.40～2.60GHz, 3.30～3.80GHz, and 5.00～5.90GHz;

Where n is the number of sampling points in the above three frequency bands, f _i is the frequency of sampling points in the frequency band, and S ₁₁ (f _i ) is the return loss value at frequency f _i ;

Objective function 2: Antenna size;

F ₂ =w×l

Taking 40 groups of antenna design parameter variables as input values, call the non-fully connected BP neural network model to predict the return loss value of each group of antenna design parameter variables at each frequency sampling point, and then solve the objective function value F ₁ according to the return loss value, Solve the objective function value F ₂ according to the design parameters;

Judging whether the objective function value obtained in step 5 meets the antenna design requirements, if so, proceed to step 7, otherwise, use MOEA/D to update and generate 40 new sets of antenna design parameter variables, and return to step 5 until the design requirements are obtained. Antenna design parameters, or up to the number of iterations set by MOEA/D;

If the antenna design results meet the two antenna design objectives, the iteration ends.

The design parameters obtained by applying the method of the present invention are as shown in Table 2, and the reflection curves of 6 antennas that meet the design objectives are as shown in Figure 3. Under different area parameters, the antennas are at 2.33～2.63GHz, 3.17 The return loss values of the three frequency bands of ～3.92GHz and 4.97～5.99GHz are all less than -10dB, which meets the performance requirements of antenna design.

Table 2 Design of 6 antenna sizes that meet the design goals

design x⁽¹⁾ x⁽²⁾ x⁽³⁾ x⁽⁴⁾ x⁽⁵⁾ x⁽⁶⁾ F₁(dB) -13.26 -12.61 -12.40 -12.05 -11.73 -11.21 F₂(mm) 885.92 839.22 825.10 794.22 777.48 726.93 L 39.2 39.4 37.0 36.6 37.2 36.9 L1 17.2 18.3 18.9 16.8 17.4 18.2 L2 11.1 10.6 10.3 10.7 12.5 12.2 L3 9.2 9.9 9.2 9.5 8.9 9.4 L4 3.6 3.4 3.5 3.1 3.0 3.1 L5 11.3 10.3 10.9 11.4 11.5 10.5 W 22.6 21.3 22.3 21.7 20.9 19.7 W1 8.1 6.9 8.2 7.2 6.7 6.5 W2 10.5 9.8 10.4 10.1 9.6 9.1 g 1.8 1.9 2.0 2.0 2.0 1.8

Secondly, the traditional electromagnetic simulation (EM) design method, MOEA/D combined with the BP network model that only optimizes structural parameters, and MOEA/D combined with the non-fully connected BP neural network model are used to design the antenna. The comparison results of the total calculation cost of the antenna are shown in Table 3 shown.

Table 3 Computational cost comparison of three antenna design methods

Finally, for the 6 sets of antenna design parameter variables obtained in the design, the BP neural network model is used to directly predict (prediction result 1) and the non-fully connected BP neural network model to predict the response value and calculate the objective function F ₁ (prediction result 2), Then directly use the simulated response value to calculate its objective function F ₁ , and its error rate comparison is shown in Table 4.

Table 4 Accuracy comparison of the two prediction methods

design x⁽¹⁾ x⁽²⁾ x⁽³⁾ x⁽⁴⁾ x⁽⁵⁾ x⁽⁶⁾ Prediction result 1 -15.03 -13.93 -13.95 -11.07 -13.76 -10.11 Prediction result 2 -13.26 -12.61 -12.40 -12.05 -11.73 -11.21 Simulation results -13.12 -12.88 -12.68 -12.24 -12.07 -11.71 error rate 1 14.56% 8.15% 10.02% 9.56% 14.00% 13.66% error rate 2 1.83% 2.10% 2.21% 1.56% 2.82% 4.27%

Claims (10)

1. An antenna design method includes the following steps:

s1, constructing an antenna initial model according to design requirements of an antenna;

s2, selecting a plurality of groups of antenna design parameters in an antenna design space as input samples, and inputting the samples into the antenna initial model obtained in the step S1 so as to obtain antenna model responses corresponding to the input samples;

s3, simulating the mapping relation between the input sample obtained in the step S2 and the antenna model response by adopting a neural network, so as to obtain a corresponding antenna agent model;

s4, constructing a plurality of groups of antenna design parameter variables, and constructing a plurality of antenna design objective functions according to antenna design requirements;

s5, inputting the antenna design parameter variables obtained in the step S4 into the antenna proxy model obtained in the step S3, so as to obtain responses corresponding to each group of antenna design parameters, and calculating objective function values corresponding to each antenna design parameter according to the obtained responses;

s6, judging the objective function value obtained in the step S5:

if the objective function value meets the preset requirement, the antenna design parameter variable corresponding to the objective function value is determined as the final antenna design parameter;

if all the objective function values do not meet the preset requirements, generating a plurality of new antenna design parameter variables, and repeating the steps S5-S6 until the objective function values meet the preset requirements, thereby obtaining the final antenna design parameters.

2. The method of claim 1, wherein the selecting step S2 selects a plurality of antenna design parameters in the antenna design space, specifically selecting a plurality of antenna design parameters in the antenna design space by using a latin hypercube sampling method.

3. The antenna design method according to claim 1, wherein the samples are input into the antenna initial model in step S2 to obtain the antenna model response corresponding to each input sample, and specifically, an electromagnetic simulation tool is used to perform simulation solution on the antenna initial model into which the samples are input to obtain the antenna model response corresponding to each input sample.

4. The antenna design method according to one of claims 1 to 3, wherein the input samples and the corresponding antenna model responses thereof are simulated by using a neural network structure in step S3, specifically, the neural network structure and parameters are optimized by using a hybrid optimization algorithm according to the input samples and the corresponding antenna model responses thereof, so as to obtain the neural network structure capable of simulating the input samples and the corresponding antenna model responses thereof, and the neural network structure is used as a final antenna proxy model.

5. The antenna design method according to claim 4, wherein the optimization of the neural network structure and parameters by the hybrid optimization algorithm is specifically performed by the following steps:

A. respectively determining the number n of input neurons of the neural network according to the antenna design parameter variables and the corresponding response vectors thereof _i And number of output neurons n _o ；

B. Determining the number n of hidden layer neurons of a neural network _h ；

C. Respectively encoding the neural network structure and the initial structure parameters, and initializing a mixed particle swarm;

D. constructing a fitness function f for representing a prediction error between a prediction response and a real response of the neural network structure to the input sample;

E. calculating a fitness function value of each mixed particle;

G. and optimizing the neural networks corresponding to the neuron data of different hidden layers to obtain the prediction error of the neural networks, and selecting the neural network with the minimum prediction error as a final antenna agent model.

6. The antenna design method according to claim 5, wherein the step C of encoding the neural network structure and the initial structure parameters and initializing the hybrid particle swarm respectively comprises the following steps:

(1) Binary coding is carried out on the neural network structure, real number coding is carried out on initial structure parameters of the neural network, and the dimensionality d of binary and real number particles is calculated by adopting the following formula:

d＝n _i ×n _h +n _h +n _h ×n _o +n _o

in the formula n _i Number of input neurons being a neural network, n _o Number of output neurons, n _h Hiding the number of layer neurons for the neural network;

(2) Generating d-dimensional binary particles representing a link switch between an input layer and a hidden layer of the neural network, a link switch between the hidden layer and an output layer and link thresholds of the hidden layer and the output layer, wherein 1 represents that the link exists, and 0 represents that the link does not exist;

(3) Generating real number particles in d dimension (0, 1) representing weights between input layer and hidden layer of the neural network, weights between the hidden layer and output layer, and thresholds of the hidden layer and output layer;

(4) And sequentially arranging the real number particles and the binary particles to form 2 d-dimensional mixed particles representing the neural network structure and the initial structure parameters, and initializing the mixed particle swarm.

7. The antenna design method according to claim 5, wherein the constructing of the fitness function f in step D is specifically to construct the fitness function f by using the following formula:

where err is the absolute average error and

where q is the number of input samples and k is 1 to n ₀ Index of variable between, Y _k (t) is the response value of each set of input samples, y _k And (t) is the predicted response value of the neural network for each set of input samples.

8. The antenna design method according to claim 5, wherein the fitness function value of each hybrid particle is calculated in step E by:

1) Correspondingly multiplying the d-dimensional real number part of each mixed particle with the binary part to obtain the structural parameters of the neural network, and then constructing the neural network by using the structural parameters;

2) And inputting each group of input samples serving as input data into the neural network to obtain a prediction response vector corresponding to the input samples, and solving a prediction error of the neural network on the input samples, wherein the prediction error is a fitness function value.

9. The antenna design method according to claim 5, wherein the neural networks corresponding to the different hidden layer neuron data in step G are optimized, specifically, optimized by using HPSO algorithm.

10. The antenna design method according to claim 4, wherein the neural network is a BP neural network, a perceptron neural network or a linear neural network; the hybrid optimization algorithm is a hybrid particle swarm algorithm or a hybrid Taguchi genetic algorithm; and S6, generating a plurality of new antenna design parameter variables, specifically, generating a plurality of new antenna design parameter variables by adopting a multi-target intelligent algorithm, wherein the multi-target intelligent algorithm is a multi-target evolutionary algorithm based on decomposition, a non-dominated sorting evolutionary algorithm, a multi-target genetic algorithm or a multi-target particle swarm algorithm.

Images