CN115329655A - Lens antenna multi-objective optimization method based on priori knowledge neural network - Google Patents
Lens antenna multi-objective optimization method based on priori knowledge neural network Download PDFInfo
- Publication number
- CN115329655A CN115329655A CN202210666376.0A CN202210666376A CN115329655A CN 115329655 A CN115329655 A CN 115329655A CN 202210666376 A CN202210666376 A CN 202210666376A CN 115329655 A CN115329655 A CN 115329655A
- Authority
- CN
- China
- Prior art keywords
- neural network
- electromagnetic response
- lens antenna
- simulation data
- vector
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 114
- 238000000034 method Methods 0.000 title claims abstract description 39
- 238000005457 optimization Methods 0.000 title claims abstract description 17
- 230000004044 response Effects 0.000 claims abstract description 92
- 238000004088 simulation Methods 0.000 claims description 61
- 239000013598 vector Substances 0.000 claims description 54
- 238000012549 training Methods 0.000 claims description 27
- 238000010606 normalization Methods 0.000 claims description 13
- 238000012545 processing Methods 0.000 claims description 9
- 238000012360 testing method Methods 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 6
- 238000009499 grossing Methods 0.000 claims description 5
- 230000008859 change Effects 0.000 claims description 4
- 238000013461 design Methods 0.000 abstract description 24
- 230000006870 function Effects 0.000 description 12
- 230000005855 radiation Effects 0.000 description 8
- 210000002569 neuron Anatomy 0.000 description 6
- 230000004913 activation Effects 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 238000013507 mapping Methods 0.000 description 3
- 230000010287 polarization Effects 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000010146 3D printing Methods 0.000 description 1
- 101100272279 Beauveria bassiana Beas gene Proteins 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 239000003989 dielectric material Substances 0.000 description 1
- 230000005672 electromagnetic field Effects 0.000 description 1
- 229910052731 fluorine Inorganic materials 0.000 description 1
- 125000001153 fluoro group Chemical group F* 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 230000003014 reinforcing effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H01—ELECTRIC ELEMENTS
- H01Q—ANTENNAS, i.e. RADIO AERIALS
- H01Q15/00—Devices for reflection, refraction, diffraction or polarisation of waves radiated from an antenna, e.g. quasi-optical devices
- H01Q15/0006—Devices acting selectively as reflecting surface, as diffracting or as refracting device, e.g. frequency filtering or angular spatial filtering devices
- H01Q15/0013—Devices acting selectively as reflecting surface, as diffracting or as refracting device, e.g. frequency filtering or angular spatial filtering devices said selective devices working as frequency-selective reflecting surfaces, e.g. FSS, dichroic plates, surfaces being partly transmissive and reflective
-
- H—ELECTRICITY
- H01—ELECTRIC ELEMENTS
- H01Q—ANTENNAS, i.e. RADIO AERIALS
- H01Q15/00—Devices for reflection, refraction, diffraction or polarisation of waves radiated from an antenna, e.g. quasi-optical devices
- H01Q15/02—Refracting or diffracting devices, e.g. lens, prism
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/06—Multi-objective optimisation, e.g. Pareto optimisation using simulated annealing [SA], ant colony algorithms or genetic algorithms [GA]
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a multi-objective optimization method of a lens antenna based on a priori knowledge neural network, which takes INN as a main body of the algorithm, adopts a reverse neural network to carry out reverse design on the lens antenna, inputs of the reverse neural network are electromagnetic responses, outputs are structural parameters, and inversion of the antenna structural parameters of multiple performance indexes is realized. Meanwhile, the invention introduces a plurality of sub-FNNs which are respectively used for providing prior knowledge corresponding to a plurality of performance indexes, the input of the forward neural network is a structural parameter, and the output is electromagnetic response. Compared with the existing lens antenna design method, the design method provided by the invention has the following advantages: 1) The antenna design efficiency is high, and the designed antenna can realize good multi-target characteristics; 2) The KBANN model provided by the method relieves the problem of large data volume requirement of the neural network and provides a new solution. The method can realize the rapid optimization of the multi-target lens antenna, and simultaneously relieve the problem of large data volume requirement of a neural network.
Description
Technical Field
The invention relates to the technical field of lens antennas, in particular to a multi-objective optimization method of a lens antenna based on a priori knowledge neural network (KBANN).
Background
The increasingly complex electromagnetic environment and application requirements require ever increasing performance of radar, guidance, communication, biomedical, electronic countermeasure and radio astronomy systems, which also means increasingly stringent requirements for antenna performance. When designing an antenna, the antenna needs to be considered to meet a plurality of specific performance indexes, such as impedance bandwidth, aperture efficiency, gain bandwidth, beam forming, polarization characteristics/bandwidth (circular polarization), and the like. Therefore, the design of modern antennas faces high challenges. On the other hand, lens antennas have the advantages of rich form change, good electromagnetic properties, and the like, and the development of 3D printing technology makes it possible to develop all-dielectric lenses of complex structures, which makes all-dielectric lens antennas excellent candidates for satisfying future antenna systems. Therefore, it is very important to research the multi-objective optimization technology of the lens antenna.
Currently, most antenna designs in domestic and foreign documents are based on a traditional manual trial and error method, the electrical performance of the antenna needs to be obtained by continuously changing the structural parameters of the antenna to perform simulation (parameter scanning), and the design effects are uneven; thus, this method is time and labor consuming and inefficient. With the further enhancement of computer computing power and the development of machine learning, machine learning (including neural network, support vector regression, gaussian process regression, etc.) methods are gradually applied to the electromagnetic field to accelerate design, and particularly, neural network-based methods achieve good effects. For example, some documents adopt the method of neural network to S of antenna 11 Predicting to realize multimode resonanceDesigning an antenna; some documents predict the transmission amplitude and the transmission phase of the frequency selective surface unit by using a neural network; in addition, there is also literature that a neural network is applied to array synthesis, and array information can be obtained quickly from a directional pattern. However, the study of neural network-based lens antenna designs, particularly multi-objective lens antenna designs, is still lacking.
Disclosure of Invention
The invention provides a lens antenna multi-objective optimization method based on a priori knowledge neural network, which can realize the rapid optimization of the multi-objective lens antenna and simultaneously relieve the problem of large data volume demand of the neural network.
In order to solve the problems, the invention is realized by the following technical scheme:
the lens antenna multi-objective optimization method based on the prior knowledge neural network comprises the following steps:
step 1, collecting N pieces of simulation data (x) by utilizing a full-wave simulation mode for a given lens antenna structure i ，y i ) (ii) a Wherein x i For a simulated structural parameter vector of dimension n, y i Is composed of
A simulated electromagnetic response vector of dimensions;
step 2, carrying out normalization on N pieces of simulation data (x) according to the maximum-minimum value i ，y i ) Carrying out normalization processing to obtain N pieces of normalized simulation data
Wherein
For an n-dimensional normalized simulated structural parameter vector,
a dimension normalized simulated electromagnetic response vector;
step 3, constructing a neural network based on prior knowledge, which consists of m forward neural networks and 1 reverse neural network; the input of the forward neural network is a structural parameter, and the output is an electromagnetic response; the input of the reverse neural network is electromagnetic response, and the output is structural parameters;
step 4, utilizing N pieces of normalized simulation data
Respectively training m forward neural networks in the neural network based on the priori knowledge to obtain m trained forward neural networks; in the training process of the forward neural network corresponding to the kth electromagnetic response attribute, the normalized simulation structure parameter vector is subjected to
As input to the forward neural network, a normalized simulated electromagnetic response vector is generated
Z corresponding to electromagnetic response k The dimension is used as the output of the forward neural network, and the training error and the testing error meet the specified requirements, so that the correspondingly trained forward neural network is obtained;
Step 6, utilizing N 0 Pieces of random data (x' j ，y′ j ) Training a reverse neural network in the neural network based on the prior knowledge to obtain a trained reverse neural network; training of an inverse neural networkRandom structure parameter vector y' j As input of the reverse neural network, a random structure parameter vector x' j As the output of the reverse neural network, and the training error and the testing error reach the specified requirements, thereby obtaining the trained reverse neural network;
step 7, satisfying the requirements of the lens antenna to be designed
Inputting the dimensional electromagnetic response vector y into a trained reverse neural network to obtain an n-dimensional target structure parameter vector x required to be designed by the lens antenna to be designed;
the above i =1,2, …, N indicates the number of set simulation data; j =1,2, …, N 0 ，N 0 Is the set number of random data; n is a radical of 0 N is greater than; n is the number of structural parameter attributes of the lens antenna to be designed; z is a radical of k K =1,2, …, m, m is the number of electromagnetic response properties of the lens antenna that needs to be satisfied, which is the number of discrete points of the kth electromagnetic response property.
In the above step 2, the structure parameter vector x is determined i The formula for normalizing the tag value of the pth structure parameter attribute in (1) is as follows:
for electromagnetic response vector y i The formula for normalizing the qth tag value of the kth electromagnetic response attribute in (1) is:
in the formula,
a normalized tag value representing the p-th structural parameter attribute of the i-th simulation data,
tag value, minx, representing the pth structural parameter attribute of the ith piece of simulation data p Minimum tag value, maxx, representing the p-th structural parameter attribute in N pieces of simulation data p Representing the maximum label value of the p-th structural parameter attribute in the N pieces of simulation data, wherein p =1,2, …, N and N represent the number of the structural parameter attributes;
a qth normalized tag value representing a kth electromagnetic response property of the ith simulation data,
q tag value, miny, representing the kth electromagnetic response attribute of the ith piece of simulation data k Minimum tag value, maxy, representing the kth electromagnetic response attribute in the N pieces of simulation data k Represents the maximum label value of the k-th electromagnetic response attribute in N pieces of simulation data, q =1,2, …, z k ，z k The number of discrete points representing the k-th electromagnetic response property, k =1,2, …, m, m being the number of electromagnetic response properties of the lens antenna that need to be satisfied; i =1,2, …, N indicates the number of simulation data.
In the step 2, the simulation data is normalized
Before normalization processing, the step-shaped smoothing simplification needs to be performed by using a step smoothing method on the tag value of the electromagnetic response with the change exceeding the set threshold.
Compared with the prior art, the invention provides a lens antenna multi-objective optimization method based on a priori knowledge neural network (KBANN). The method takes INN as a main body of the algorithm, adopts a reverse neural network to carry out reverse design on the lens antenna, inputs of the reverse neural network are electromagnetic responses, outputs are structural parameters, and inversion of the antenna structural parameters with multiple performance indexes is realized. In consideration of multi-objective design requirements, a reverse neural network is difficult to obtain a good training effect by using a small amount of data, but a large amount of data needs to consume a large amount of time, so that the lens antenna can meet the requirements of a plurality of performance indexes, the invention introduces a plurality of sub-FNNs respectively used for providing prior knowledge corresponding to the plurality of performance indexes. The input of the positive neural network is a structural parameter, and the output is an electromagnetic response. Firstly, the forward neural network is easy to train, and compared with the reverse neural network, the forward neural network can obtain better training effect under less data; secondly, the forward network is usually set as a single target, has a simple structure, and can be better trained compared with a multi-target reverse network. Compared with the existing lens antenna design method, the design method provided by the invention has the following advantages: 1) The antenna has high design efficiency, and the designed antenna can realize good multi-target characteristics; 2) The KBANN model provided by the method relieves the problem of large data volume requirement of the neural network and provides a new solution. The method can realize the rapid optimization of the multi-target lens antenna, and simultaneously relieve the problem of large data volume requirement of a neural network.
Drawings
Fig. 1 is a schematic diagram of the KBANN model.
Fig. 2 is a schematic diagram of a lens antenna model, wherein (a) is a side view and (b) is a top view.
Fig. 3 is an embodied KBANN model.
FIG. 4 shows the S of the lens antenna designed in this embodiment 11 Curve line.
Fig. 5 is a gain axis ratio curve of the lens antenna designed in this embodiment.
Fig. 6 is a radiation efficiency curve of the lens antenna designed by the present embodiment.
Fig. 7 shows the radiation pattern of the lens antenna designed in this embodiment at 65GHz, (a) is xoz planar pattern, and (b) is yoz planar pattern.
Fig. 8 shows the radiation pattern of the lens antenna designed in this embodiment at 75GHz, (a) is xoz planar pattern, and (b) is yoz planar pattern.
Fig. 9 shows the radiation pattern of the lens antenna designed in this embodiment at 85GHz, (a) is xoz planar pattern, and (b) is yoz planar pattern.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to specific examples.
Neural networks used in electromagnetism include Forward Neural Networks (FNNs) and Inverse Neural Networks (INNs). Compared with FNN, INN is trained once, and a target can be input to directly obtain output parameters, so that INN is more efficient for the design problem of the lens antenna. However, the application of INN in antenna design faces an important problem that data collection is time consuming, and especially for complex mapping relationships, the data volume requirement is larger. The design of the lens antenna can be further accelerated by solving the problem, and meanwhile, technical support is provided for the application of the neural network in the field of antenna electromagnetism. The invention provides a lens antenna multi-objective optimization method which combines a reverse neural network and a forward neural network to form a neural network based on priori knowledge, and is shown in figure 1.
According to design requirements, the lens antenna of the embodiment mainly comprises two parts, namely a lens and a feed source antenna, has the working frequency of 60-90GHz, and has two functions, wherein one function is to convert linear polarized waves into circularly polarized waves, and the other function is to improve the gain of the feed source antenna. Based on the above requirements, a lens structure as shown in fig. 2 is given, which mainly comprises three parts: the bottom is provided with a cylindrical sleeve and a small square sleeve which are mainly used for reinforcing the lens and fixing the lens on the horn; the middle part is a cylinder consisting of dielectric grids (epsilon 1=2.9, tan delta = 0.01), the thickness of the dielectric grids is w, and gaps among the dielectric grids are g; the middle part is used as the most important part of the lens, is mainly used for realizing linear-circular polarization conversion and can also improve the antenna gain; the top is approximately conical in shape, mainly used to achieve high gain, and also to strengthen the lens. The feed source antenna is a horn antenna, and the model is LB-12-15-AwithWR12.
Based on a given lens antenna structure, the invention provides a lens antenna multi-objective optimization method based on a priori knowledge neural network, which comprises the following steps:
step 1, collecting N pieces of simulation data (x) by utilizing a full-wave simulation mode for a given lens antenna structure i ，y i ). Where i =1,2, …, N represents the number of simulation data; x is a radical of a fluorine atom i The method comprises the steps of (1) obtaining an n-dimensional simulation structure parameter vector, wherein n is the number of structure parameter attributes of a lens antenna to be designed; y is i Is composed of
Since each electromagnetic response attribute of the dimensional simulated electromagnetic response vector is represented by an electromagnetic response curve, each electromagnetic response curve needs to be discretized to form the simulated electromagnetic response vector, in which case z is k K =1,2, …, m, m is the number of electromagnetic response properties of the lens antenna that need to be satisfied, which is the number of discrete points of the kth electromagnetic response property.
In this embodiment, the adopted full-wave simulation mode is a CST-MATLAB joint simulation mode, that is, after the upper and lower intervals of the structural parameters are determined, the combination of the parameter values randomly generated by MATLAB software is substituted into the CST software to perform the structural modeling simulation to obtain the corresponding electromagnetic response value, and then the electromagnetic response value is extracted according to the sampling principle. The number of pieces of simulation data collected, N, was 300, the structure parameter properties included g, w, hlow, R, and h as shown in fig. 2, i.e., N =5, and the electromagnetic response properties (i.e., optimization target) included axial ratio, gain, and S 11 I.e. m =3.
Step 2, simulation data (x) is subjected to normalization principle according to maximum-minimum value i ，y i ) Carrying out normalization processing to obtain normalized simulation data
Wherein
For an n-dimensional normalized simulated structural parameter vector,
is composed of
A normalized simulated electromagnetic response vector of dimensions.
On one hand, the purpose of normalization is to distribute all variables originally belonging to different intervals in a uniform interval; on the other hand, aiming at a single variable, because one variable is usually multidimensional, the normalization operation can change the distribution interval from the original large interval to a small interval, which is beneficial to the learning of the subsequent neural network.
For the structure parameter vector x i The formula for normalization processing of the tag value of the pth structural parameter attribute in (1) is as follows:
in the formula,
a normalized tag value representing the p-th structural parameter attribute of the i-th simulation data,
tag value, minx, representing the p-th structural parameter attribute of the ith simulation data p Minimum tag value, maxx, representing the p-th structure parameter attribute in N pieces of simulation data p Representing the maximum label value of the p-th structure parameter attribute in N pieces of simulation data, wherein p =1,2, …, N and N represent the number of the structure parameter attributes; i =1,2, …, N indicates the number of simulation data.
For the electromagnetic response vector y i The formula for normalizing the qth tag value of the kth electromagnetic response property in (1) is:
in the formula,
a qth normalized tag value representing a kth electromagnetic response property of the ith simulation data,
q tag value, miny, representing the kth electromagnetic response attribute of the ith simulation data k Minimum tag value, maxy, representing the kth electromagnetic response attribute in the N simulation data k Represents the maximum label value of the k-th electromagnetic response attribute in N pieces of simulation data, q =1,2, …, z k ，z k Discrete point numbers representing the k-th electromagnetic response attribute, k =1,2, …, m, m being the number of electromagnetic response attributes of the lens antenna that need to be satisfied; i =1,2, …, N indicates the number of simulation data.
Taking into account simulation data (x) i ，y i ) The tag value of some electromagnetic responses in the simulation data (x) is changed very violently, which is not beneficial to the learning of the neural network i ，y i ) Before normalization processing, the tag values of the electromagnetic responses with severe changes (namely, the changes exceed a set threshold) need to be subjected to step-shaped smooth simplification by using a step smoothing method, so that partial targets (such as a directional diagram, | S) are solved 11 |) describe difficult problems.
And 3, constructing a neural network based on prior knowledge, which consists of 3 forward neural networks and 1 reverse neural network. The input of the forward neural network is a structural parameter, and the output is an electromagnetic response. The input of the reverse neural network is electromagnetic response, and the output is structural parameters. As shown in fig. 3.
The input of the reverse neural network is electromagnetic response, and the output is structural parameters. In this embodiment, the inverse neural network includes an input layer, an output layer, and three hidden layers. The INN input can be written as
Wherein,
representing the axial ratio in the electromagnetic response,
representing the gain in the electromagnetic response,
representing | S in the electromagnetic response 11 |，z 1 ＝z 2 ＝z 3 =31, i.e.
Is 93D. The output of INN can be written as a vector
n is the number of structural parameters. When the network model is determined, the input-to-output mapping relationship can be expressed by the following relation:
wherein,
represents the output value of the input layer neuron,
represents the output value of the hidden layer neuron,
l represents the output value of the output layer neuron. l represents the l-th layer, C l Represents the number of neurons of the l-th layer,
is the connection weight between the c-th neuron and the d-th neuron in the adjacent layer.
And
determine input and outputThe training of the neural network adjusts the two parameters to make the mapping relation as accurate as possible. f. of l (x) Representative is the activation function of the l-th layer.
The output layer activation function of INN is a variant of tanh (x):
by the above equation, the output range is limited to [0,1], consistent with the normalized data. For the activation function of the hidden layer, the trained iterative attempts are finally determined as the relu (x) function:
f(x)＝relu(x)＝max(0,x) (3)
in order to guide the training of the neural network, a loss function (measuring the difference between the predicted value and the true value) needs to be defined. The loss function is defined as the MSE function:
wherein
Is the true label for the ith sample.
Is the predicted value of the ith sample, and N is the number of samples. The neural network is trained essentially by adjusting the weights (w) and the bias (b) such that the loss value is reduced, since reducing the loss value means that the gap between the predicted value and the true value is reduced.
The input of the forward neural network is a structural parameter, and the output is an electromagnetic response. In the present embodiment, | S 11 The I-FNN comprises an input layer, an output layer and two hidden layers, the AR-FNN comprises an input layer, an output layer and one hidden layer, and the Gain-FNN comprises an input layer, an output layer and one hidden layer.
The activation function of the output layer is the same as formula (2), and the activation of the hidden layerThe living function is the same as equation (3). The three sub-FNNs have the same set of structural parameters
The output of the three sub-FNNs is
Respectively represent axial ratio, gain and | S 11 L. The loss function is the same as equation (4). The goal of the FNN is to train so that the predicted electromagnetic response is less than the true electromagnetic response.
Step 4, utilizing N pieces of normalized simulation data
Respectively training 3 forward neural networks in the neural networks based on the priori knowledge to obtain 3 trained forward neural networks; in the training process of the forward neural network corresponding to each electromagnetic response attribute: normalizing the simulated structure parameter vector
As an input to the forward neural network; the normalized simulated electromagnetic response vector
Z corresponding to electromagnetic response k Dimension is used as the output of the forward neural network (when training the forward neural network corresponding to the axial ratio, the dimension will be used as the output of the forward neural network
As an output; when training the positive neural network corresponding to the gain, the training method is to
As an output; in training | S 11 If the positive neural network corresponds to |, the network will be
As an output); and make training errors and testing errorsThe difference reaches the specified requirement, so that a correspondingly trained forward neural network is obtained.
Number of random data set N 0 Chosen according to the tuning experience of neural networks, is usually much larger than N. The specific operation is to randomly generate N by utilizing MATLAB 0 Combining the group structure parameters, inputting FNN to obtain N 0 Group output electromagnetic response, N 0 Group structure parameter combination and N 0 Group output electromagnetic response group N 0 Strip random data (x' j ，y′ j ). In the present embodiment, the number of pieces of random data N 0 Is 50000 strips.
Step 6, utilizing N 0 Strip random data (x' j ，y′ j ) Training a reverse neural network in the neural network based on the prior knowledge to obtain a trained reverse neural network; in the training process of the reverse neural network, a random structure parameter vector y' j As input of the reverse neural network, a random structure parameter vector x' j And as the output of the inverse neural network, the training error and the testing error are made to meet the specified requirements, thereby obtaining the trained inverse neural network.
Step 7, satisfying the requirements of the lens antenna to be designed
The electromagnetic response vector y of the dimension is input to the trainingAnd obtaining an n-dimensional target structure parameter vector x required to be designed by the lens antenna to be designed in the trained reverse neural network.
The design targets and INN input values of the present embodiment are shown in Table 1. Since the training data of the INN is normalized, the input target electromagnetic response of the INN also needs to be normalized. The specific input target electromagnetic response value setting takes into account the design target, the electromagnetic response of the INN, and the characteristics of the lens antenna itself. It is also worth mentioning that for INN, a set of electromagnetic response parameters is input and the output is unique. However, for a particular design goal, there are typically many sets of input electromagnetic response parameters that are satisfactory. Only one set of output parameters is required for the designer. For the input electromagnetic response of Table 1, the resulting output structural parameter is P target ＝[31.52,13.43,0.731,1.24,43.6]。
TABLE 1 comparison of design objectives, INN inputs, and CST simulation values
The structural parameters obtained by the method are subjected to structural modeling and simulation in CST to verify the effect of the method.
The processing was performed using an Objet500 Connex 3D printer, wherein the printer support material was RGD 837 veropurewite, and the processed lens was tested.
FIG. 4 shows the S of the lens antenna designed in this embodiment 11 The | curve. Simulated and tested | S 11 All | are lower than-16 dB, which shows that the lens antenna has good matching at 60-90 GHz.
Fig. 5 shows the gain axial ratio curve of the lens antenna designed by this embodiment. The gain of the horn lens antenna is improved by about 10dB compared with that of the feed source antenna, and the gain curve of the lens antenna in the whole working bandwidth (60-90 GHz) is relatively flat, so that 40% of 3-dB gain bandwidth is realized. From the results, the simulated axial ratio is lower than 2dB in 60-90GHz, and is consistent with the design target. For the test results, the test results were good for the low frequency band, while the axial ratio of 84-90GHz was 3-4.5dB. The reason for the simulation and test of the high frequency part may be the instability of the electromagnetic property of the dielectric material.
Fig. 6 is a radiation efficiency curve of the lens antenna designed by the present embodiment. The radiation efficiency curve shows that the radiation efficiency of the antenna is greater than 67% and reaches up to 83%. This indicates that the energy of the lens-antenna pair is radiated effectively.
Fig. 7-9 show the radiation patterns of the lens antenna designed by the present embodiment at different frequencies, and as a result, the antenna is a left-handed circularly polarized antenna, and the sidelobe levels are all below-20 dB.
It should be noted that, although the above-mentioned embodiments of the present invention are illustrative, the present invention is not limited thereto, and therefore, the present invention is not limited to the above-mentioned specific embodiments. Other embodiments, which can be made by those skilled in the art in light of the teachings of the present invention, are considered to be within the scope of the present invention without departing from its principles.
Claims (3)
1. The lens antenna multi-objective optimization method based on the prior knowledge neural network is characterized by comprising the following steps of:
step 1, collecting N pieces of simulation data (x) by utilizing a full-wave simulation mode for a given lens antenna structure i ，y i ) (ii) a Wherein x i For a simulated structural parameter vector of dimension n, y i Is composed of
A dimensional simulated electromagnetic response vector;
step 2, carrying out normalization on N pieces of simulation data (x) according to the maximum-minimum value i ，y i ) Carrying out normalization processing to obtain N pieces of normalized simulation data
Wherein
For an n-dimensional normalized simulated structural parameter vector,
is composed of
A normalized simulated electromagnetic response vector of dimensions;
step 3, constructing a priori knowledge-based neural network consisting of m forward neural networks and 1 reverse neural network; the input of the forward neural network is a structural parameter, and the output is an electromagnetic response; the input of the reverse neural network is electromagnetic response, and the output is structural parameters;
step 4, utilizing N pieces of normalized simulation data
Respectively training m forward neural networks in the neural network based on the priori knowledge to obtain m trained forward neural networks; in the training process of the forward neural network corresponding to the kth electromagnetic response attribute, the normalized simulation structure parameter vector is subjected to
As input to the forward neural network, a normalized simulated electromagnetic response vector is generated
Z corresponding to electromagnetic response k Dimension is used as the output of the forward neural network, and the training error and the testing error reach the specified requirements, so that the correspondingly trained forward neural network is obtained;
step 5, firstly randomly generating N 0 N-dimensional random structure parameter vector x' j (ii) a Then, a random structure parameter vector x' j Respectively inputting the data into m trained forward neural networks to respectively obtain m z k A random electromagnetic response vector of dimensions; then m are z k Electromagnetic sound of dimensionShould vector splice into
Electromagnetic response vector y 'of dimension' j (ii) a Thereby obtaining N 0 Strip random data (x' j ，y′ j )；
Step 6, utilizing N 0 Strip random data (x' j ，y′ j ) Training a reverse neural network in the neural network based on the prior knowledge to obtain a trained reverse neural network; in the training process of the reverse neural network, a random structure parameter vector y' j As input to the inverse neural network, a random structure parameter vector x' j As the output of the reverse neural network, and the training error and the testing error reach the specified requirements, thereby obtaining the trained reverse neural network;
step 7, meeting the requirements of the lens antenna to be designed
Inputting the dimensional electromagnetic response vector y into a trained reverse neural network to obtain an n-dimensional target structure parameter vector x required to be designed by the lens antenna to be designed;
the above i =1,2, …, N indicates the number of set simulation data; j =1,2, …, N 0 ，N 0 Is the set number of random data; n is a radical of 0 N is greater than; n is the number of structural parameter attributes of the lens antenna to be designed; z is a radical of k K =1,2, …, m, m is the number of electromagnetic response properties of the lens antenna that needs to be satisfied, which is the number of discrete points of the kth electromagnetic response property.
2. The multi-objective optimization method for lens antenna based on a priori knowledge neural network as claimed in claim 1, wherein in step 2,
for the structure parameter vector x i The formula for normalization processing of the tag value of the pth structural parameter attribute in (1) is as follows:
for electromagnetic response vector y i The formula for normalizing the qth tag value of the kth electromagnetic response property in (1) is:
in the formula,
a normalized tag value representing the p-th structural parameter attribute of the i-th simulation data,
tag value, minx, representing the p-th structural parameter attribute of the ith simulation data p Minimum tag value, maxx, representing the p-th structural parameter attribute in N pieces of simulation data p Representing the maximum label value of the p-th structure parameter attribute in N pieces of simulation data, wherein p =1,2, …, N and N represent the number of the structure parameter attributes;
a qth normalized tag value representing a kth electromagnetic response property of the ith simulation data,
q tag value, miny, representing the kth electromagnetic response attribute of the ith simulation data k Minimum tag value, maxy, representing the kth electromagnetic response attribute in the N pieces of simulation data k Represents the maximum label value of the k-th electromagnetic response attribute in N pieces of simulation data, q =1,2, …, z k ，z k Discrete points representing the kth electromagnetic response property, k =1,2, …, m, m being the number of electromagnetic response properties of the lens antenna that need to be satisfied; i =1,2, …, N represents the number of simulation data.
3. The multi-objective lens antenna optimization method based on the priori knowledge neural network of claim 1, wherein in the step 2, normalized simulation data are subjected to
Before normalization processing, the step-shaped smoothing simplification needs to be performed by using a step smoothing method on the tag value of the electromagnetic response with the change exceeding the set threshold.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210666376.0A CN115329655A (en) | 2022-06-13 | 2022-06-13 | Lens antenna multi-objective optimization method based on priori knowledge neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210666376.0A CN115329655A (en) | 2022-06-13 | 2022-06-13 | Lens antenna multi-objective optimization method based on priori knowledge neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115329655A true CN115329655A (en) | 2022-11-11 |
Family
ID=83916316
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210666376.0A Pending CN115329655A (en) | 2022-06-13 | 2022-06-13 | Lens antenna multi-objective optimization method based on priori knowledge neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115329655A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116911349A (en) * | 2023-09-13 | 2023-10-20 | 华南师范大学 | Optical nano antenna structure prediction network training method, prediction method and device |
-
2022
- 2022-06-13 CN CN202210666376.0A patent/CN115329655A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116911349A (en) * | 2023-09-13 | 2023-10-20 | 华南师范大学 | Optical nano antenna structure prediction network training method, prediction method and device |
| CN116911349B (en) * | 2023-09-13 | 2024-01-09 | 华南师范大学 | Optical nano antenna structure prediction network training method, prediction method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wei et al. | Fully automated design method based on reinforcement learning and surrogate modeling for antenna array decoupling | |
| CN108173578B (en) | Array antenna simulation multi-beam forming method | |
| CN115081325B (en) | Lens antenna multi-objective optimization method based on particle swarm and genetic hybrid algorithm | |
| CN109033647B (en) | Near-field sparse antenna array optimization method based on L1 norm constraint | |
| CN109063266A (en) | A kind of skin antenna electromechanics collaborative design method based on parallel Bayes optimization | |
| CN113030931B (en) | MIMO radar waveform generation method based on manifold optimization | |
| Koziel et al. | Accelerated gradient-based optimization of antenna structures using multifidelity simulations and convergence-based model management scheme | |
| CN115329655A (en) | Lens antenna multi-objective optimization method based on priori knowledge neural network | |
| Wei et al. | Automated antenna design via domain knowledge-informed reinforcement learning and imitation learning | |
| CN111709192B (en) | Planar inverted F antenna resonant frequency prediction method based on semi-supervised learning | |
| Fan et al. | Optimal design of microwave devices by fitness-estimation-based particle swarm optimization algorithm | |
| Guney et al. | Computation of resonant frequency for equilateral triangular microstrip antennas using the adaptive neuro‐fuzzy inference system | |
| Xue et al. | A novel intelligent antenna synthesis system using hybrid machine learning algorithms | |
| CN117113544A (en) | Sum and difference wave beam optimization method for array antenna with cover | |
| He et al. | A novel grey wolf optimizer and its applications in 5G frequency selection surface design | |
| Wu et al. | AI Deep Learning Optimization for Compact Dual-Polarized High-Isolation Antenna Using Backpropagation Algorithm | |
| WO2024049492A2 (en) | Tunable system designs for dynamic metamaterial | |
| Raya et al. | Design of inset fed rectangular shaped microstrip patch antenna using deep neural networks | |
| Jian et al. | Millimeter wave microstrip antenna design based on swarm intelligence algorithm in 5G | |
| Zhang et al. | Nonuniform antenna array design by parallelizing three-parent crossover genetic algorithm | |
| Gao et al. | Gradient-based optimization method for interference suppression of linear arrays by the amplitude-only and phase-only control | |
| Zheng et al. | Design of monopole antennas based on progressive Gaussian process | |
| Xu et al. | Optimal design of conical horn antenna based on GP model with coarse mesh | |
| Guney et al. | New narrow aperture dimension expressions obtained by using a differential evolution algorithm for optimum gain pyramidal horns | |
| Fan et al. | Optimal design of Yagi microstrip antenna based on particle swarm optimization with fitness estimation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |