CN110162847A - Machine learning auxiliary antenna design method based on addition feature policy - Google Patents
Machine learning auxiliary antenna design method based on addition feature policy Download PDFInfo
- Publication number
- CN110162847A CN110162847A CN201910368034.9A CN201910368034A CN110162847A CN 110162847 A CN110162847 A CN 110162847A CN 201910368034 A CN201910368034 A CN 201910368034A CN 110162847 A CN110162847 A CN 110162847A
- Authority
- CN
- China
- Prior art keywords
- machine learning
- algorithm
- addition
- value
- design method
- 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.)
- Granted
Links
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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Hardware Design (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Variable-Direction Aerials And Aerial Arrays (AREA)
Abstract
The invention discloses a kind of machine learning auxiliary antenna design methods based on addition feature policy, including (1) to construct training dataset;(2) it concentrates addition frequency to be characterized value in training data obtained in the previous step, obtains the training dataset with frequecy characteristic;(3) training dataset with frequecy characteristic is trained using machine learning algorithm, obtains agent model;(4) agent model is optimized using evolution algorithm, obtains the corresponding antenna parameter combination of optimal target prediction value;(5) the corresponding antenna parameter of optimal objective predicted value is combined and carries out full-wave simulation calculating, judged whether to reach loop termination condition according to simulation result, it continues cycling through if necessary, after updating training dataset, repeats the calculating process of step (2) to (5) until loop termination.Frequecy characteristic addition strategy of the present invention, improves the precision of prediction for the agent model that machine learning obtains and improves convergence speed of the algorithm.
Description
Technical field
The present invention relates to a kind of antenna structure design method more particularly to a kind of machine learning based on addition feature policy
Auxiliary antenna design method.
Background technique
In more than ten years in past, the optimization method based on agent model is widely used in antenna, passive device and circuit
The fields such as design, and achieve good effect.Such method can improve because the high emulation cost and member of full-wave simulation inspire
The problem of optimization overlong time caused by the multioperation number of formula algorithm.The target of optimization method based on agent model is to build
The agent model of a vertical low-cost, to predict the performance of device at the point of the possibility in design space.Machine learning is
A method of effectively establishing agent model, and traditional antenna optimization method based on machine learning there are convergence rate compared with
Slowly, the problems such as actual multiple-objection optimization can not be adapted to.
Summary of the invention
Goal of the invention: in view of the above problems, the present invention proposes that a kind of machine learning based on addition feature policy assists day
Line design method is able to ascend the precision of prediction of Antenna Design agent model, accelerates the computational efficiency of algorithm, utilizes machine learning
Obtain optimal antenna structure characteristic.
Technical solution: the technical scheme adopted by the invention is that a kind of machine learning based on addition feature policy assists day
Line design method, comprising the following steps:
(1) training dataset is constructed;The structural characteristic parameter for influencing antenna radiation characteristics is extracted, above and below parameters
Stochastical sampling in section is limited, full-wave simulation calculating is carried out using 3 D electromagnetic field emulation tool HFSS, forms training dataset.
(2) it concentrates addition frequency to be characterized value in training data obtained in the previous step, obtains the training with frequecy characteristic
Data set;0.5GHz is divided between its frequency point added.
(3) training dataset with frequecy characteristic is trained using machine learning algorithm, obtains agent model;Institute
The machine learning algorithm stated is that Gaussian process returns (GPR, Gaussian Process Regression) algorithm.
(4) optimization aim is selected, agent model is optimized using evolution algorithm, obtains optimal target prediction value pair
The antenna parameter combination answered;Wherein, optimization aim is simple target, may be configured as the reflection coefficient in bandwidth 38-47.5GHz
In -10dB hereinafter, target prediction value is the predicted value of reflection coefficient, fitness function m1Are as follows:
Wherein, M is the number of added Eigen-frequencies f,For the prediction of the reflection coefficient at i-th of Frequency point
Value, p1For low confidence Boundary algorithm LCB constant,For the standard deviation of the predicted value of the reflection coefficient at i-th of Frequency point.p1
It should suitably reduce for low confidence Boundary algorithm LCB constant, select p1Preferable calculating effect can be obtained by belonging to (0,0.2).
Optimization aim may be multiple target, optimize task at this time are as follows:
Wherein, N is the number of optimization aim, wkFor the weight constant of k-th of target, OkIt (x) is k-th of target through low
Fitness after confidence Boundary algorithm LCB prescreening,For the reference point value of k-th of target.The optimization aim can choose
The reflection coefficient being set as in bandwidth 38-47.5GHz | S11 | value is to -10dB or less and the value of gain is to 4.8dBi or more.
(5) the corresponding antenna parameter of optimal objective predicted value is combined and carries out full-wave simulation calculating, sentenced according to simulation result
It is disconnected whether to reach loop termination condition, it continues cycling through if necessary, after updating training dataset, repeats step (2) to (5)
Calculating process is until loop termination.
The utility model has the advantages that 1) addition of feature used by algorithm strategy will in the initial and each iterative process of algorithm
Data set of the data set of feature as machine learning is added, with the precision of prediction for the agent model that hoisting machine study obtains;
2) algorithm implements complete evolution algorithm process, the optimum point of prediction is obtained, compared to tradition in primary iterative process
Method, can search out possible optimum point in each iteration to more maximum probability;3) algorithm multiple target application in,
Using the prescreening method based on reference point, actually required optimization aim, the practicability of boosting algorithm can be absorbed in.
Detailed description of the invention
Fig. 1 is the flow chart of auxiliary antenna design method of the present invention;
Fig. 2 is the antenna structure view for verifying algorithm proposed by the present invention;
Fig. 3 is using the single object optimization comparative result figure before and after adding feature policy;
Fig. 4 is the predicted value and simulation value of resulting antenna after single goal and multi-objective optimization algorithm optimization of the present invention
And the measured value for the antenna processed under single object optimization size;
Fig. 5 is the picture in kind for the antenna processed under single object optimization size.
Specific embodiment
Further description of the technical solution of the present invention with reference to the accompanying drawings and examples.
The present embodiment illustrates the practicability and superiority of proposed algorithm with true antenna structure example.Such as Fig. 2
It is shown, a kind of schematic diagram of conventional back chamber gap broadband antenna structure based on substrate integration wave-guide is given, the day is influenced
The important feature parameter of cable architecture radiance is as shown in table 1, provides the initial value of 9 antenna parameters to be optimized altogether and provides
Upper and lower bound when optimization.
Table 1
Separately below to the optimization task of single goal and multiple target, specific embodiment is stated.Based on above-mentioned antenna structure,
The flow chart of machine learning auxiliary antenna optimization method of the present invention based on addition feature policy is as shown in Figure 1, be more
The iterative optimization procedure of step, comprising the following steps:
(1) training dataset is constructed.
The characteristic parameter for influencing antenna radiation characteristics is extracted, the stochastical sampling in the bound section of parameters utilizes
HFSS carries out full-wave simulation calculating, forms training dataset, by taking 9 antenna parameters to be optimized gone out given in table 1 as an example, benefit
With the section as defined in the bound of parameter to be optimized Latin Hypercube Sampling (LHS, Latin Hypercube Sampling)
Interior stochastical sampling chooses 10 groups of numerical value, carries out full-wave simulation calculating, resulting simulation result and antenna parameter phase group using HFSS
It closes, forms training dataset, so far complete the process of initialization training set.
(2) it concentrates addition frequency to be characterized value in training data obtained in the previous step, obtains the training with frequecy characteristic
Data set.It is added characteristic value by frequency selection purposes in this example, the data after converting raw data set to addition characteristic value
Collection, frequency point interval select 0.5GHz.
(3) training dataset with frequecy characteristic is trained using machine learning algorithm, obtains agent model.Tool
Body, model is trained using GPR, the independent variable of the dimensional parameters and frequency composition of establishing antenna is corresponding
Connection between antenna performance, that is, dependent variable.The step in obtained agent model be cheap agent model.
(4) optimization aim is selected, using acquired cheap agent model, using evolution algorithm to the dimensional parameters of antenna
Optimization obtains optimum prediction value and its combination of corresponding antenna parameter.
Using single object optimization method, it is anti-in optimization design bandwidth 38-47.5GHz that the optimization task of single goal, which is arranged,
Penetrate coefficient | S11 | value to -10dB hereinafter, optimization aim herein can be selected according to design object, as optimized in bandwidth
Antenna gain, axis is than different antenna performances such as, directional diagrams.Its fitness function is selected as through too low confidence boundary method
(LCB, Lower Confidence Bounding) pretreated functional value, with combine provided by agent model it is anti-
Penetrate the mean value and potentiality of coefficient predictors.Herein due to the increase of addition feature bring prediction variance, the constant to LCB is needed
It is corresponding to reduce, therefore select p1=(0,0.2) selects the constant p of LCB in this example1It is 0.1.Fitness function m herein1It can
Selection are as follows:
Wherein, M is the number of added Eigen-frequencies f,For the reflection coefficient at i-th of Frequency point | S11 |
Predicted value, p1For LCB constant,For the reflection coefficient at i-th of Frequency point | S11 | predicted value standard deviation.
(5) optimum prediction is worth corresponding antenna parameter combination progress full-wave simulation calculating, is judged whether according to calculated result
Reach loop termination condition, i.e., whether meets set optimization task, be herein anti-in optimization design bandwidth 38-47.5GHz
Penetrate coefficient | S11 | value to -10dB or less.
After antenna parameter combination and predicted value at the optimum point for obtaining prediction, emulation meter is carried out to it using HFSS
Calculate, and by emulation obtain antenna performance (be herein the reflection coefficient at each frequency point | S11 | value) combined with antenna parameter, it is right
Training dataset is updated.Judge whether updated training dataset reaches termination condition, decides whether to continue cycling through.Such as
Fruit needs to continue cycling through the process for then repeating step (2) to (5).
Utilize the optimum results of single object optimization algorithm:
Fig. 3 gives in the Comparative result using the single object optimization before and after addition feature policy, it can be seen that is adopting
After addition feature policy, algorithm has just reached goal-selling merely through 22 iteration, far faster than not using addition feature policy
Algorithm.Table 2 gives the time-consuming comparison of two kinds of distinct methods, after the single object optimization using addition feature policy, algorithm
The fitness that can reach -10.1 for time-consuming 2.5 hours, meets design object;And in the single object optimization of non-addition feature policy
Under method, algorithm is 7.7 hours time-consuming, undergoes 100 iterative process to also fail to reach design object, terminates at -9.8 adaptation
Degree.Fig. 4 gives the predicted value and simulation value of the resulting antenna after single goal algorithm optimization, excellent using single object optimization algorithm
The antenna size dissolved is processed resulting antenna picture and is presented in Fig. 5, and the left side and the right are to process obtained antenna respectively
Front and back, actual measurement | S11 | data are also presented in Fig. 4.
Table 2
In step (4), agent model is optimized using evolution algorithm, the calculation of multiple-objection optimization can also be used
Method.The optimization task of multiple target is set for the reflection coefficient in optimization design bandwidth 38-47.5GHz | S11| value to -10dB with
Under and gain value to 4.8dBi or more.Reference point, which is arranged, isThe optimization task of multiple target can be converted
Are as follows:
Wherein, N is the number of target, is herein 2;wkFor the weight constant of k-th of target, OkIt (x) is k-th of target
Fitness (the constant p of selection LCB after LCB prescreening1For 0.1), i.e. fitness function defined in formula (1),For
The reference point value of k-th of target.
Utilize the optimum results of multi-objective optimization algorithm:
Table 3 gives in the Comparative result using the multiple-objection optimization before and after addition feature policy, special using addition
After the multiple-objection optimization for levying strategy, design object can reach within algorithm time-consuming 1.8 hours, and in the monocular of non-addition feature policy
It marks under optimization method, algorithm time-consuming can be only achieved design object in 4.6 hours.Fig. 4 gives through single goal and multiple-objection optimization calculation
The predicted value and simulation value of resulting antenna, the antenna size gone out using single object optimization algorithm optimization process institute after method optimization
The antenna picture obtained is presented in Fig. 5, actual measurement | S11 | data are also presented in Fig. 4.
Table 3
Claims (8)
1. a kind of machine learning auxiliary antenna design method based on addition feature policy, which comprises the following steps:
(1) structural characteristic parameter for influencing antenna radiation characteristics is extracted, training dataset is constructed;
(2) it concentrates addition frequency to be characterized value in training data obtained in the previous step, obtains the training data with frequecy characteristic
Collection;
(3) training dataset with frequecy characteristic is trained using machine learning algorithm, obtains agent model;
(4) optimization aim is selected, agent model is optimized using evolution algorithm, it is corresponding to obtain optimal target prediction value
Antenna parameter combination;
(5) the corresponding antenna parameter of optimal objective predicted value is combined and carries out full-wave simulation calculating, be according to simulation result judgement
It is no to reach loop termination condition, it continues cycling through if necessary, after updating training dataset, repeats the calculating of step (2) to (5)
Process is until loop termination.
2. the machine learning auxiliary antenna design method according to claim 1 based on addition feature policy, feature exist
In building training dataset described in step 1 includes following procedure: in the bound section of each structural characteristic parameter
Interior stochastical sampling, composition characteristic parameter sets simultaneously carry out full-wave simulation calculating, calculated result and corresponding structural characteristic parameter
Collectively constitute training dataset.
3. the machine learning auxiliary antenna design method according to claim 1 based on addition feature policy, feature exist
0.5GHz is divided between, addition frequecy characteristic described in step 2, the frequency point of addition.
4. the machine learning auxiliary antenna design method according to claim 1 based on addition feature policy, feature exist
In: machine learning algorithm described in step 3 is Gaussian process regression algorithm.
5. the machine learning auxiliary antenna design method according to claim 1 based on addition feature policy, feature exist
In: agent model is optimized using evolution algorithm described in step 4, optimization aim is set as simple target, evolves
Fitness function m in algorithm1Are as follows:
Wherein, M is the number of added Eigen-frequencies,For the optimization aim predicted value at i-th of Frequency point, p1For
Low confidence Boundary algorithm LCB constant,For the standard deviation of the optimization aim predicted value at i-th of Frequency point;The low confidence
Boundary algorithm LCB constant p1Belong to (0,0.2).
6. the machine learning auxiliary antenna design method according to claim 5 based on addition feature policy, feature exist
In: the optimization aim is set as the reflection coefficient in bandwidth 38-47.5GHz in -10dB hereinafter, the optimization aim predicted value
For the predicted value of reflection coefficient.
7. the machine learning auxiliary antenna design method according to claim 1 based on addition feature policy, feature exist
In: agent model is optimized using evolution algorithm described in step 4, optimization aim is set as multiple target, and optimization is appointed
Business are as follows:
Wherein, N is the number of optimization aim, wkFor the weight constant of k-th of target, OkIt (x) is k-th of target through low confidence
Fitness after Boundary algorithm LCB prescreening,For the reference point value of k-th of target.
8. the machine learning auxiliary antenna design method according to claim 7 based on addition feature policy, feature exist
In: the optimization aim is set as the reflection coefficient in bandwidth 38-47.5GHz | S11 | value to -10dB or less and the value of gain
To 4.8dBi or more.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910368034.9A CN110162847B (en) | 2019-05-05 | 2019-05-05 | Machine learning auxiliary antenna design method based on feature adding strategy |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910368034.9A CN110162847B (en) | 2019-05-05 | 2019-05-05 | Machine learning auxiliary antenna design method based on feature adding strategy |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110162847A true CN110162847A (en) | 2019-08-23 |
CN110162847B CN110162847B (en) | 2023-08-25 |
Family
ID=67633375
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910368034.9A Active CN110162847B (en) | 2019-05-05 | 2019-05-05 | Machine learning auxiliary antenna design method based on feature adding strategy |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110162847B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110728034A (en) * | 2019-09-24 | 2020-01-24 | 东南大学 | Antenna rapid multi-target modeling method using multistage cooperative machine learning |
CN111159935A (en) * | 2019-12-11 | 2020-05-15 | 同济大学 | BP neural network parameter calibration method based on LHS |
CN111159915A (en) * | 2020-01-03 | 2020-05-15 | 山东天岳电子科技有限公司 | Parameter optimization method and device for device design |
CN111985150A (en) * | 2020-07-06 | 2020-11-24 | 东南大学 | Multilayer electronic device robustness optimization design method using machine learning auxiliary optimization |
CN111985064A (en) * | 2020-08-31 | 2020-11-24 | 华中科技大学 | Agent-assisted optimization design method and system for permanent magnet motor |
CN113050086A (en) * | 2021-06-01 | 2021-06-29 | 中国南方电网有限责任公司超高压输电公司广州局 | Ground penetrating radar system, control method, device, equipment and storage medium |
CN113609677A (en) * | 2021-08-05 | 2021-11-05 | 东南大学 | Multipath-based machine learning auxiliary antenna design method |
CN113889737A (en) * | 2021-09-30 | 2022-01-04 | 西华大学 | Substrate integrated waveguide parameter optimization method and structure based on reinforcement learning |
Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106897511A (en) * | 2017-02-17 | 2017-06-27 | 江苏科技大学 | Annulus tie Microstrip Antenna Forecasting Methodology |
-
2019
- 2019-05-05 CN CN201910368034.9A patent/CN110162847B/en active Active
Patent Citations (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106897511A (en) * | 2017-02-17 | 2017-06-27 | 江苏科技大学 | Annulus tie Microstrip Antenna Forecasting Methodology |
Non-Patent Citations (1)
Title |
---|
许永秀等: "基于两阶段高斯过程的天线建模方法", 《计算机应用技术》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110728034A (en) * | 2019-09-24 | 2020-01-24 | 东南大学 | Antenna rapid multi-target modeling method using multistage cooperative machine learning |
CN110728034B (en) * | 2019-09-24 | 2022-11-08 | 东南大学 | Antenna rapid multi-target modeling method using multistage cooperative machine learning |
CN111159935A (en) * | 2019-12-11 | 2020-05-15 | 同济大学 | BP neural network parameter calibration method based on LHS |
CN111159915A (en) * | 2020-01-03 | 2020-05-15 | 山东天岳电子科技有限公司 | Parameter optimization method and device for device design |
CN111985150A (en) * | 2020-07-06 | 2020-11-24 | 东南大学 | Multilayer electronic device robustness optimization design method using machine learning auxiliary optimization |
CN111985064A (en) * | 2020-08-31 | 2020-11-24 | 华中科技大学 | Agent-assisted optimization design method and system for permanent magnet motor |
CN113050086A (en) * | 2021-06-01 | 2021-06-29 | 中国南方电网有限责任公司超高压输电公司广州局 | Ground penetrating radar system, control method, device, equipment and storage medium |
CN113609677A (en) * | 2021-08-05 | 2021-11-05 | 东南大学 | Multipath-based machine learning auxiliary antenna design method |
CN113609677B (en) * | 2021-08-05 | 2024-01-23 | 东南大学 | Multipath-based machine learning auxiliary antenna design method |
CN113889737A (en) * | 2021-09-30 | 2022-01-04 | 西华大学 | Substrate integrated waveguide parameter optimization method and structure based on reinforcement learning |
CN113889737B (en) * | 2021-09-30 | 2022-04-08 | 西华大学 | Substrate integrated waveguide parameter optimization method and structure based on reinforcement learning |
Also Published As
Publication number | Publication date |
---|---|
CN110162847B (en) | 2023-08-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110162847A (en) | Machine learning auxiliary antenna design method based on addition feature policy | |
CN106054123B (en) | A kind of sparse L battle arrays and its arrival direction estimation method | |
CN107333238B (en) | Indoor fingerprint rapid positioning method based on support vector regression | |
CN104615854B (en) | A kind of beam-broadening and side lobe suppression method based on sparse constraint | |
CN110166154B (en) | Software radio frequency spectrum monitoring and identifying method based on neural network | |
CN106680780A (en) | Radar optimal waveform design method based on radio frequency stealth in frequency spectrum shared environment | |
CN108984985B (en) | Antenna structure design method based on neural network | |
CN110824415A (en) | Sparse direction-of-arrival angle estimation method based on multiple-sending and multiple-receiving array | |
CN104023395B (en) | The scattering signal source locating method changed in extensive MIMO based on beam space | |
CN102395194B (en) | ANFIS (Adaptive Neural Fuzzy Inference System) indoor positioning method based on improved GA(Genetic Algorithm) optimization in WLAN (Wireless Local Area Network) environment | |
CN110232212B (en) | Antenna design optimization method and system based on bat algorithm | |
CN108667502A (en) | A kind of spatial modulation antenna selecting method based on machine learning | |
CN108173578B (en) | Array antenna simulation multi-beam forming method | |
CN109063266A (en) | A kind of skin antenna electromechanics collaborative design method based on parallel Bayes optimization | |
CN109061578B (en) | Notch directional diagram waveform comprehensive design method based on MIMO radar | |
CN109189776A (en) | A kind of Method of Data with Adding Windows | |
CN110147590B (en) | Spiral antenna design method based on adaptive evolution optimization algorithm | |
WO2022001148A1 (en) | Non-cooperative game-based radar communication integrated networking system power control method | |
CN108983142A (en) | A kind of near field sources direction-finding method based on quantum whale Optimization Mechanism | |
CN106650171A (en) | Transformer partial discharge ultrahigh frequency detection Hilbert fractal antenna optimization method | |
CN109946673B (en) | Networked radar node selection method based on Monte Carlo tree search | |
CN114204971B (en) | Iterative aggregate beam forming design and user equipment selection method | |
CN111651936A (en) | FOA-GRNN-based modeling design method for ultra-wideband antenna with double trap characteristics | |
CN105093200B (en) | Target Wave arrival direction estimating method outside a kind of grid based on amendment dictionary | |
CN112149360A (en) | ALO-LMBP neural network-based double-notch characteristic ultra-wideband antenna reverse modeling research method |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |