CN112257373A - Snake-shaped PCB antenna return loss prediction method based on three-body training algorithm - Google Patents

Snake-shaped PCB antenna return loss prediction method based on three-body training algorithm Download PDF

Info

Publication number
CN112257373A
CN112257373A CN202011267345.5A CN202011267345A CN112257373A CN 112257373 A CN112257373 A CN 112257373A CN 202011267345 A CN202011267345 A CN 202011267345A CN 112257373 A CN112257373 A CN 112257373A
Authority
CN
China
Prior art keywords
pcb antenna
model
training
serpentine
return loss
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
Application number
CN202011267345.5A
Other languages
Chinese (zh)
Other versions
CN112257373B (en
Inventor
高婧
田雨波
陈学志
张天亮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jiangsu University of Science and Technology
Original Assignee
Jiangsu University of Science and Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Jiangsu University of Science and Technology filed Critical Jiangsu University of Science and Technology
Priority to CN202011267345.5A priority Critical patent/CN112257373B/en
Publication of CN112257373A publication Critical patent/CN112257373A/en
Application granted granted Critical
Publication of CN112257373B publication Critical patent/CN112257373B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/30—Circuit design
    • G06F30/39—Circuit design at the physical level
    • 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
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
    • Y02E60/00—Enabling technologies; Technologies with a potential or indirect contribution to GHG emissions mitigation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computer Hardware Design (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Geometry (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a snake-shaped PCB antenna return loss prediction method based on a three-body training algorithm, which comprises the steps of establishing three prediction models based on the mapping relation between snake-shaped PCB antenna size parameters and resonant frequency, selecting the optimal prediction model by adopting the three-body training algorithm, expanding training set samples, and finally training the optimal prediction model by utilizing the expanded training set to obtain a trained snake-shaped PCB antenna return loss prediction model; and predicting the return loss sampling point of the snake-shaped PCB antenna with the input size parameters by using the model. According to the method, a prediction model with high precision can be trained under the condition that a small number of marked samples exist, and then the return loss of the snake-shaped PCB antenna is predicted.

Description

Snake-shaped PCB antenna return loss prediction method based on three-body training algorithm
Technical Field
The invention belongs to the technical field of serpentine PCB antenna design and optimization, and particularly relates to a method for predicting return loss of a serpentine PCB antenna according to size parameters of the serpentine PCB antenna.
Background
In the modeling and optimization design process of microwave devices, electromagnetic software such as hfss (high Frequency Structure simulator), cst (computer Simulation technology) and the like is generally used to simulate and analyze the performance of the devices. The electromagnetic simulation software needs to perform complex calculation during analysis and solution, and the time consumption is long. Particularly, when the structure of the microwave device is complicated, the conventional method using electromagnetic software consumes much labor and time. With the rise of Machine Learning, rapid computer aided design methods, such as Gaussian Process (GP), Support Vector Machine (SVM), Kernel Extreme Learning Machine (KELM), etc., have been developed, and these algorithms are used to assist in designing microwave devices.
GP is gradually developed on the basis of Bayesian neural network research, and the GP has good adaptability to the problems of small sample processing, high dimension, nonlinearity and the like. The SVM is a machine learning method proposed according to a statistical learning theory, and is suitable for solving the problems of nonlinearity, small samples, high dimensionality and the like. The ELM (Extreme Learning Machine, ELM) is a simple and effective single hidden layer feedforward neural network Learning algorithm, a kernel function is introduced into the ELM to form a kernel limit Learning Machine (KELM) with the least square optimal solution, and the method has the advantages of few adjustable parameters, high convergence speed and good generalization performance. The three alternative models have certain research results in the field of antenna optimization, and feasibility is provided for the application of the three alternative models in the field of electromagnetic devices.
The traditional machine learning technology relies on a large number of labeled samples for training, the existing modeling for electromagnetic behavior is based on a supervised learning mode, and the labeled training samples used for modeling are based on simulation software such as HFSS. Since it is time consuming for the simulation software to mark one sample, it also takes a long time to construct the training set.
Disclosure of Invention
The purpose of the invention is as follows: the invention discloses a snake-shaped PCB antenna return loss prediction method based on a three-body training algorithm, which can train a prediction model with higher precision under the condition of a small number of marked samples, and further train and predict the snake-shaped PCB antenna return loss.
The technical scheme is as follows: the invention adopts the following technical scheme:
the snake-shaped PCB antenna return loss prediction method based on the three-body training algorithm comprises a training stage and a prediction stage, wherein the training stage comprises the following steps:
step 1, obtaining labeled samples, wherein each labeled sample is represented as (X, Y, Z); wherein X is a p-dimensional vector and represents p size parameters of the serpentine PCB antenna; y is the resonant frequency of the serpentine PCB antenna with the size of X; z is a q-dimensional vector representing a serpentine PCB antenna S of size X11Q sample point values on the curve (return loss);
dividing the marked samples into a training set and a test set, and constructing an unmarked sample set U, wherein each unmarked sample is a p-dimensional vector representing p size parameters of the snake-shaped PCB antenna;
step 2, constructing three serpentine PCB antenna return loss prediction models which are respectively recorded as M1, M2 and M3; the input of each snake-shaped PCB antenna return loss prediction model is a p-dimensional vector representing snake-shaped PCB antenna size parameters, and the output is the resonant frequency of the snake-shaped PCB antenna;
step 3, training the return loss prediction models of the three snake-shaped PCB antennas by adopting a training set to obtain three initial models which are respectively recorded as M10,M20,M30(ii) a Using test set pair M10,M20,M30Respectively testing to obtain initial testing errors of each initial model, and respectively recording the initial testing errors as e10,e20,e30(ii) a Let the initial best model MI0Is and min (e 1)0,e20,e30) A corresponding model; the iteration time t is 0;
step 4, selecting num samples from the unlabeled sample set U
Figure BDA0002776571220000021
Three prediction models M1 respectively input to the current iterationt,M2t,M3tIn, the output of the model is respectively
Figure BDA0002776571220000022
Forming three sets of pseudo-label data with num samples
Figure BDA0002776571220000023
n＝1,2,…,num；
Step 5, three pseudo label data set pairs M1 are adoptedt,M2t,M3tPerforming cross training to obtain 6 cross training models
Figure BDA00027765712200000211
Figure BDA00027765712200000212
Indicating the use of the jth pseudo-mark data set
Figure BDA0002776571220000024
For the ith prediction model MitCarrying out a trained cross training model; i, j belongs to {1,2,3}, and i is not equal to j;
step 6, selecting num from the test settestOne sample to 6 cross training models
Figure BDA0002776571220000025
Performing test, and calculating test error of each model
Figure BDA0002776571220000026
Minimize the test error
Figure BDA0002776571220000027
The corresponding cross training model is
Figure BDA0002776571220000028
Will be provided with
Figure BDA0002776571220000029
Corresponding pseudo-marked data set
Figure BDA00027765712200000210
And numtestAdding each test sample into a training set, and adding the numtestIndividual test samples are deleted from the test set;
if it is not
Figure BDA0002776571220000031
Figure BDA0002776571220000032
For the best model, the iteration is ended, and the process jumps to step 9, where ErrorthIs a preset error threshold;
the iteration time t is increased by one;
step 7, training M1, M2 and M3 by using the updated training set to obtain three updated models which are respectively marked as M1t,M2t,M3t(ii) a Using the updated test set pair M1t,M2t,M3tRespectively testing to obtain the current iteration test error of each model, and respectively recording as e1t,e2t,e3t(ii) a Let the current iteration best model MItIs and min (e 1)t,e2t,e3t) A corresponding model;
and 8, if the minimum test error of the current iteration and the minimum test error of the last iteration meet the conditions:
min(e1t,e2t,e3t)>min(e1t-1,e2t-1,e3t-1) And min (e 1)t-1,e2t-1,e3t-1)<Errorth
Then MIt-1The iteration is finished for the optimal model, and the step 9 is skipped, otherwise, the step 4 is skipped for the next iteration;
step 9, modifying the output of the optimal model selected by iteration into a q-dimensional vector, training the modified optimal model by adopting an updated training set, inputting the p-dimensional vector X in the training sample, and outputting the q-dimensional vector Z in the training sample to obtain a trained snake-shaped PCB antenna return loss prediction model;
the prediction phase comprises:
inputting p size parameters of the serpentine PCB antenna to be predicted into a trained serpentine PCB antenna return loss prediction model, wherein the output q-dimensional vector is predicted S11Q sample point values on the curve.
In the step 1, p size parameters of the serpentine PCB antenna in the marked sample and the unmarked sample are determined by adopting an orthogonal test.
In the step 1, HFSS analysis software is adopted to carry out simulation analysis on the snakelike PCB antenna with the size parameter of X to obtain S11Curve line.
The three serpentine PCB antenna return loss prediction models are respectively as follows: the prediction model based on Gaussian process regression, the prediction model based on support vector regression and the prediction model based on kernel limit learning machine.
The snakelike PCB antenna has 6 size parameters which are respectively the length L of the tail end of the resonant branch1A first coupling distance L2A first coupling length L3A second coupling length L4A second coupling distance L5Length of short circuit branch6。
Based on the Gaussian process, Gaussian kernel functions are adopted in the prediction models of the support vector machine and the kernel extreme learning machine.
Has the advantages that: the snake-shaped PCB antenna return loss prediction method disclosed by the invention has the following advantages:
(1) the existing three-body training method is improved, 3 different models are trained in a crossed mode by using the same unlabeled sample, and the 3 models are updated by using pseudo-labeled data with high accuracy and corresponding test data until satisfactory accuracy is achieved. In the training process, only a small number of labeled samples are utilized, so that the frequency of calling HFSS to obtain accurate labeled data is reduced, and the labor and the time are saved;
(2) an iteration condition is set, the number of introduced unmarked data is controlled, and excessive accumulation of noise caused by excessive introduction of unmarked data is prevented, and the accuracy of the model is reduced; meanwhile, ideal model precision can be effectively achieved in a short time.
(3) And after the iterative training is finished, obtaining an optimal prediction model, expanding a training set, retraining the prediction model again, predicting the performance of the snakelike PCB antenna by using the trained prediction model, and accurately fitting the performance index of the designed antenna.
Drawings
FIG. 1 is a flow chart of a serpentine PCB antenna return loss prediction method disclosed in the present invention;
FIG. 2 is a schematic diagram of a serpentine PCB antenna structure;
FIG. 3 is a three-dimensional perspective view of a serpentine PCB antenna in HFSS;
FIG. 4 is a test error graph of an iterative process in an embodiment;
FIG. 5 is a comparison graph of the return loss curve obtained by the prediction method disclosed in the present invention and the return loss curve obtained by HFSS simulation.
Detailed Description
The invention is further elucidated with reference to the drawings and the detailed description.
The snake-shaped PCB antenna return loss prediction method based on the three-body training algorithm, as shown in FIG. 1, comprises a training stage and a prediction stage, wherein the training stage comprises:
step 1, obtaining labeled samples, wherein each labeled sample is represented as (X, Y, Z); wherein X is a p-dimensional vector and represents p size parameters of the serpentine PCB antenna; y is the resonant frequency of the serpentine PCB antenna with the size of X; z is a q-dimensional vector representing a serpentine PCB antenna S of size X11Q sample point values on the curve;
in this embodiment, the method disclosed in the present invention is specifically described by taking the serpentine microstrip antenna shown in fig. 2 as an example. The structure of the snake-shaped microstrip antenna is derived from an inverted-F antenna, integrates the advantages of the inverted-F antenna structure, bends the resonance branches, realizes the miniaturization of the antenna, and is widely applied to personal interface equipment such as a wireless mouse, a keyboard and the like. Fig. 2 is a schematic structural diagram of a serpentine PCB antenna, and fig. 3 is a three-dimensional perspective view of the serpentine PCB antenna in HFSS. Considering that the influence of the branch length on the microstrip antenna is large, the number of the size parameters of the serpentine PCB antenna selected by the embodiment is 6, and the size parameters are respectively the length L of the tail end of the resonant branch1A first coupling distance L2A first coupling length L3A second coupling length L4A second coupling distance L5Length of short circuit branch6As shown in fig. 2. The antenna fixing parameters are shown in table 1.
TABLE 1
Name of variable Variable value/mm Name of variable Variable value/mm
W1 0.90 D3 0.30
W2 0.50 D4 0.50
D1 0.50 D5 1.40
D2 0.30 D6 1.70
In Table 1, W1Width of transmission line, W, for serpentine PCB antenna short-circuited stubs2For the width of the transmission line of the resonant stub of the antenna, D1，D2，D3Respectively the distance between the antenna and the edge of the PCB; d4The length of the antenna short circuit branch is superposed with the upper grounding plate of the PCB; d5For the distance of the antenna feed branch from the short-circuit stub, D6The distance between the antenna feed branch and the resonant stub.
The value ranges and step values of the 6 size parameters of the serpentine PCB antenna are determined, and as shown in table 2, the value ranges and step values of the 6 size parameters of the serpentine PCB antenna in the labeled sample and the unlabeled sample determined in this embodiment are determined.
TABLE 2
Figure BDA0002776571220000051
And finally determining that 32 size parameters of the serpentine PCB antenna in the marked sample and the unmarked sample exist by adopting a partial orthogonal test according to the value range and the step value of the size parameters in the table 2. 22 of them were selected as labeled samples and the other 10 were unlabeled samples. Respectively carrying out simulation analysis on the 22 antennas marked with the samples by using HFSS analysis software to obtain corresponding S11Curve line. Calculate each S11Resonant frequency point of the curve and for the obtained S11The curve is sampled, every S in this example11The number of curve sampling points is 301, namely Z is a 301-dimensional vector. Thus, 22 labeled samples are obtained, and the labeled samples are divided into a training set and a test set, where the number of samples in the training set is 12 and the number of samples in the test set is 10 in this embodiment. The other 10 unlabeled samples constitute an unlabeled sample set U, each unlabeled sample being a 6-dimensional vector representing 6 dimensional parameters of the serpentine PCB antenna.
Step 2, constructing three serpentine PCB antenna return loss prediction models which are respectively recorded as M1, M2 and M3; the input of each snake-shaped PCB antenna return loss prediction model is a p-dimensional vector representing snake-shaped PCB antenna size parameters, and the output is the resonant frequency of the snake-shaped PCB antenna;
in this embodiment, the return loss prediction models of the three serpentine PCB antennas are respectively: a prediction model based on Gaussian Process Regression (GPR), a prediction model based on support vector Regression, a prediction model based on a kernel-limit learning machine.
(1) Prediction model based on Gaussian process regression
The GP essentially describes the covariance of the predicted data by the covariance of the input data. Gaussian Process Regression (GPR) is a Process of inferring the functional relationship between the input vector and the target output and finally obtaining the conditional distribution of the target output. The properties of the mean function and covariance function of the GP are determined by a set of hyper-parameters, and the maximum likelihood function can be used to find the optimal hyper-parameter. And solving the partial derivative of the hyper-parameter by establishing a log-likelihood function of the conditional probability of the training sample, and then searching the optimal solution of the hyper-parameter by adopting a conjugate gradient optimization method. After the optimal hyper-parameter is obtained, the prediction can be carried out by utilizing the trained GP. Given a new input, the maximum possible predicted a posteriori distribution of the output is inferred.
(2) Support vector regression-based prediction model
The SVM seeks the best compromise between the complexity of sample information and learning ability according to a VC (virtual channel) dimension theory of statistics and a structure risk minimum principle, and obtains the best generalization ability. For the linear inseparable condition, a nonlinear mapping algorithm is used, namely, a kernel function is introduced, the linear inseparable samples in the low-dimensional space are converted into linear separable samples in the high-dimensional space, and the SVM has specific advantages in solving small samples, nonlinearity and high-dimensional modes. Support Vector Regression (SVR) applies SVM to machine learning such as function fitting, and after training of a training set, SVR can obtain input target output.
(3) Prediction model based on kernel extreme learning machine
An Extreme Learning Machine (ELM) is an effective single hidden layer feedforward neural network Learning algorithm, and a unique optimal solution can be generated only by setting the number of hidden layer nodes of a network. KeLM introduces kernel function based on ELM, and solves the problem of random initialization of ELM algorithm. The KELM forms an extreme learning machine with a least square optimal solution, and has the advantages of less adjustable parameters, high convergence speed and good generalization performance. The kernel function constructs a mapping relation between a sample original space and a high-dimensional feature space, so that the sample can be linearly divided in the high-dimensional space, and a linear kernel, a Gaussian kernel, a Sigmoid kernel and the like are generally adopted. In this embodiment, a gaussian kernel function is used.
Step 3, training the return loss prediction models of the three snake-shaped PCB antennas by adopting a training set to obtain three initial models which are respectively recorded as M10,M20,M30(ii) a Using test set pair M10,M20,M30Respectively testing to obtain initial testing errors of each initial model, and respectively recording the initial testing errors as e10,e20,e30(ii) a Let the initial best model MI0Is and min (e 1)0,e20,e30) A corresponding model; the iteration time t is 0;
in this embodiment, the average Relative Error (MRE) of the whole test set is used as the test Error, and is calculated as follows:
Figure BDA0002776571220000071
where N is the total number of samples in the test set, Ytest(n) is the resonant frequency of the nth sample in the test set, Ypred(n) is the resonant frequency of the nth test sample predicted by the predictive model.
As shown in table 3, is the initial test error.
TABLE 3
Model (model) GP SVM KELM
Initial error 0.1352 0.1323 0.1328
Because the number of samples in the training set in the initial stage is small, the prediction precision of the initial model obtained in the step is not high, and the initial test error is large.
Step 4, selecting num samples from the unlabeled sample set U
Figure BDA0002776571220000072
Three prediction models M1 respectively input to the current iterationt,M2t,M3tIn, the output of the model is respectively
Figure BDA0002776571220000081
Forming three sets of pseudo-label data with num samples
Figure BDA0002776571220000082
n is 1,2, …, num; in this example, num is 1, i.e., 1 unlabeled sample is selected at a time.
Step 5, three pseudo label data set pairs M1 are adoptedt,M2t,M3tPerforming cross training to obtain 6 cross training models
Figure BDA0002776571220000083
Indicating the use of the jth pseudo-mark data set
Figure BDA0002776571220000084
For the ith prediction model MitCarrying out a trained cross training model; i, j belongs to {1,2,3}, and i is not equal to j;
step 6, selecting num from the test settestOne sample to 6 cross training models
Figure BDA0002776571220000085
Performing test, and calculating test error of each model
Figure BDA0002776571220000086
Minimize the test error
Figure BDA0002776571220000087
The corresponding cross training model is
Figure BDA0002776571220000088
Will be provided with
Figure BDA0002776571220000089
Corresponding pseudo-marked data set
Figure BDA00027765712200000810
And numtestAdding each test sample into a training set, and adding the numtestIndividual test samples are deleted from the test set; in this example, numtest＝1。
If it is not
Figure BDA00027765712200000811
Figure BDA00027765712200000812
For the best model, the iteration is ended, and the process jumps to step 9, where ErrorthThe error threshold value is a preset error threshold value, and the error threshold value in the embodiment is 5 e-02;
the iteration time t is increased by one;
step 7, training M1, M2 and M3 by using the updated training set to obtain three updated models which are respectively marked as M1t,M2t,M3t(ii) a Using the updated test set pair M1t,M2t,M3tRespectively testing to obtain the current iteration test error of each model, and respectively recording as e1t,e2t,e3t(ii) a Let the current iteration best model MItIs and min (e 1)t,e2t,e3t) A corresponding model;
and 8, if the minimum test error of the current iteration and the minimum test error of the last iteration meet the conditions:
min(e1t,e2t,e3t)>min(e1t-1,e2t-1,e3t-1) And min (e 1)t-1,e2t-1,e3t-1)<Errorth
Then MIt-1The iteration is finished for the optimal model, and the step 9 is skipped, otherwise, the step 4 is skipped for the next iteration;
step 9, modifying the output of the optimal model selected by iteration into a q-dimensional vector, training the modified optimal model by adopting an updated training set, inputting the p-dimensional vector X in the training sample, and outputting the q-dimensional vector Z in the training sample to obtain a trained snake-shaped PCB antenna return loss prediction model;
table 4 records the test error after each iteration of the iteration process, and fig. 4 is a graph thereof.
TABLE 4
Number of iterations t 1 2 3 4 5 6 7 8 9
Error of test 0.1321 0.1005 0.0888 0.0872 0.0865 0.0716 0.0685 0.0465 0.0526
As can be seen from table 4 and fig. 4, the test error of the previous 8 iterations gradually decreases, the test error increases after the 9 th iteration, and the test error after the 8 th iteration is smaller than the preset error threshold. Therefore, the iteration is ended at the 9 th time, and the optimal model corresponding to the minimum test error after the 8 th iteration is the optimal prediction model optimized by the iteration. In this embodiment, the optimal prediction model is a prediction model based on a kernel-based extreme learning machine.
After 8 iterations, the number of samples for the training set increased from 12 to 28.
And modifying the prediction model based on the support vector regression, changing the output of the prediction model into a 301-dimensional vector, and training the modified model by adopting the updated training set with 28 samples to obtain the finally trained snake-shaped PCB antenna return loss prediction model.
The prediction phase comprises:
inputting p size parameters of the serpentine PCB antenna to be predicted into a trained serpentine PCB antenna return loss prediction model, wherein the output q-dimensional vector is predicted S11Q sample point values on the curve.
In this example, the pair size parameter is [4,2.5,4.5,1.5,3,4.5 ]]S of serpentine PCB antenna11Predicting a curve, performing curve fitting on 301 sampling points obtained by prediction, and obtaining S by using HFSS electromagnetic simulation software11The curves were compared. As shown in FIG. 5, the dotted line in the figure is S obtained by the prediction method disclosed in the present invention11The curve, solid line is S obtained by HFSS simulation software11The curve and the curve are highly matched, and the method for predicting the return loss of the snake-shaped PCB antenna disclosed by the invention is verified to have higher accuracy.

Claims (6)

1. The snake-shaped PCB antenna return loss prediction method based on the three-body training algorithm is characterized by comprising a training stage and a prediction stage, wherein the training stage comprises the following steps:
step 1, obtaining labeled samples, wherein each labeled sample is represented as (X, Y, Z); wherein X is a p-dimensional vector and represents p size parameters of the serpentine PCB antenna; y is the resonant frequency of the serpentine PCB antenna with the size of X; z is a q-dimensional vector representing a serpentine PCB antenna S of size X11Q sample point values on the curve;
dividing the marked samples into a training set and a test set, and constructing an unmarked sample set U, wherein each unmarked sample is a p-dimensional vector representing p size parameters of the snake-shaped PCB antenna;
step 2, constructing three serpentine PCB antenna return loss prediction models which are respectively recorded as M1, M2 and M3; the input of each snake-shaped PCB antenna return loss prediction model is a p-dimensional vector representing snake-shaped PCB antenna size parameters, and the output is the resonant frequency of the snake-shaped PCB antenna;
step 3, training the return loss prediction models of the three snake-shaped PCB antennas by adopting a training set to obtain three initial models which are respectively recorded as M10,M20,M30(ii) a Using test set pair M10,M20,M30Respectively testing to obtain initial testing errors of each initial model, and respectively recording the initial testing errors as e10,e20,e30(ii) a Let the initial best model MI0Is and min (e 1)0,e20,e30) A corresponding model; the iteration time t is 0;
step 4, unmarked sampleSelect num samples in set U
Figure FDA0002776571210000011
Three prediction models M1 respectively input to the current iterationt,M2t,M3tIn, the output of the model is respectively
Figure FDA0002776571210000012
Forming three sets of pseudo-label data with num samples
Figure FDA0002776571210000013
Step 5, three pseudo label data set pairs M1 are adoptedt,M2t,M3tPerforming cross training to obtain 6 cross training models
Figure FDA0002776571210000014
Figure FDA0002776571210000015
Indicating the use of the jth pseudo-mark data set
Figure FDA0002776571210000016
For the ith prediction model MitCarrying out a trained cross training model; i, j belongs to {1,2,3}, and i is not equal to j;
step 6, selecting num from the test settestOne sample to 6 cross training models
Figure FDA0002776571210000017
Performing test, and calculating test error of each model
Figure FDA0002776571210000018
Minimize the test error
Figure FDA0002776571210000019
The corresponding cross training model is
Figure FDA00027765712100000110
Will be provided with
Figure FDA00027765712100000111
Corresponding pseudo-marked data set
Figure FDA00027765712100000112
And numtestAdding each test sample into a training set, and adding the numtestIndividual test samples are deleted from the test set;
if it is not
Figure FDA0002776571210000021
Figure FDA0002776571210000022
For the best model, the iteration is ended, and the process jumps to step 9, where ErrorthIs a preset error threshold;
the iteration time t is increased by one;
step 7, training M1, M2 and M3 by using the updated training set to obtain three updated models which are respectively marked as M1t,M2t,M3t(ii) a Using the updated test set pair M1t,M2t,M3tRespectively testing to obtain the current iteration test error of each model, and respectively recording as e1t,e2t,e3t(ii) a Let the current iteration best model MItIs and min (e 1)t,e2t,e3t) A corresponding model;
and 8, if the minimum test error of the current iteration and the minimum test error of the last iteration meet the conditions:
min(e1t,e2t,e3t)>min(e1t-1,e2t-1,e3t-1) And min (e 1)t-1,e2t-1,e3t-1)<Errorth
Then MIt-1The iteration is ended and the step 9 is skipped to for the optimal model, otherwise the step 4 is skipped to carry out the next iteration；
Step 9, modifying the output of the optimal model selected by iteration into a q-dimensional vector, training the modified optimal model by adopting an updated training set, inputting the p-dimensional vector X in the training sample, and outputting the q-dimensional vector Z in the training sample to obtain a trained snake-shaped PCB antenna return loss prediction model;
the prediction phase comprises:
inputting p size parameters of the serpentine PCB antenna to be predicted into a trained serpentine PCB antenna return loss prediction model, wherein the output q-dimensional vector is predicted S11Q sample point values on the curve.
2. The serpentine PCB antenna return loss prediction method of claim 1, wherein in the step 1, p size parameters of the serpentine PCB antenna in the labeled sample and the unlabeled sample are determined by an orthogonal test.
3. The serpentine PCB antenna return loss prediction method of claim 1, wherein in the step 1, HFSS analysis software is adopted to perform simulation analysis on the serpentine PCB antenna with a size parameter X to obtain S11Curve line.
4. The serpentine PCB antenna return loss prediction method of claim 1, wherein the three serpentine PCB antenna return loss prediction models are respectively: the prediction model based on Gaussian process regression, the prediction model based on support vector regression and the prediction model based on kernel limit learning machine.
5. The serpentine PCB antenna return loss prediction method of claim 1, wherein the serpentine PCB antenna has 6 size parameters, each being a resonant branch tip length L1A first coupling distance L2A first coupling length L3A second coupling length L4A second coupling distance L5Length of short circuit branch6。
6. The serpentine PCB antenna return loss prediction method of claim 4, wherein Gaussian kernel functions are adopted in the prediction models of the support vector machine and the kernel limit learning machine based on the Gaussian process.
CN202011267345.5A 2020-11-13 2020-11-13 Snake-shaped PCB antenna return loss prediction method based on three-body training algorithm Active CN112257373B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011267345.5A CN112257373B (en) 2020-11-13 2020-11-13 Snake-shaped PCB antenna return loss prediction method based on three-body training algorithm

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011267345.5A CN112257373B (en) 2020-11-13 2020-11-13 Snake-shaped PCB antenna return loss prediction method based on three-body training algorithm

Publications (2)

Publication Number Publication Date
CN112257373A true CN112257373A (en) 2021-01-22
CN112257373B CN112257373B (en) 2022-05-17

Family

ID=74265531

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011267345.5A Active CN112257373B (en) 2020-11-13 2020-11-13 Snake-shaped PCB antenna return loss prediction method based on three-body training algorithm

Country Status (1)

Country Link
CN (1) CN112257373B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113191037A (en) * 2021-03-23 2021-07-30 深圳市信维通信股份有限公司 Insertion loss prediction method based on form and position tolerance and terminal

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050181732A1 (en) * 2004-02-18 2005-08-18 Kang Joseph H. Method and apparatus for determining at least an indication of return loss of an antenna
CN104573216A (en) * 2014-12-31 2015-04-29 深圳职业技术学院 Antenna performance optimizing method and device
CN106844924A (en) * 2017-01-12 2017-06-13 桂林电子科技大学 Method based on Response Surface Method and genetic algorithm optimization PCB microstrip line constructions
CN110348106A (en) * 2019-07-08 2019-10-18 江苏科技大学 A kind of wireless power transmission EFFICIENCY PREDICTION method based on Gaussian process
CN111460665A (en) * 2020-04-01 2020-07-28 江苏科技大学 Antenna design method based on Kriging agent model and multi-target genetic algorithm
CN111651936A (en) * 2020-05-27 2020-09-11 辽宁工程技术大学 FOA-GRNN-based modeling design method for ultra-wideband antenna with double trap characteristics
CN111709192A (en) * 2020-07-07 2020-09-25 江苏科技大学 Plane inverted F-shaped antenna resonant frequency prediction method based on semi-supervised learning

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20050181732A1 (en) * 2004-02-18 2005-08-18 Kang Joseph H. Method and apparatus for determining at least an indication of return loss of an antenna
CN104573216A (en) * 2014-12-31 2015-04-29 深圳职业技术学院 Antenna performance optimizing method and device
CN106844924A (en) * 2017-01-12 2017-06-13 桂林电子科技大学 Method based on Response Surface Method and genetic algorithm optimization PCB microstrip line constructions
CN110348106A (en) * 2019-07-08 2019-10-18 江苏科技大学 A kind of wireless power transmission EFFICIENCY PREDICTION method based on Gaussian process
CN111460665A (en) * 2020-04-01 2020-07-28 江苏科技大学 Antenna design method based on Kriging agent model and multi-target genetic algorithm
CN111651936A (en) * 2020-05-27 2020-09-11 辽宁工程技术大学 FOA-GRNN-based modeling design method for ultra-wideband antenna with double trap characteristics
CN111709192A (en) * 2020-07-07 2020-09-25 江苏科技大学 Plane inverted F-shaped antenna resonant frequency prediction method based on semi-supervised learning

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
JING GAO 等: ""Resonant Frequency Modeling of Microwave Antennas Using Gaussian Process Based on Semisupervised Learning"", 《HTTPS://IDEAS.REPEC.ORG/A/HIN/COMPLX/3485469.HTML》 *
胡晓朋 等: "基于分层优化类电磁机制算法的天线优化设计", 《计算机仿真》 *
许永秀: "高斯过程模型及其在天线设计中的应用", 《中国优秀高级论文全文数据库（信息科技辑）》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113191037A (en) * 2021-03-23 2021-07-30 深圳市信维通信股份有限公司 Insertion loss prediction method based on form and position tolerance and terminal

Also Published As

Publication number Publication date
CN112257373B (en) 2022-05-17

Similar Documents

Publication Publication Date Title
Belakaria et al. Uncertainty-aware search framework for multi-objective Bayesian optimization
Gu et al. Maximum likelihood estimation for spatial models by Markov chain Monte Carlo stochastic approximation
CN112101630A (en) Multi-target optimization method for injection molding process parameters of thin-wall plastic part
JP7111671B2 (en) LEARNING APPARATUS, LEARNING SYSTEM AND LEARNING METHOD
Koziel et al. Improved modeling of microwave structures using performance-driven fully-connected regression surrogate
CN115809624B (en) Automatic analysis design method for integrated circuit microstrip line transmission line
CN112257373B (en) Snake-shaped PCB antenna return loss prediction method based on three-body training algorithm
Ye et al. Learning sparse gradients for variable selection and dimension reduction
CN114297925A (en) Power amplifier design method based on feasible domain contraction Bayesian optimization
Sedaghat et al. Compressed machine learning-based inverse model for design optimization of microwave components
Thakare et al. Microstrip antenna design using artificial neural networks
Koziel et al. Variable-fidelity CFD models and co-Kriging for expedited multi-objective aerodynamic design optimization
CN113066528B (en) Protein classification method based on active semi-supervised graph neural network
CN111967184B (en) Multi-target antenna design method based on sequence model
CN111651911B (en) Lumped element impedance sensitivity rapid calculation method and optimization method
Chen et al. Machine learning for microstrip patch antenna design: Observations and recommendations
CN111709192B (en) Planar inverted F antenna resonant frequency prediction method based on semi-supervised learning
CN110866838A (en) Network representation learning algorithm based on transition probability preprocessing
Yuan et al. A novel fault diagnosis method for second-order bandpass filter circuit based on TQWT-CNN
CN113987962A (en) Microstrip antenna optimization method and system based on student T process
Feng et al. A FOM/ROM hybrid approach for accelerating numerical simulations
Abdelhamid et al. Deep Investigation of Machine Learning Techniques for Optimizing the Parameters of Microstrip Antennas
Touloupas et al. Mixed-Variable Bayesian Optimization for Analog Circuit Sizing using Variational Autoencoders
Ma et al. A machine learning methodology for inferring network S-parameters in the presence of variability
Han et al. Structure parameter estimation method for microwave device using dimension reduction network

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