CN112597727B

CN112597727B - Novel rapid and efficient filter small sample modeling and optimizing method

Info

Publication number: CN112597727B
Application number: CN202011572732.XA
Authority: CN
Inventors: 喻梦霞; 陈林; 汪家兴; 李桂萍
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2020-12-25
Filing date: 2020-12-25
Publication date: 2023-04-21
Anticipated expiration: 2040-12-25
Also published as: CN112597727A

Abstract

The invention discloses a novel, rapid and efficient filter small sample modeling and optimizing method, and provides a novel, rapid and efficient filter small sample modeling and model optimizing method, which is a method for forward modeling of a filter based on a Gaussian process of priori knowledge and reverse optimization of a particle swarm based on KBGP, aiming at the defects that modeling optimization calculation is long in the original modeling optimizing method and the actual filter circuit design and sample data size is limited. In forward modeling, the method trains an ANNs coarse network using coarse sample data; then, the output vector of the ANNs is used as priori knowledge to be injected into the GPR, so that sample data of the GPR can be reduced, and a high-precision circuit model is obtained; during reverse optimization, the best global input vector is obtained by using a Particle Swarm Optimization (PSO) method to quickly optimize the input vector of the model.

Description

Novel rapid and efficient filter small sample modeling and optimizing method

Technical Field

The invention relates to the field of simulation modeling and model optimization of filters, in particular to a novel rapid and efficient method for modeling and model optimization of small samples of filters, which is a method for forward modeling of the filters and reverse optimization based on particle swarm (KBGP) of Gaussian process based on priori knowledge.

Background

With the development of wireless communication technology, various new technologies are continuously developed, filters are also continuously developed to higher frequencies, the performance and the circuit structure of the filters are more complex, and the original modeling optimization method can not meet the requirements of modern design gradually. The model of modern filter design needs to be able to describe not only the behavior characteristics, but also accurately reflect the variation of the variables that causes the change of the filter performance. The modern filter accurate modeling optimization technology is a full-wave analysis method based on electromagnetic theory, an ideal circuit is obtained through professional electromagnetic simulation software (such as CST, HFSS and the like), but the full-wave analysis method consumes more calculation resources and takes a long time, the structure is complex, the precision is improved, the calculation times and the resource consumption are exponentially increased along with the increase of the working frequency of the filter, the calculation time of the software is too long, the calculation result is often not converged or does not meet the precision requirement, and the method is difficult in actual complex engineering application, so that a rapid and efficient modeling optimization method is urgently needed for design to adapt to the continuously changing design requirements.

The advent of machine learning provides new ideas for modeling optimization of filters. If the model generated by machine learning is used for replacing electromagnetic simulation, the modeling optimization of the modern filter can be efficiently performed by utilizing the characteristics of high machine learning precision, short calculation time and less occupied memory. Artificial neural networks (ans) and Gaussian Process Regression (GPR) are two efficient machine learning modeling optimization methods that have been increasingly applied in recent years to modeling and optimization of filters. The ANNs model can be built by using a large amount of sample data, and complex relations which are difficult to describe between geometric structure sizes and electric parameters of the filter can be approximated. However, the network structure of ANNs is difficult to determine, and especially the selection of the number of network layers and the number of neurons lacks theoretical support, so that too many training samples are needed, the time is too long, and quick modeling is difficult to realize. Knowledge neural networks (KBNN) are an improved method of ANNs, which, although able to reduce training samples to some extent, have limited effectiveness when training samples accurately are not sufficient. GPR is taken as a powerful tool, the model training time is short, the precision is high, the GPR can be applied to modeling optimization of a complex circuit, the shortcoming of ANNs can be exactly overcome, more importantly, GPR is suitable for training of small samples, however, the GPR is large in calculation excess, and the modeling efficiency is low. Thus, the use of a machine learning method alone cannot meet the need for rapid and accurate modeling optimization of modern filters.

Aiming at the problems of long time consumption and limited sample data volume of modeling optimization calculation in the original modeling optimization method and the actual filter circuit design, a novel, rapid and efficient filter small sample modeling and model optimization method is provided, and the method is a method (KBGP) for forward modeling of a filter based on a Gaussian process of priori knowledge and reverse optimization based on a particle swarm. The novel method (KBBGP) has the advantages of small sample number, high modeling and optimizing speed and high accuracy, can solve the problem of modeling and optimizing the original circuit, and realizes the rapid, efficient and high-accuracy modeling and optimizing of the filter.

Disclosure of Invention

The invention aims at: the invention provides a novel, rapid and efficient filter small sample modeling and optimizing method, and aims at the defects that modeling and optimizing calculation are long in time consumption and sample data size is limited in the original modeling and optimizing method and the actual filter circuit design. In forward modeling, the method trains an ANNs coarse network using coarse sample data; then, the output vector of the ANNs is used as priori knowledge to be injected into the GPR, so that sample data of the GPR can be reduced, and a high-precision circuit model is obtained; during reverse optimization, the best global input vector is obtained by using a Particle Swarm Optimization (PSO) method to quickly optimize the input vector of the model.

The technical scheme adopted by the invention is as follows:

the KBGP method provided by the invention has two stages: forward modeling and reverse optimization, see fig. 1. In forward modeling, the coarse model is a low-precision model, namely a multi-layer perceptive neural network (MLPANN), and training sample data of the coarse model is derived from simulation software ADS. ADS is electromagnetic software based on a road, simulation speed is high, simulation precision is low, and enough sample data can be obtained in extremely short time by using ADS to construct a rough model. And then, by utilizing the GPR and combining the prior knowledge output by the rough model, mapping between the rough model and the results obtained by full-wave electromagnetic simulation is quickened, the accurate convergence of the GPR is realized, and the forward modeling process of the KBGP is completed. During reverse optimization, the PSO global optimization characteristic is utilized, the output result of the GPR is used as fitness, the position of particles in the PSO algorithm is updated, the global optimal solution can be obtained in a few seconds, and the reverse optimization process of the KBGPR is completed.

In order to achieve the above object, the present invention has the following technical scheme:

a novel rapid and efficient filter small sample modeling and model optimizing method comprises the following steps:

step 1: and establishing a rough model. Firstly, determining an input vector and an output vector of a rough model, forming an input vector by input variables and forming an output vector by output variables; then, constructing an equivalent rough road model by utilizing ADS, and rapidly acquiring sample data of a rough network; finally, training the network and detecting whether the precision of the rough network model meets the requirement;

step 2: a GPR sample set is established. And (3) uniformly sampling by using CST software to obtain a small amount of full-wave electromagnetic simulation results. Taking the prediction result of ANN as priori knowledge, and combining with CST full-wave simulation result to form GPR sample data;

step 3: the GPR model was trained. Invoking GPR sample data to train a GPR model; then, calling CST to generate a test sample, and checking the reliability of the proposed method through average absolute error to complete the modeling process of the model;

step 4: optimization of the KBBGP model. And taking the output result of the GPR model as the PSO fitness, continuously updating the position of particles in a PSO algorithm, and completing the optimization process of the whole KBGP model in a few seconds.

In summary, due to the adoption of the technical scheme, the beneficial effects of the invention are as follows:

(1) The modeling technology of the KBGP method is provided, so that the limitation of the original neural network model on the huge sample number can be eliminated, the sample number is greatly reduced, and the modeling technology has high calculation precision and calculation efficiency in solving the problem of a small sample rapid filter model;

(2) The forward modeling process of the KBGP method provided by the invention comprises a coarse model and a Gaussian Process (GPR). According to the method, the characteristics of ANNs and GPR are combined, and small sample modeling is rapidly realized under the condition that model accuracy can be guaranteed;

(3) The PSO model is used alone, so that local convergence is easy to fall into, and the calculation speed is low; according to the invention, the output result of the KBGP model is taken as the fitness, and the positions of particles in the PSO algorithm are updated, so that the model can be reversely optimized, the optimal input vector value can be quickly found, and the local optimization is avoided;

(4) The KBGP method provided by the invention not only can reduce the modeling training sample and modeling time, but also can improve the prediction precision of the model, and shows that the KBGP method has important value in filter design.

Drawings

For a clearer description of the technical solutions of embodiments of the present invention, the drawings that are needed in the embodiments will be briefly described below, it being understood that the following drawings only illustrate some embodiments of the present invention and should not be considered limiting in scope, and other related drawings can be obtained according to these drawings without inventive effort for a person skilled in the art, wherein:

fig. 1 is a schematic diagram of the overall modeling optimization process of the novel KBGP;

FIG. 2 is a forward modeling flow diagram of the novel KBGP;

FIG. 3 is a schematic diagram of a band reject filter;

FIG. 4 shows the |S of the different methods when the same test sample is used ₂₁ An I curve;

fig. 5 is a filter optimization result based on the KBGP method.

Detailed Description

The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the particular embodiments described herein are illustrative only and are not intended to limit the invention, i.e., the embodiments described are merely some, but not all, of the embodiments of the invention. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.

Thus, the following detailed description of the embodiments of the invention, as presented in the figures, is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be made by a person skilled in the art without making any inventive effort, are intended to be within the scope of the present invention.

It is noted that relational terms such as "first" and "second", and the like, are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

The features and capabilities of the present invention are described in further detail below in connection with examples.

Example 1

Based on the technical indexes of the band-stop filter, the process and modeling schematic diagrams shown in fig. 1 and 2 are adopted in the flow, the circuit structure shown in fig. 3 is adopted, and the specific modeling optimization comprises the following steps:

step 1: and establishing a rough model of the micro-strip band-stop filter based on ANN. Index requirements of band reject filter:

|S21|≥0.92(-0.72dB)，for 2.8GHz≤f≤3.6GHz

|S21|≤0.05(-26dB)，for 4.1GHz≤f≤4.9GHz

|S21|≥0.92(-0.72dB)，for 5.4GHz≤f≤6.3GHz

determining input vectors for a coarse model

(/>

The ith frequency point representing the jth geometry), see sample data for the microstrip band reject filter of table 1, the output vector is |s ₂₁ |。/>

Table 1 sample data for microstrip band reject filter

And selecting a Multilayer Neural Network Structure (MLPANNs) by the rough model, selecting an ADS simulation result of 100 groups of geometric structures as sample data by using a full-factor orthogonal experimental method, and training the rough MLPANN model, wherein the set test error precision is 0.7dB.

Step 2: a GPR sample set is established.

And generating an input vector, and calling the CST to perform full-wave electromagnetic simulation. Taking the prediction result of ANN as priori knowledge and combining with CST simulation result to form a GPR sample set;

step 3: the GPR model was trained.

GPR covariance function expression:

wherein the method comprises the steps of

σ _f ，σ _m The characteristic length and the standard deviation of the signal, respectively.

Firstly, selecting GPR sample sets with different numbers, training GPR, testing generalization errors by using test samples, and comparing the generalization errors with the original method, wherein when the number of the samples is different in FIG. 4 and Table 2, the average errors of the novel KBGP method and the original method are shown;

number of samples

6

9

12

15

18

21

24

KBNN

20.47％

9.47％

3.54％

2.28％

1.37％

1.01％

0.76％

GPR

2.39％

1.93％

1.49％

1.36％

1.32％

1.14％

1.05％

Novel KBGP

1.28％

1.15％

0.76％

0.68％

0.61％

0.56％

0.48％

Table 2 average error of novel KBGP method and original method for different sample numbers

Step 4: model optimization, finding the best input vector of the circuit model.

The optimization circuit inputs a parameter vector in combination with a Particle Swarm Optimization (PSO) algorithm. Setting the fitness function as follows according to the index requirement:

fit＝3*S ₂₁ i (I) _Passband +0.8*(0.93-S ₂₁ I (I) _{Low stop band} )+1.2*(0.93-S ₂₁ I (I) _{High stop band} )

In the PSO algorithm optimizing process, the fit function should be made to take the minimum value as much as possible. After optimizing, the input vector of the circuit model can be obtained, and the optimizing result is shown in fig. 5.

Step 5: and (5) model evaluation analysis.

As can be seen from fig. 4 and table 2, compared with the original methods such as the knowledge neural network (KBNN) and the gaussian regression process (GPR), the fitness of the KBGP result and the CST result is best, when the number of samples is 12, the error has reached 0.76%, and the KBGP modeling method has the highest fitness, the KBGP error is the lowest, the accuracy is the highest, and the samples are the least. The method has high modeling accuracy and small sample number, and is particularly suitable for actual complex engineering.

As can be seen from fig. 5, after the optimization by the method, the KBGP result is completely consistent with the full-wave CST accurate result, so that the model error is further reduced, and the model accuracy is improved; therefore, the KBGP modeling and the optimization method thereof can not only reduce the training sample and modeling time of the modeling, but also ensure the model precision, and the prediction precision is higher, which shows that the modeling optimization method provided by the invention has important value in the filter design optimization, as shown by the consumed time of the novel KBGP method of the table 3 and the original method, the method replaces the original filter modeling optimization method, a reliable model can be quickly and accurately built by using a small amount of sample data, the modeling time is greatly reduced, and the purposes of high-efficiency and quick modeling optimization are realized, as shown in the tables 2 and 3.

Algorithm	KBNN	GPR	Novel KBGP
				Time(s)	9.46	22.4	3.13

TABLE 3 time spent on novel KBGP method and original method

The above description is only of the preferred embodiments of the present invention and is not intended to limit the scope of the present invention, and any modifications, equivalents, improvements and modifications within the spirit and principles of the invention will become apparent to those skilled in the art.

Claims

1. A novel rapid and efficient modeling and optimizing method for small samples of a filter is characterized by comprising the following steps of: the method comprises the following steps:

step 1: establishing a rough model, firstly, determining an input vector and an output vector of the rough model, forming an input vector by input variables and an output vector by output variables; then, constructing an equivalent rough road model by utilizing ADS, and rapidly acquiring sample data of a rough network; training a network and detecting whether the precision of the rough network model meets the requirement;

step 2: establishing a GPR sample set, uniformly sampling by using CST software to obtain a small amount of full-wave electromagnetic simulation results, taking the prediction result of ANN as priori knowledge, and combining the CST full-wave simulation results to form GPR sample data;

step 3: invoking GPR sample data to train a GPR model; calling CST to generate a test sample, and checking the reliability of the proposed method through average absolute error to complete the modeling process of the model;

step 4: and (3) optimizing the KBGPR model, taking the output result of the GPR model as the PSO fitness, and continuously updating the positions of particles in a PSO algorithm to complete the whole optimizing process of the KBGPR model.

2. The novel, fast and efficient modeling and optimization method for small samples of a filter according to claim 1, wherein the method is characterized in that: the coarse model is a coarse model of the ANN-based microstrip band-stop filter.

3. The novel, fast and efficient modeling and optimization method for small samples of a filter according to claim 1, wherein the method is characterized in that: the training GPR model in step S3 uses a GPR covariance function expression:

wherein the method comprises the steps of

4. The novel, fast and efficient modeling and optimization method for small samples of a filter according to claim 1, wherein the method is characterized in that: in the step S4, during the optimizing process of the PSO algorithm, the fit function takes an infinitely close minimum value.

5. The novel, fast and efficient modeling and optimization method for small samples of a filter according to claim 1, wherein the method is characterized in that: and (5) after the KBGP model optimization process is completed in the step (4), performing model evaluation analysis in the step (5).