Filter optimization design method based on neural network under guidance of coupling matrix

Abstract

The invention belongs to the field of filter optimization of coupling matrixes and neural networks, and discloses a filter optimization design method of a neural network based on guidance of a coupling matrix, and physical elements and mechanisms are introduced into the neural network. The method comprises the following steps: firstly, determining filter optimization indexes, randomly generating sample data in a range near the size of the optimization indexes, and screening data samples by adopting a sample screening method to obtain high-quality samples; then, extracting a coupling matrix from the obtained S parameter according to a vector fitting algorithm, and forming new coupling information M'by combining the center frequency f0 and the bandwidth BW; a full-connection neural network is used as a medium of filter device size parameters and output frequency response, the size of a to-be-optimized device of the filter is input, and coupling information M'representing S parameters is output. And the forward simulator is connected with a genetic optimization algorithm GA to realize optimization design of a specific index parameter filter. According to the method, the optimization design time is effectively saved, the optimization efficiency is improved, and the optimization dimension is reduced.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of filter optimization of coupling matrix and neural network, and particularly relates to a filter optimization design method based on neural network guided by coupling matrix. BACKGROUND

[0002] A microwave filter is an electronic device used in the microwave frequency band, mainly for selectively passing or rejecting signals within a specific frequency range. They play a crucial role in the fields of microwave communication, radar, satellite communication, wireless local area networks, etc.

[0003] In the microwave frequency band (usually defined as between 1 GHz and 300 GHz), the characteristics of signals are closely related to the transmission line characteristics of radio frequency and microwave circuits. Microwave filters are usually composed of elements such as inductors and capacitors, and can adopt various structures such as cavity filters, microstrip filters, plug-in filters, etc.

[0004] The main functions of microwave filters include:

[0005] (1) Frequency selectivity: Microwave filters can selectively pass or reject signals within a specific frequency range. This is crucial for filtering out unwanted interference signals or selecting specific communication frequency bands.(2) Signal attenuation: Microwave filters can be used to reduce or suppress the intensity of signals within a specific frequency range, achieving the purpose of signal conditioning or interference suppression.(3) Group delay: Microwave filters may introduce a certain delay when passing signals, which is a factor to be considered in some applications, especially in radar systems.(4) Insertion loss: Microwave filters will introduce a certain signal loss when passing through the filter, so the loss and filtering effect need to be balanced in design.

[0006] Traditional design methods are usually based on filter theory and electromagnetic field theory, combined with experience and manual adjustment of parameters to design filters that meet specific requirements. However, the design process in the microwave frequency band is tedious, time-consuming, and difficult. SUMMARY

[0007] The present application aims to provide a filter optimization design method based on neural network guided by coupling matrix to solve the above technical problems.

[0008] To solve the above technical problems, the specific technical scheme of the filter optimization design method based on neural network guided by coupling matrix of the present application is as follows:

[0009] A filter optimization design method based on neural network guided by coupling matrix, comprising the following steps:

[0010] Step 1, define filter optimization design index, determine optimization target;

[0011] Step 2, target index data set selection definition;

[0012] Step 3, acquisition of high-quality sample data definition;

[0013] Step 4, vector fitting algorithm extracts coupling matrix information from scattering parameters;

[0014] Step 5, neural network model training:

[0015] Step 6, optimization of geometric parameters.

[0016] Further, the step of defining filter optimization design index includes center frequency, bandwidth, return loss, insertion loss, order, and the center frequency f0 of the optimization index is 3.60GHz, 4.00GHz, and 4.20GHz, and the bandwidth BW is 0.30GHz, 0.40GHz, and 0.60GHz.

[0017] Further, the step 2 includes ADS electromagnetic simulation software for rough fast electromagnetic simulation for specified optimization index, obtaining device size value range, and then realizing joint calling between Python and CST, randomly selecting device size parameters within the rough value range, and the filter device topology is a 4th order open resonant ring, and the geometric parameters are L1, L2, L3, L4, L5, s1, s2, s3, s4, w, a total of 10 to-be-optimized sensitive geometric parameters.

[0018] Further, the step 3 adopts Bayesian optimization to obtain high-quality sample data, and this process is divided into two modules:

[0019] (1) Define fitness function fitness(x) = c(x) / BW, and the definition of c(x) is as follows: S(x) represents the frequency output response S parameter, and BW represents the filter bandwidth. If the given dB value standard cannot be reached

[0020]

[0021] (2) Bayesian optimization of Gaussian process as follows:

[0022]

[0023] f(x) ~ GP[m(x), (x, x')]

[0024] GP stands for Gaussian process, m(x) is the mean function, which means the average prediction of the target function at position x, k(x, x′) is the covariance function (also called kernel function), which means the correlation between the target values ​​at position x and position x′. After (1) and (2) in step 3, a high-quality scattering parameter sample is obtained. Further, Bayesian optimization is used to optimize S 11 The quality of the signal is judged by setting the threshold to -15dB. Further, step 4 includes: the original S parameter is composed of multiple continuous frequency points, and only 16 coupling information are needed to represent the S parameter curve. According to the transmission line theory, the influence of the phase change when the signal propagates on the transmission line is removed:

[0025]

[0026] In the formula Indicates the phase of S11, and The expression of the coupling matrix obtained by combining the admittance parameter transformation is as follows:

[0027]

[0028] Where S represents the source end, L represents the load end, and the main diagonal is the main coupling of the M matrix, which reflects the degree of mutual coupling or reflection between the various ports inside the filter.

[0029] Furthermore, the vector fitting algorithm is constructed based on a matrix of the (N+2)×(N+2) order, and performs a transverse matrix transformation.

[0030] Furthermore, step 5 includes rotating the values ​​in the coupling matrix to obtain a folded coupling matrix. Due to the principle of symmetry, non-repeated non-zero elements are selected to integrate the center frequency and bandwidth information to reconstruct the coupling information, and a fully connected neural network FCNN is used to establish a nonlinear relationship training model between the filter geometric parameters and the coupling information.

[0031] Furthermore, the fully connected neural network has a total of 6 layers, including an input layer, a hidden layer and an output layer. The input layer is a vector of size 10, representing the geometric size of the filter, followed by multiple hidden layers, each layer containing a linear layer, batch normalization and LeakyReLU activation function, gradually expanding the dimension from 10 to 402. Batch normalization is used in the network structure to improve the training stability of the model, and the LeakyReLU activation function introduces nonlinearity.

[0032] Further, the step 6 includes using the trained neural network model as a front electromagnetic simulator to replace the electromagnetic simulation software CST, inputting the filter geometric parameters to obtain predicted coupling information, and then using a genetic algorithm to constantly approach the predicted coupling information and the target coupling information, the basic flow including the steps of initialization of a population, fitness evaluation, selection, crossover, mutation and population updating, each approximation of the M matrix coupling information being regarded as an optimization process, the genetic algorithm gradually evolving in the solution space through the simulation evolution process to find a potential optimal solution of the problem, and finally outputting the geometric parameters Geometry meeting the optimization design requirements opt ,

[0033] fitness(i) = |M ideal (i) - M opt (i) |

[0034] M opt = [M ss , M s1 , M 11 , M 12 ,..., M ll , f, BW]

[0035] M ideal = [M ss ', M s1 ', M 11 ', M 12 ',..., M ll ', f', BW']

[0036] fitness(i) represents an optimization evaluation function of the genetic algorithm, and i represents the optimization round number.

[0037] The filter optimization design method based on the neural network guided by the coupling matrix has the following advantages:

[0038] The neural network filter optimization method based on the coupling matrix gives a specific solution for the optimization of a microstrip coupling filter. It includes two parts: (1) neural network training, the network is configured to select the filter geometric parameters as the input and the coupling information as the output, and reflect and fit the nonlinear relationship thereof. (2) optimization algorithm iteration. Finally, the optimized device geometric size is obtained after the minimum optimization error is obtained, and experimental tests prove the accuracy of the optimization method. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a topological structure of a microstrip band-pass filter to be optimized and a geometric size to be optimized.

[0040] Figure 2is a structural schematic diagram of a neural network.

[0041] Figure 3 is a coupling element schematic diagram of effective non-zero elements extracted from S parameters.

[0042] Figure 4 is a schematic diagram of a 4-resonator cross-coupling mechanism constituting a filter.

[0043] Figure 5 is a flowchart of a filter optimization design method based on a neural network guided by a coupling matrix. DETAILED DESCRIPTION

[0044] In order to better understand the purpose, structure and function of the present application, the present application, a filter optimization design method based on a neural network guided by a coupling matrix, will be further described in detail below in combination with the drawings.

[0045] The filter optimization design method based on a neural network guided by a coupling matrix of the present application comprises the following steps:

[0046] Step 1, define filter optimization design indicators: center frequency, bandwidth, return loss, insertion loss, order, and determine the optimization target. The center frequency f0 of the optimization indicator is 3.60 GHz, 4.00 GHz, 4.20 GHz, and the bandwidth BW is 0.30 GHz, 0.40 GHz, 0.60 GHz.

[0047] Step 2, target indicator dataset selection definition: ADS electromagnetic simulation software performs rough and fast electromagnetic simulation on the specified optimization indicators to obtain device size value ranges. Then, joint calling between Python and CST is realized, and device size parameters are randomly selected within the rough value range. The filter device topology is a 4th-order open resonant ring, and its geometric parameters are L1, L2, L3, L4, L5, s1, s2, s3, s4, w, a total of 10 sensitive geometric parameters to be optimized.

[0048] Step 3, high-quality sample data acquisition definition: due to the uncertainty of randomly generated device size parameters, Bayesian optimization is used to obtain high-quality sample data. This process is divided into two modules:

[0049] (1) define the fitness function fitness(x) = c(x) / BW, and the definition of c(x) is as follows: S(x) represents the frequency output response S parameter, and BW represents the filter bandwidth. If the given dB value standard cannot be reached

[0050]

[0051] (2) Bayesian optimization of Gaussian process as follows:

[0052]

[0053] f(x) ~ GP[m(x), (x, x')]

[0054] GP represents Gaussian process, m(x) is the mean function, which represents the average prediction of the target function value at position x, k(x, x') is the covariance function (also called kernel function), which represents the correlation between the target values at position x and position x'. The (1) (2) after step 3 can obtain high-quality scattering parameter samples. As shown in Figure 3 , in order to screen high-quality S parameter samples, the pros and cons of S 11 are judged by Bayesian optimization, and the threshold is set to -15dB.

[0055] Step 4, vector fitting algorithm extracts coupling matrix information from scattering parameters. The original S parameter is composed of multiple consecutive frequency points, and now only 16 coupling information is needed to represent the curve of S parameter. According to the transmission line theory, the influence of the change of phase of the signal when propagating on the transmission line is removed:

[0056]

[0057] In the formula , represents the phase of S11, and combined with the admittance parameter transformation can obtain the expression of the coupling matrix as follows:

[0058]

[0059] Where S represents the source end, L represents the load end, and the main diagonal line is the main coupling of the M matrix, which reflects the mutual coupling degree or reflection degree between each port inside the filter.

[0060] The vector fitting algorithm is based on the construction of (N+2) × (N+2) order matrix, and the transformation of transverse matrix is carried out.

[0061] Step 5, neural network model training: the values in the coupling matrix are transformed by matrix rotation to obtain the folded coupling matrix. Due to the principle of symmetry, the non-repeated non-zero elements are selected to reconstruct the information of center frequency and bandwidth as coupling information. The full connection neural network (FCNN) is used to establish the nonlinear relationship between the filter geometric parameters and the coupling information to train the model.

[0062] The fully connected neural network has a total of 6 layers, including an input layer, hidden layers (including linear layers, batch normalization and LeakyReLU activation function modules), and an output layer. The input layer is a vector of size 10, representing the geometric dimensions of the filter, followed by multiple hidden layers, each containing a linear layer, batch normalization and LeakyReLU activation function, gradually expanding the dimension from 10 to 402. Batch normalization is used in the network structure to improve the training stability of the model, and LeakyReLU activation function introduces nonlinearity.

[0063] Step 6, optimization of geometric parameters: the trained neural network model serves as a pre-processor for electromagnetic simulators, replacing the electromagnetic simulation software Computational Simulation Technology (CST). The input filter geometric parameters can quickly obtain the predicted coupling information, followed by a genetic algorithm to continuously approach the predicted coupling information and the target coupling information. The basic process includes steps such as population initialization, fitness evaluation, selection, crossover, mutation and population update. Each approximation of the M matrix coupling information can be regarded as an optimization process, and through the simulation of the evolutionary process, the genetic algorithm gradually evolves in the solution space to find the potential optimal solution of the problem. When combining the forward simulator and genetic algorithm to search for the optimal size parameter, this combined method can comprehensively explore the design space and improve the efficiency of optimization, finally outputting the geometric parameters Geometry opt .

[0064] fitness(i) = |M ideal (i) - M opt (i)|

[0065] M opt = [M ss , M s1 , M 11 , M 12 ,..., M ll , f, BW]

[0066] M ideal = [M ss ', M s1 ', M 11 ', M 12 ',..., M ll ', f', BW']

[0067] fitness(i) represents the optimization evaluation function of the genetic algorithm, and i represents the optimization round number.

[0068] The optimization algorithm iteration round epoch is set to 200 rounds, and the population size is 50.

[0069] AsFigure 1 The physical structure model of the filter is constructed, and the topology is selected. The optimization index is defined: the bandwidth is 0.30 GHz, 0.40 GHz, and 0.60 GHz, and the center frequency is 3.60 GHz, 4.00 GHz, and 4.20 GHz. The three bandwidths and the three center frequencies are respectively matched to form nine optimization targets. Figure 2 The structure of the entire network is shown, and there are a total of 6 layers. The hidden layer includes 4 fully connected neural network layers, and a Sigmoid activation function is used in the output layer because the S parameter ranges from 0 to 1. The obtained S parameter is subjected to key information extraction using a vector fitting algorithm, and the M matrix after transformation is as shown in the matrix arrangement structure Figure 4 The red information part of the matrix arrangement structure is reserved as coupling information together with the center frequency and the bandwidth. Finally, as shown in the flowchart Figure 5 The neural network is used as an electromagnetic simulator to perform overall optimization to obtain optimal device geometric parameters. The population number of the genetic algorithm is set to 1500, the maximum iteration number is 200, the mutation probability is 0.1, and the crossover probability is 0.5.

[0070] It can be understood that the present application is described through some embodiments, and those skilled in the art know that various changes or equivalent replacements can be made to the features and embodiments without departing from the spirit and scope of the present application. In addition, under the guidance of the present application, the features and embodiments can be modified to adapt to specific conditions and materials without departing from the spirit and scope of the present application. Therefore, the present application is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of the present application are within the scope of protection of the present application.

Claims

1. A filter optimization design method based on a neural network guided by a coupling matrix, characterized in that: The following steps are involved: Step 1: Define the filter optimization design indicators and determine the optimization goals; Step 2: Select and define the target indicator data set; Step 3: Obtain and define high-quality sample data; Step 4: A vector fitting algorithm extracts coupling matrix information from the scattering parameters; Step 5: Neural network model training: Step 6: Optimization of geometric parameters.

2. The filter optimization design method based on a neural network guided by a coupling matrix according to claim 1 is characterized in that: The steps define the filter optimization design indicators including center frequency, bandwidth, return loss, insertion loss, and order. The center frequencies f0 of the optimization indicators are 3.60 GHz, 4.00 GHz, and 4.20 GHz, and the bandwidths BW are 0.30 GHz, 0.40 GHz, and 0.60 GHz, respectively.

3. The filter optimization design method based on a neural network guided by a coupling matrix according to claim 1, characterized in that: The step 2 includes using the ADS electromagnetic simulation software to perform a rough and fast electromagnetic simulation on the specified optimization index to obtain the device size value range, and then implementing a joint call between Python and CST to randomly select the device size parameters within the rough value range. The filter device topology is a 4th-order open resonant ring, and its geometric parameters are L1, L2, L3, L4, L5, s1, s2, s3, s4, w, a total of 10 sensitive geometric parameters to be optimized.

4. The filter optimization design method based on a neural network guided by a coupling matrix according to claim 1, characterized in that: Step 3 uses Bayesian optimization to obtain high-quality sample data. This process is divided into two modules: (1) Define the fitness function fitness(x) = c(x) / BW, where c(x) is defined as follows, S(x) represents the frequency output response S parameter, and BW represents the filter bandwidth. If the given dB value standard is not met, (2) The Bayesian optimization of Gaussian process is as follows: f(x)~GP[m(x), (x, x′)] GP stands for Gaussian process, m(x) is the mean function, which indicates the average prediction of the target function value at position x, and k(x, x′) is the covariance function (also called kernel function), which indicates the correlation between the target values ​​at position x and position x′. After steps 3 (1) and (2), high-quality scattering parameter samples are obtained.

5. The filter optimization design method based on a neural network guided by a coupling matrix according to claim 4 is characterized in that: Using Bayesian optimization to optimize S 11 The quality of the sound is judged and the threshold is set to -15dB.

6. The filter optimization design method based on a neural network guided by a coupling matrix according to claim 1, characterized in that: The step 4 includes: the original S parameter is composed of multiple continuous frequency points, and only 16 coupling information are needed to represent the S parameter curve. According to the transmission line theory, the influence of the phase change when the signal propagates on the transmission line is removed: In the formula Indicates the phase of S11, and The expression of the coupling matrix obtained by combining the admittance parameter transformation is as follows: Where S represents the source end, L represents the load end, and the main diagonal is the main coupling of the M matrix, which reflects the degree of mutual coupling or reflection between the various ports inside the filter.

7. The filter optimization design method based on a neural network guided by a coupling matrix according to claim 6, characterized in that: The vector fitting algorithm is constructed based on a matrix of the (N+2)×(N+2) order, and performs a lateral matrix transformation.

8. The filter optimization design method based on a neural network guided by a coupling matrix according to claim 1, characterized in that: Step 5 involves rotating the values ​​in the coupling matrix to obtain a folded coupling matrix. Due to the principle of symmetry, non-repeated non-zero elements are selected to integrate the center frequency and bandwidth information to reconstruct the coupling information. The fully connected neural network FCNN is used to establish a nonlinear relationship training model between the filter geometric parameters and the coupling information.

9. The filter optimization design method based on a neural network guided by a coupling matrix according to claim 8, characterized in that: The fully connected neural network has a total of 6 layers, including an input layer, a hidden layer, and an output layer. The input layer is a vector of size 10, representing the geometric size of the filter, followed by multiple hidden layers, each of which contains a linear layer, batch normalization, and a LeakyReLU activation function, gradually expanding the dimension from 10 to 402. Batch normalization is used in the network structure to improve the training stability of the model, and the LeakyReLU activation function introduces nonlinearity.

10. The filter optimization design method based on a neural network guided by a coupling matrix according to claim 1, characterized in that: The step 6 includes using the trained neural network model as a front-end electromagnetic simulator to replace the electromagnetic simulation software CST, inputting the filter geometric parameters to obtain the predicted coupling information, and then using the genetic algorithm to continuously approximate the predicted coupling information and the target coupling information. The basic process includes initializing the population, fitness evaluation, selection, crossover, mutation and population update steps. Each approximation of the M matrix coupling information is regarded as an optimization process. By simulating the evolutionary process, the genetic algorithm gradually evolves in the solution space, finds the potential optimal solution to the problem, and finally outputs the geometric parameters Geometry that meet the optimization design requirements. opt , fitness(i)=|M ideal (i)-M opt (i)| M opt =[M ss ,M s1 ,M 11 ,M 12 ,...,M ll ,f,BW] M ideal =[M ss ′,M s1 ′,M 11 ′,M 12 ′,...,M ll ′,f′,BW′] fitness(i) represents the optimization evaluation function of the genetic algorithm, and i represents the number of optimization rounds.

Images