A knowledge-driven particle swarm optimization method for microwave filter tuning

Abstract

The application provides a microwave filter debugging method based on knowledge-driven particle swarm optimization, mainly including data collection, data-driven knowledge model establishment, debugging knowledge extraction, knowledge-driven particle swarm optimization, debugging process model construction and debugging according to the optimization result. The microwave filter debugging method based on knowledge-driven particle swarm optimization has the beneficial effects that a series of evaluation functions are set for each performance index of the microwave filter, the filter performance is effectively comprehensively evaluated, the knowledge model is established through a large amount of data, the debugging knowledge is mined, the particle swarm optimization algorithm is driven by using the debugging knowledge, the optimization can be effectively avoided from falling into local optimization by setting the optimization range of each adjustable parameter, and the debugging efficiency is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of microwave filter tuning technology, and in particular to a knowledge-driven particle swarm optimization method for microwave filter tuning. Background Technology

[0002] As a top priority in my country's "new infrastructure" construction, 5G base station construction relies heavily on microwave filters, which are core frequency selection devices whose filtering performance significantly impacts frequency selection quality. During production, due to design errors and manufacturing tolerances, microwave filter performance typically requires tuning to meet requirements. Traditional manual tuning methods are inefficient and costly. With the rapid development of machine learning and artificial intelligence, intelligent optimization and control have been applied to filter tuning processes, including fuzzy logic-based tuning, reinforcement learning-based tuning, and swarm intelligence-based tuning. Fuzzy logic-based tuning designs tuning rules based on experience, changing the adjustable variables of the microwave filter one by one. This method ignores the coupling relationships between variables, resulting in low tuning efficiency. Reinforcement learning-based tuning discretizes continuous adjustable variables, leading to low tuning accuracy. Swarm intelligence-based tuning can be achieved through particle swarm optimization. This technique simultaneously changes all adjustable variables in a continuous space, fully considering the coupling between variables. However, the complex relationship between adjustable variables and performance indicators, coupled with numerous algorithm iterations, limits the improvement of tuning efficiency.

[0003] In manual debugging, experienced debuggers predict the optimal range of adjustable variables based on their experience and make finer adjustments within this range. This mechanism, known as the attention mechanism, can accelerate debugging efficiency. This invention proposes a knowledge-driven particle swarm optimization method for microwave filter debugging. It learns debugging knowledge from debugging data and uses a knowledge-driven particle swarm optimization algorithm for debugging, significantly improving debugging efficiency. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a knowledge-driven particle swarm optimization method for microwave filter tuning, which mainly includes:

[0005] S1: The adjustable variable u of the adjustable component on the microwave filter is changed multiple times. * Obtain the S matrix S * The S-parameters s were sampled and measured using a vector network analyzer to construct a debug dataset;

[0006] S2: Based on the aforementioned debugging dataset, establish a knowledge model using a convolutional neural network, with performance metric I as its input. * Amplitude-frequency response R a and phase frequency response The output is an adjustable variable u *The input of the knowledge model is calculated by the S-parameters s. The knowledge model includes a convolutional layer, an activation layer, a pooling layer, a fully connected layer, and an output layer. The knowledge model is trained using the stochastic gradient descent method to obtain a trained knowledge model.

[0007] S3: Input the required performance metrics I into the trained knowledge model. * Amplitude-frequency response R a and phase frequency response The corresponding outputs constitute debugging knowledge;

[0008] S4: Use debugging knowledge to determine the optimization range of the particle swarm optimization algorithm. Based on the optimization range, performance indicators and requirements, design the evaluation function, as well as the population size, inertia weight, acceleration weight, stopping criteria and debugging parameters, and perform iterative optimization.

[0009] S5: Transform the S matrix S * Convert to Y parameter y p The debugging process model is constructed using an Elman neural network, and the optimization results in step S4 are combined to obtain the optimization results of the adjustable variables, which are used to reflect the performance indicators of the microwave filter.

[0010] S6: Once the debugging standard is met, the actual microwave filter is debugged according to the optimization results of step S5, and the overall performance of the microwave filter is comprehensively evaluated.

[0011] Furthermore, the loss function used to train the knowledge model is:

[0012]

[0013] Where N is the number of samples in the training set, and m is the adjustable variable u of the microwave filter. * The dimension of , where t represents the t-th variable among m variables, n represents the n-th sample in the debugging data, and m, n, and t are all positive integers greater than or equal to 1. Let be the error between the t-th adjustable variable of the n-th sample and the model output, and:

[0014]

[0015] in, This is the model output for the nth sample and the tth variable. It is the sample value of the t-th variable in the n-th sample;

[0016] When the value of the loss function is less than the allowable loss error, i.e., the maximum loss value δ loss At that time, the knowledge model training is complete, that is:

[0017] Ll2loss<δloss .

[0018] Furthermore, the debugging knowledge represents the adjustable variables corresponding to the performance that meet the indicator requirements, predicted by the knowledge model. Let the knowledge matrix be denoted as The range of variation for each adjustable variable is denoted as Where b represents the amount of knowledge, that is, the number of performance indicators that meet the requirements input into the trained knowledge model, t represents the t-th adjustable variable, t = 1, ..., m, and m is the number of adjustable variables.

[0019] Furthermore, the range of variation for each adjustable variable is determined by the range of the corresponding adjustable variable in the output, i.e. b represents the amount of knowledge, which is the number of performance indicators that meet the requirements input into the trained knowledge model, and t represents the t-th adjustable variable, t = 1, ..., m, where m is the number of adjustable variables.

[0020] Furthermore, the Y matrix is ​​represented as:

[0021]

[0022]

[0023] ω is the angular frequency, i is the imaginary unit, and λ is the angular frequency. k It is the k-th pole of the Y matrix, denoted as λ. k =[λ1,...,λ m ] T r 11 k It is Y 11 and Y 21 The k-th residue is denoted as r. 11 =[r 11 1 ,...,r 11 m ] T and r 21 =[r 21 1 ,...,r 21 m ] T The Y matrix is ​​composed of λ k r 11 and r 12 The parameters y are determined jointly and then obtained through vector fitting. p =[y 1 ,y 2 ,y 3 ] T , denoted as:

[0024] y 1=imag(λ k ),y 2 =real(r) 11 ),y 3 =real(r) 21 ),

[0025] Here, real and img represent the real part and the imaginary part, respectively, and the Y parameter serves as the input to the debugging process model.

[0026] Furthermore, the debugging process of the model is as follows:

[0027] Three variables were established, each mapped to imag(λ). k ),real(r 11 ) and real(r 21 The parameters of the three models are determined using gradient descent, and the loss function is:

[0028]

[0029] Where q = 1, 2, 3, m is the number of adjustable variables, Ns represents the number of times the adjustable variable sample is input, n represents the nth input sample, and k represents the kth adjustable variable; y p (k)=[y 1 (k),y 2 (k),y 3 (k)] T y p () and y p* () represent the expected output value and predicted value of the debugging process model for the input sample, respectively. Each input will only yield one set of Y parameters. The debugging process models M1, M2, and M3 are defined as follows:

[0030] M1:y 1 (n)=g1(u * (n)),M2:y 2 (n)=g2(u * (n)),M3:y 3 (n)=g3(u * (n)); y 1 (n), y 2 (n)

[0031] y 3 (n) represent the expected output value y obtained from the nth sample input. 1 y 2 y 3 u * (n) represents the adjustable variable sample of the nth input, i.e. Training of models M1, M2, and M3 will stop when the following criteria are met or the maximum number of debugging attempts is reached:

[0032] L q <δ Lq ,

[0033] δ Lq This represents the allowable error of the loss function.

[0034] Furthermore, the process of comprehensively evaluating the overall performance of the filter is as follows:

[0035] (1)f1: describes the actual center bandwidth w c and target center bandwidth w c * The difference between them is set as:

[0036]

[0037] (2)f2: describes the actual center bandwidth W c and target center bandwidth W c * The difference between them is set as:

[0038]

[0039] (3) f3: Describes the actual echo loss ζ and the target echo loss ζ * The difference between them is set as:

[0040]

[0041] Where α1 is the scalability coefficient;

[0042] (4) f4: Presents amplitude-frequency response A s21 The maximum value of (ω) in the passband is set as:

[0043]

[0044] Where ω is the angular frequency and α2 is the scaling factor 2;

[0045] (5) f5: Related to the number of peaks, reflecting the resonant state of the filter, set as:

[0046]

[0047] Nc is the number of valleys in the amplitude-frequency response. When all resonant holes and adjustable variables are in good condition, the number of valleys is m, that is, the number of valleys is equal to the number of adjustable variables. Therefore, 0≤f5≤1. The smaller the value of f5, the better the resonance state of the resonant hole.

[0048] The objective function is set as follows:

[0049]

[0050] Among them, w Nf It is f Nf The weight, f Nf =1,2,3,4,5.

[0051] The beneficial effects of the technical solution provided by this invention are: by setting the optimization range of each adjustable parameter, the optimization can be effectively avoided from getting stuck in local optima, and the debugging efficiency can be significantly improved. Attached Figure Description

[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0053] Figure 1 This is a flowchart of a microwave filter debugging method based on knowledge-driven particle swarm optimization in an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram of the Elman neural network structure in an embodiment of the present invention.

[0055] Figure 3 This is a schematic diagram illustrating the construction of the debugging process model in an embodiment of the present invention.

[0056] Figure 4 This is a simulation diagram of the debugging platform in an embodiment of the present invention.

[0057] Figure 5 This is the output of the knowledge model in this embodiment of the invention.

[0058] Figure 6 These are the debugging results of the methods disclosed in the embodiments of this invention.

[0059] Figure 7 This is the debugging result of the knowledge-free particle swarm algorithm in this embodiment of the invention.

[0060] Figure 8 These are the electromagnetic simulation results in the embodiments of this invention. Detailed Implementation

[0061] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0062] Please refer to Figure 1 , Figure 1This is a flowchart of a microwave filter debugging method based on knowledge-driven particle swarm optimization in an embodiment of the present invention. The method mainly consists of data acquisition, data-driven knowledge model building, extraction of debugging knowledge, knowledge-driven particle swarm optimization, construction of a debugging process model, and debugging based on the optimization results. The specific steps are as follows:

[0063] First, the adjustable variable u of the adjustable component on the microwave filter was changed multiple times. * Data was collected to obtain the S matrix S. * The S-parameters s are obtained using a vector network analyzer. Here, the multi-feature fusion modeling method from patent CN113158541B is cited to process the data and construct a knowledge model. Wherein, s... 11 =a 11 +b 11 i represents the reflectivity of the input signal energy, s 21 =a 21 +b 21 i represents the energy transfer rate of the input signal, a 11 and a 12 s 11 and s 21 The real part of b 11 and b 21 s 11 and s 21 The imaginary part. The amplitude and phase are calculated according to formulas (1) and (2), and then combined with the sampling frequency f, the amplitude-frequency response is obtained. and phase frequency response

[0064]

[0065]

[0066] Then, set the target performance indicators according to the communication requirements; the performance indicators include the center frequency f. c Bandwidth W and return loss ζ are calculated using S-parameters; A s21 When the value is -3dB, the corresponding sampling frequencies are the upper and lower cutoff frequencies f1 and f2, respectively. Then the center frequency f... c for:

[0067] f c =(f1+f2) / 2, (3)

[0068] Bandwidth is:

[0069] W = f1 - f2. (4)

[0070] Meeting performance requirements means that the actual performance indicators satisfy the following relationships:

[0071]

[0072] |WW * |≤δ W (6)

[0073] ζ≤ζ * (7)

[0074] in, W * and ζ * These are the target center frequency, target bandwidth, and target return loss, respectively. and δ w These are the allowable errors for the center frequency and bandwidth, respectively.

[0075] Then, based on the debugging dataset, a knowledge model is established based on a convolutional neural network. This model includes an input layer, a convolutional layer, a pooling layer, an activation function layer, a fully connected layer, and an output layer.

[0076] The loss function used to build and train the knowledge model is:

[0077]

[0078] Where N is the number of samples in the training set, and m is the adjustable variable u of the microwave filter. * The dimension of , where t represents the t-th variable among m variables, n represents the n-th sample in the debugging data, and m, n, and t are all positive integers greater than or equal to 1. Let the error between the t-th adjustable variable of the n-th sample and the model output be:

[0079]

[0080] in, This is the model output for the nth sample and the tth variable. It is the sample value of the t-th variable in the n-th sample.

[0081] When the value of the loss function is less than the allowable loss error, i.e., the maximum loss value δ loss At that time, the knowledge model training is complete, that is:

[0082]

[0083] Next, acquire debugging knowledge. Input a large number of performance metrics that meet the requirements into the trained model. * Amplitude-frequency response R a and phase frequency response The corresponding outputs constitute debugging knowledge. Debugging knowledge represents the adjustable variables corresponding to the performance that meet the performance requirements, predicted by the knowledge model. Let the knowledge matrix be denoted as The range of variation for each adjustable variable is denoted as Where b represents the amount of knowledge, that is, the number of performance indicators that meet the requirements input into the trained knowledge model, t represents the t-th adjustable variable, t = 1, ..., m, and m is the number of adjustable variables.

[0084] The debugging knowledge is used to determine the optimization range of the particle swarm optimization algorithm. The optimization range is determined by the variation range of all adjustable variables in the debugging knowledge. The decision, wherein the range of variation for each adjustable variable is determined by... Decision, that is

[0085]

[0086] b represents the amount of knowledge, which is the number of performance indicators that meet the requirements input into the trained knowledge model, and t represents the t-th adjustable variable, t = 1, ..., m, where m is the number of adjustable variables.

[0087] To improve debugging efficiency, a neural network is used to establish a debugging process model. The method for constructing an electromechanical characteristic model of a microwave cavity filter, as described in patent CN109783905B, is cited here to establish the debugging process model. Based on the S matrix S... * and adjustable variable u * Establish a debugging process model, due to the S matrix S * The S matrix is ​​a high-dimensional and complex matrix that can only reflect some of the electromechanical characteristics of microwave filters, making it difficult to directly use the S matrix to build a model. Therefore, the S matrix is... * The equivalent admittance matrix (Y matrix) converted into a microwave filter is Y = [Y]. 11 ,Y 21 The voltage and current relationship at the equivalent node of the filter is described using the following transformation:

[0088]

[0089]

[0090] Where E is related to s 11 A matrix of the same dimension with all elements equal to 1. The Y matrix can also be represented as:

[0091]

[0092]

[0093] ω is the angular frequency, i is the imaginary unit, and λ is the angular frequency. k It is the k-th pole of the Y matrix, denoted as λ. k =[λ1,...,λ m] T r 11 k It is Y 11 and Y 21 The k-th residue is denoted as r. 11 =[r 11 1 ,…,r 11 m ] T and r 21 =[r 21 1 ,...,r 21 m ] T The Y matrix is ​​composed of λ k r 11 and r 12 The decision is made jointly. The Y parameters y are then obtained through vector fitting. p =[y 1 ,y 2 ,y 3 ] T , denoted as:

[0094] y 1 =imag(λ k ),y 2 =real(r) 11 ),y 3 =real(r) 21 ), (16)

[0095] Here, real and img represent the real and imaginary parts, respectively. The Y parameter contains the mechanical properties of the microwave filter, and its dimension is lower than that of the Y matrix and S matrix. Therefore, the Y parameter is more suitable as input to the debugging process model.

[0096] Because the Y parameter y p With adjustable variable u * The relationship between them is non-linear, so a neural network is used to build the debugging process model. Furthermore, debugging is a continuous decision-making process, and different Y parameters and adjustable variables are closely related to continuity; therefore, as... Figure 2 As shown, an Ellmann neural network with memory is used, connecting local feedback to the artificial neural network. The input to the current hidden layer consists of the output of the previous hidden layer and the current network input. Furthermore, the activation function accompanying the hidden layer enhances the network's non-linear mapping capability.

[0097] The mapping between adjustable variables and S-parameters, and the construction of the debugging process model, as follows: Figure 3 As shown. Due to imag(λ), real(r) 11 ), and real(r 21The distributions of λ and imag are different. To obtain a more accurate mapping relationship, three mappings are established, one for each adjustable variable and the other for imag(λ). k ),real(r 11 ) and real(r 21 The parameters of the three models are determined using gradient descent, and the loss function is:

[0098]

[0099] Where q = 1, 2, 3, m is the number of adjustable variables, Ns represents the number of times the adjustable variable sample is input, n represents the nth input sample, and k represents the kth adjustable variable; y p (k)=[y 1 (k),y 2 (k),y 3 (k)] T y p () and y p* () represent the expected output value and predicted value of the debugging process model for the input sample, respectively. Each input will only yield one set of Y parameters. The debugging process models M1, M2, and M3 are defined as follows:

[0100] M1:y 1 (n)=g1(u * (n)),M2:y 2 (n)=g2(u * (n)),M3:y 3 (n)=g3(u * (n)); y 1 (n), y 2 (n)

[0101] y 3 (n) represent the expected output value y obtained from the kth sample input. 1 y 2 y 3 u * (n) represents the adjustable variable sample of the k-th input, i.e. Training of models M1, M2, and M3 will stop when the following criteria are met or the maximum number of debugging attempts is reached:

[0102]

[0103] δ Lq This represents the allowable error of the loss function. Next, the Y matrix is ​​obtained from the output of the debugging process model through equations (14) and (15), and the S-parameters s are obtained from the Y matrix through equations (12) and (13). Finally, the debugging process model can quickly and accurately evaluate the performance indicators of the microwave filter.

[0104] In addition, a series of evaluation functions were set to comprehensively evaluate the overall performance of the filter.

[0105] (1)f1: describes the actual center bandwidth w c and target center bandwidth w c * The difference between them is set as:

[0106]

[0107] (2)f2: describes the actual center bandwidth W c and target center bandwidth W c * The difference between them is set as:

[0108]

[0109] (3) f3: Describes the actual echo loss ζ and the target echo loss ζ * The difference between them is set as:

[0110]

[0111] Wherein, α1 is the scalability coefficient, and this parameter value can be set manually;

[0112] (4) f4: Presents A s21 The maximum value of (ω) in the passband is set as:

[0113]

[0114] Where ω is the angular frequency, and α2 is the scalability coefficient, which can be set manually;

[0115] (5) f5: Related to the number of peaks, reflecting the resonant state of the filter, set as:

[0116]

[0117] Nc is the number of valleys in the amplitude-frequency response. When all resonant holes and adjustable variables are in good condition, the number of valleys is m, that is, the number of valleys is equal to the number of adjustable variables. Therefore, 0≤f5≤1. The smaller the value of f5, the better the resonance state of the resonant hole.

[0118] The objective function is set as follows:

[0119]

[0120] Among them, w Nf It is f Nf The weights can be set manually, f Nf=1,2,3,4,5.

[0121] The improved algorithm flow is as follows:

[0122]

[0123] The stop criterion is when the number of iterations reaches the upper limit, and the tuning criterion is when f3 is less than the set value δ. f3 The value of f5 is 0, that is:

[0124] f3≤δ f3 f5 = 0. (18)

[0125] like Figure 4 As shown, this invention establishes an electromagnetic simulation model of a microwave filter based on the three-dimensional electromagnetic simulation software (High-Frequency StructureSimulator) and calculates the S-parameters. MATLAB is used to establish a microwave filter debugging knowledge model and a debugging knowledge model to implement a knowledge-driven particle swarm optimization algorithm.

[0126] The microwave filter in the simulation model has six resonant holes. Due to its symmetrical structure, that is... Adjustable variables The performance metrics of a filter include:

[0127] (1) Center frequency ω c * =2.610GHz, allowable error δ ωc =0.001GHz.

[0128] (2) Bandwidth W * =0.1930GHz, allowable error δ W =0.0005GHz.

[0129] (3) Return loss ζ * = -20dB.

[0130] First, a debugging knowledge model is constructed based on a convolutional neural network. The structure of the knowledge model is shown in Table 1, using debugging data {u * Simulation of the knowledge model is performed using s}.

[0131] Table 1 Knowledge Model Structure

[0132]

[0133] Given 21 types of inputs to the knowledge model, the parameters can be changed as follows: Figure 5As shown. The optimization ranges for all variables are as follows: θ°(1)=[0.0467mm,0.0643mm], θ°(2)=[-0.0064mm,-0.0237mm], θ°(3)=[-0.0302mm,-0.0449mm].

[0134] The parameter optimization settings are as follows:

[0135] (1) Population size: Population size is the number of particles, which is set to 50.

[0136] (2) Inertia weight c0: The inertia weight ranges from [0.4, 0.9]. The larger the value, the more global the research. Setting it to 0.9 can avoid getting trapped in local optima and improve the global search capability.

[0137] (3) Acceleration weights c1 and c2: Acceleration weights represent the rate of change of velocity. The larger the weight, the faster the velocity changes. To improve global search capability, c1 and c2 are set to 1.5 and 2, respectively.

[0138] (4) Stopping criterion: The maximum number of iterations is 100.

[0139] (5) Tuning standard: δ f3 =0.6.

[0140] (6) Scalability coefficients: α1 = 3; α2 = 10; α3 = 0.3; α4 = 10.

[0141] The amplitude-frequency response debugged using the algorithm proposed in this patent is as follows: Figure 6 As shown, for comparison, the debugging results of the knowledge-free particle swarm optimization algorithm are as follows: Figure 7 As shown. The optimization range of the knowledge-free particle swarm optimization algorithm is [2.15, 2.35], which is larger than that of the knowledge-driven optimization particle swarm optimization algorithm proposed in this invention. Apart from this, the initial settings of the two algorithms are the same, and the comparison results are as follows. Figure 6 , 7 As shown:

[0142] Figure 6 In the diagram, (a) represents the amplitude-frequency response after the first iteration. (b) represents the amplitude-frequency response after the second iteration. (c) represents the amplitude-frequency response after the third iteration. Figure 7 In the diagram, (a) represents the amplitude-frequency response after the first iteration. (b) represents the amplitude-frequency response after the second to fifth iterations. (c) represents the amplitude-frequency response after the sixth iteration. (d) represents the amplitude-frequency response after the seventh iteration.

[0143] Depend on Figure 6As can be seen, the frequency response metric met the requirements after the first iteration, which is the result of the application of debugging knowledge. Therefore, the particle swarm optimization algorithm proposed in this patent is far more efficient than particle swarm optimization algorithms without knowledge-driven mechanisms.

[0144] To ensure the generalizability of the conclusions and reduce randomness, we debugged the filter simulation model ten times each using the two algorithms. The results are as follows: Figure 8 As shown, when debugging using the knowledge-free particle swarm optimization algorithm, there were three successful experiments and seven failed experiments. In the failed experiments, after the same number of iterations, the particle swarm's position was trapped in a local optimum. However, when debugging using the improved algorithm, all experiments were successful, and the number of iterations and the time consumption were far less than those of the knowledge-free particle swarm optimization algorithm. Therefore, the knowledge-driven particle swarm optimization algorithm greatly improves debugging accuracy and efficiency.

[0145] The beneficial effects of this invention are: a series of evaluation functions are set for various performance indicators of microwave filters, which effectively evaluates the filter performance comprehensively; a knowledge model is established through a large amount of data, and debugging knowledge is mined; the debugging knowledge is used to drive the particle swarm optimization algorithm; by setting the optimization range of each adjustable parameter, the optimization can be effectively avoided from getting trapped in local optima, and the debugging efficiency is significantly improved.

[0146] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A microwave filter tuning method based on knowledge-driven particle swarm optimization, characterized in that: include: S1: The adjustable variable of the adjustable component is changed multiple times on the microwave filter. u * Obtain the S matrix S * Sampling and measurement using a vector network analyzer S parameter s Build a debug dataset; S2: Based on the aforementioned debugging dataset, establish a knowledge model using a convolutional neural network, with performance metrics as its input. I * Amplitude-frequency response R a and phase frequency response R φ The output is an adjustable variable. u * The input to the knowledge model is... S parameter s The calculation shows that the knowledge model includes convolutional layers, activation layers, pooling layers, fully connected layers, and an output layer. The knowledge model is trained using the stochastic gradient descent method to obtain a trained knowledge model. S3: Input the required performance metrics into the trained knowledge model. I * Amplitude-frequency response R a and phase frequency response R φ The corresponding outputs obtained constitute debugging knowledge; S4: Use debugging knowledge to determine the optimization range of the particle swarm optimization algorithm. Based on the optimization range, performance indicators and requirements, design the evaluation function, as well as the population size, inertia weight, acceleration weight, stopping criteria and debugging parameters, and perform iterative optimization. S5: Calculate the S matrix S * Convert to Y parameters The debugging process model is constructed using an Elman neural network, and the optimization results in step S4 are combined to obtain the optimization results of the adjustable variables, which are used to reflect the performance indicators of the microwave filter. S6: Once the debugging standard is met, the actual microwave filter is debugged according to the optimization results of step S5, and the overall performance of the microwave filter is comprehensively evaluated.

2. The microwave filter tuning method for knowledge-driven particle swarm optimization as described in claim 1, characterized in that: In step S2, the loss function used to train the knowledge model is: in, N It is the number of samples in the training set. m For microwave filter adjustable variables u * dimensionality t express m The first variable in the nth variable t indivual, n Indicates the first in the debug data n One sample, m , n and t All are positive integers greater than or equal to 1. For the first n The first sample t The error between each adjustable variable and the model output, and: in, It is the first n The first sample t The model output results for each variable. It is the first n The first sample t Sample values ​​of each variable; When the value of the loss function is less than the allowable loss error, i.e., the maximum loss value At that time, the knowledge model training is complete, that is: 。 3. The microwave filter tuning method based on knowledge-driven particle swarm optimization as described in claim 1, characterized in that: In step S3, the debugging knowledge represents the adjustable variables corresponding to the performance that meet the indicator requirements, predicted by the knowledge model. Let the knowledge matrix be denoted as The range of variation for each adjustable variable is denoted as ,in, b This is expressed as the quantity of knowledge, specifically the number of performance metrics that meet the requirements input into the trained knowledge model. t Represented as the first t One adjustable variable, t = 1,…, m , m This represents the number of adjustable variables.

4. The microwave filter tuning method based on knowledge-driven particle swarm optimization as described in claim 1, characterized in that: In step S4, the range of change for each adjustable variable is determined by the range of the corresponding adjustable variable in the output, i.e. , , b This is expressed as the quantity of knowledge, specifically the number of performance metrics that meet the requirements input into the trained knowledge model. t Represented as the first t One adjustable variable, t = 1,…, m , m This represents the number of adjustable variables.

5. The microwave filter tuning method for knowledge-driven particle swarm optimization as described in claim 1, characterized in that: In step S5, the Y matrix is ​​represented as: ω It is the angular frequency, and i is the imaginary unit. λ k yes Y The first of the matrix k There are three extreme points, denoted as [a_1, a_2, a_3, a_4, a_5, a_6, a_7, a_8, a_9, a_1, a_1, a_1, a_2 ... , r 11 k yes Y 11 and Y 21 The k The number of residues are denoted as follows: and The Y matrix is ​​composed of λ k , r 11 and r 21 The parameters of Y are determined jointly and then obtained through vector fitting. , denoted as: Here, real and img represent the real part and the imaginary part, respectively, and the Y parameter serves as the input to the debugging process model.

6. The microwave filter tuning method for knowledge-driven particle swarm optimization as described in claim 1, characterized in that: The debugging process for the model is as follows: Three variables were established, each mapped to an imag variable. λ k ), real( r 11 ) and real( r 21 The model uses gradient descent to determine the parameters of the three models, and the loss function is: in, q =1,2,3, m The number of adjustable variables. Ns This indicates the number of times the adjustable variable sample input is performed. n Indicates the first n Second input sample, k Indicates the first k One adjustable variable; , ()and () represent the expected output value and predicted value of the debugging process model for the input sample, respectively. Each input will only yield one set of Y parameters. The debugging process models M1, M2, and M3 are defined as follows: , , They represent the first n The expected output value obtained from the sample input. , , , Indicates the first n The second input of adjustable variable samples, i.e. When the following criteria are met or the maximum number of debugging attempts is reached, the training of models M1, M2, and M3 will stop: This represents the allowable error of the loss function.

7. The microwave filter tuning method for knowledge-driven particle swarm optimization as described in claim 1, characterized in that: The process of comprehensively evaluating the overall performance of the filter is as follows: (1) f 1: Describe the actual center frequency w c and target center frequency w c * The difference between them is set as: (19) (2) f 2: Describe the actual center bandwidth W c and target center bandwidth W c * The difference between them is set as: (20) (3) f 3: Describe the actual return loss ζ and target echo loss ζ * The difference between them is set as: (21) in, The scalability factor is 1; (4) f 4: Presents amplitude-frequency response A s21 ( ω The maximum value in the passband is set as follows: (22) in, ω It is angular frequency. The scalability factor is two; (5) f 5: Related to the number of peaks, reflecting the resonant state of the filter, set as follows: (23) Nc The number of troughs in the amplitude-frequency response is given by the number of troughs when all resonant apertures (i.e., the moduli) achieve good performance. m That is, the number of troughs equals the number of adjustable variables, therefore , The smaller the value, the better the resonance state of the resonant hole; The objective function is set as follows: (24) in, w Nf yes f Nf The weight, f Nf =1,2,3,4,5.

Images