A neural network-based microwave filter coupling matrix extraction method
By constructing a convolutional neural network model based on a neural network approach, the problems of computational complexity and insufficient accuracy in microwave filter coupling matrix extraction are solved, achieving fast and high-precision coupling matrix extraction and improving the efficiency of filter design and tuning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- UNIV OF ELECTRONICS SCI & TECH OF CHINA
- Filing Date
- 2022-09-21
- Publication Date
- 2026-05-08
AI Technical Summary
Existing methods for extracting the coupling matrix of microwave filters suffer from computational complexity and insufficient accuracy. In particular, the Cauchy method is fast but has low out-of-band accuracy, while the optimization method has high accuracy but is slow, making it difficult to optimize quickly and efficiently during filter design and tuning.
A neural network-based approach is adopted, which involves constructing a convolutional neural network model, training it with a large amount of data, and quickly extracting the coupling matrix of a microwave filter. This includes establishing a forward dataset, a training dataset, building the model, and extracting the coupling matrix in reverse. The training process is optimized using the SmoothL1 loss function.
It achieves fast and high-precision extraction of microwave filter coupling matrix, reduces data acquisition time, improves the real-time performance and accuracy of filter design and tuning, and simplifies the complex formula derivation process.
Smart Images
Figure CN115481591B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of microwave technology and artificial intelligence applications, and specifically relates to a method for extracting the coupling matrix of a microwave filter. It provides a neural network-based method for extracting the coupling matrix of a microwave filter, which can be used for the design and tuning of microwave filters. Background Technology
[0002] The development of modern wireless communication technology places increasingly higher demands on the design and tuning efficiency of microwave filters. While comprehensive synthesis methods exist for filter synthesis and parameter extraction, the initial full-wave simulation results of the synthesized filters often exhibit poor frequency response performance and unpredictability, requiring further optimization and consuming considerable time to achieve the desired results. In this optimization process, extracting the filter's coupling matrix using S-parameters allows the determination of the next tuning direction based on the difference between the extracted coupling matrix and the designed ideal coupling matrix. Furthermore, for the designed physical filter, comparing the measured difference between the S-parameter-extracted coupling matrix and the ideal coupling matrix during tuning can also clarify the tuning direction of the physical filter, thus assisting in both manual and fully automated tuning.
[0003] Currently, the most common methods for extracting filter coupling matrices are the Cauchy method and the optimization method. Each has its advantages and disadvantages: the Cauchy method is fast but has lower out-of-band accuracy; the optimization method has higher accuracy but is slow and requires significant time for optimization. In contrast, neural network technology combines fast extraction speed with relatively consistent overall regression accuracy. Therefore, applying neural networks to filter coupling matrix extraction, with the assistance of computer technology and through learning from large amounts of data, holds promise for solving the increasing complexity and time-consuming problems in this field. Summary of the Invention
[0004] The purpose of this invention is to provide a method for extracting the coupling matrix of a microwave filter based on a neural network, in order to solve the problems of computational complexity and insufficient accuracy of existing methods such as the Cauchy method and optimization method.
[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:
[0006] A method for extracting the coupling matrix of a microwave filter based on a neural network, characterized by comprising the following steps:
[0007] S1: Establish a positive dataset extraction program for calculating S-parameters of microwave filter coupling matrix;
[0008] S2: Extract the training dataset containing S-parameters and microwave filter coupling matrix through a program;
[0009] S3: Construct a convolutional neural network model;
[0010] S4: Train the convolutional neural network model using the training dataset;
[0011] S5: Use the trained convolutional neural network model to perform inverse extraction of the microwave filter coupling matrix.
[0012] Furthermore, the specific process of step S1 is as follows:
[0013] S1.1: Let the coupling matrix to be solved be M, and its corresponding microwave filter have the order N, bandwidth BW, center frequency F0, quality factor Q, and solution starting frequency F. s The solution terminates at frequency F. e Solving for the frequency interval is Step;
[0014] S1.2: The order N1 is calculated as N1 = N + 2, and the number of frequency solutions P is calculated.
[0015] S1.3: Construct an N1-order matrix R:
[0016]
[0017] S1.4: Construct an N1-order matrix U:
[0018]
[0019] S1.5: For each solution frequency F i Normalize: F i ′ for F i The normalization result;
[0020] S1.6: Calculate the solution frequency F i The impedance matrix Z at the location i : Where j is the imaginary unit, then the impedance matrix Z i The inverse of Z i -1 ;
[0021] S1.7: Let matrices A and B be matrices Z. i -1 Submatrix:
[0022]
[0023]
[0024] S1.8: Calculate the solution frequency F i S at the location 11 For S11 (F i )=1+2j[Z i -1 ] 1,1 ;
[0025] S1.9: Calculate the solution frequency F i S at the location 21 For S 21 (F i )=-2j[Z i -1 ] N1,1 ;
[0026] S1.10: Calculate the solution frequency F i Group latency
[0027] Furthermore, the specific process of step S2 is as follows:
[0028] S2.1: Based on the design objective of the Nth-order filter, the N1-order ideal coupling matrix M is synthesized using the generalized Chebyshev method. ideal ;
[0029] S2.2: In its ideal coupling matrix M ideal Within the ±error range, N is randomly generated. sample The set of coupling matrices, where error is the possible detuning range of the filter;
[0030] S2.3: Calculate N using the positive dataset extraction procedure. sample S corresponding to the group coupling matrix 11 S 21 TD parameters, and S of each coupling matrix 11 S 21 The training dataset is constructed by using the TD parameter combination as input and the coupling matrix as the label vector.
[0031] Furthermore, the training samples include the following two types:
[0032] First type of training sample: S 11 The real part, S 11 The imaginary part, S 21 The real part and S 21 The virtual part is used as a channel to form four channels of input data;
[0033] The second type of training sample: S 11 S 21 Each of the input channels TD and TD is used as a channel, and the data is expressed in dB to form a three-channel input data.
[0034] Furthermore, in step S3, the convolutional neural network model includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected in sequence.
[0035] Furthermore, the input layer uses either four-channel or three-channel input data; the first and second convolutional layers use the ELU function as the activation function, and the kernel size is 3×1; the first, second, and third pooling layers all use max pooling, with a filter size of 2×1; the first fully connected layer has 1024 neurons, the second fully connected layer has 256 neurons, and the third fully connected layer has 64 neurons, all using the ELU function as the activation function; the output layer is a coupling matrix, output in the form of a one-dimensional vector.
[0036] Furthermore, the specific training process for step S4 is as follows:
[0037] S4.1: Use the input data from the training samples as the input to the neural network model, and the neural network model outputs the prediction vector;
[0038] S4.2: Set the SmoothL1 loss function to the loss function of the neural network:
[0039]
[0040] Where x is the average of the absolute values of the element differences between the prediction vector and the label vector;
[0041] S4.3: Set the target to terminate neural network training and complete model training: Stop training when the number of training rounds reaches the preset threshold or the loss function reaches the preset threshold. If the number of training rounds reaches the preset threshold but the loss function has not yet reached the preset threshold, repeat step S2, collect more sample data, and retrain until the loss function reaches the preset threshold.
[0042] Furthermore, the specific process of step S5 is as follows: the S-parameters and TD-parameters of the microwave filter to be extracted are obtained through testing or simulation, and the input data of the neural network model is formed. The corresponding coupling matrix is output by the trained neural network model.
[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0044] This invention provides a method for extracting the coupling matrix of a microwave filter based on a neural network. By introducing a neural network into the process of extracting the coupling matrix of a microwave filter based on S-parameters, only a large amount of data is needed for training, avoiding complex formula derivations and simplifications, making it convenient and fast. After training, the trained convolutional neural network model can quickly extract the coupling matrix of the microwave filter from the S-parameters in reverse, demonstrating strong real-time performance. Furthermore, based on the dataset construction method proposed in this invention, data acquisition is fast, thus avoiding the significant time spent in the data acquisition stage. Attached Figure Description
[0045] Figure 1 The example shows the S-parameter plot corresponding to the sixth-order coupling matrix.
[0046] Figure 2 The S-parameter plot is shown for an example of an eighth-order coupling matrix.
[0047] Figure 3 This is a schematic diagram of the structure of the one-dimensional convolutional neural network in this invention.
[0048] Figure 4 This is a flowchart of the microwave filter coupling matrix extraction method based on neural networks in this invention.
[0049] Figure 5 This is the equivalent circuit diagram of a fourth-order cavity filter.
[0050] Figure 6 for Figure 5 The S-parameter diagram of the equivalent circuit of the fourth-order cavity filter is shown.
[0051] Figure 7 for Figure 5 The S-parameters of the equivalent circuit of the fourth-order cavity filter shown are in the detuned state.
[0052] Figure 8 For the present invention to Figure 5 The S-parameter curves and comparison figures of the fourth-order cavity filter after coupling matrix extraction are shown. Detailed Implementation
[0053] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0054] Example 1
[0055] This embodiment provides a method for extracting the coupling matrix of a microwave filter based on a neural network, such as... Figure 4As shown; whereby, the neural network, as a mathematical model of distributed parallel information processing in machine learning that mimics the behavioral characteristics of animal neural networks, can achieve the purpose of processing information by adjusting the interconnections between a large number of internal nodes, depending on the complexity of the system; based on this, this embodiment specifically includes the following steps:
[0056] S1: A program for extracting the positive dataset for calculating S-parameters by establishing the microwave filter coupling matrix according to the formula;
[0057] S2: Extract the training dataset containing S-parameters and microwave filter coupling matrix through a program;
[0058] S3: Construct a convolutional neural network model;
[0059] S4: Train the convolutional neural network model using the training dataset;
[0060] S5: Use the trained convolutional neural network model to perform inverse extraction of the microwave filter coupling matrix.
[0061] Furthermore, the specific process of step S1 is as follows:
[0062] S1.1: Let the coupling matrix to be solved be M, and its corresponding microwave filter have the order N, bandwidth BW, center frequency F0, quality factor Q, and solution starting frequency F. s The solution terminates at frequency F. e The frequency interval is solved in Step; this embodiment provides an example of the sixth-order coupling matrix corresponding to a fourth-order filter and the eighth-order coupling matrix corresponding to a sixth-order filter, but the application scope of this invention is not limited to these two forms; the sixth-order coupling matrix corresponding to the fourth-order filter is shown in Table 1, and the corresponding S-parameters are as follows. Figure 1 As shown;
[0063] Table 1
[0064] Matrix1 S 1 2 3 4 L S 0 1.035154 0 0 0 0 1 1.035154 0 0.91058 0 0 0 2 0 0.91058 0 0.699925 0 0 3 0 0 0.699925 0 0.91058 0 4 0 0 0 0.91058 0 1.035154 L 0 0 0 0 1.035154 0
[0065] The eighth-order coupling matrix corresponding to the sixth-order filter is shown in Table 2, and its corresponding S-parameters are as follows: Figure 2 As shown;
[0066] Table 2
[0067] Matrix2 S 1 2 3 4 5 6 L S 0 1.029418 0 0 0 0 0 0 1 1.029418 0 0.898316 0 0 0 0 0 2 0 0.898316 0 0.68232 0 0 0 0 3 0 0 0.68232 0 0.371118 0 0.753555 0 4 0 0 0 0.371118 0 0.110428 0 0 5 0 0 0 0 0.110428 0 0.489005 0 6 0 0 0 0.753555 0 0.489005 0 1.029418 L 0 0 0 0 0 0 1.029418 0
[0068] S1.2: The order N1 is calculated as N1 = N + 2, and the number of frequency solutions P is calculated.
[0069] S1.3: Construct an N1-order matrix R:
[0070]
[0071] S1.4: Construct an N1-order matrix U:
[0072]
[0073] S1.5: For each solution frequency F i Normalize: F i ′ for F i The normalization result;
[0074] S1.6: Calculate the solution frequency F i The impedance matrix Z at the location i : Where j is the imaginary unit, then the impedance matrix Z i The inverse of Z i -1 ;
[0075] S1.7: Let matrices A and B be matrices Z. i -1 Submatrix:
[0076]
[0077]
[0078] S1.8: Calculate the solution frequency F i S at the location 11 For S 11 (F i )=1+2j[Z i -1 ] 1,1 ;
[0079] S1.9: Calculate the solution frequency F i S at the location 21 for
[0080] S1.10: Calculate the solution frequency F i Group latency
[0081] Furthermore, the specific process of step S2 is as follows:
[0082] S2.1: Based on the design objective of the Nth-order filter, the N1-order ideal coupling matrix M is synthesized using the generalized Chebyshev method. ideal ;
[0083] S2.2: In its ideal coupling matrix M ideal N is randomly generated within the range of ±error. sampleThe set of coupling matrices, where error is the possible detuning range of the filter;
[0084] S2.3: Calculate N using step S1. sample S corresponding to the group coupling matrix 11 S 21 TD parameters, and S of each coupling matrix 11 S 21 The TD parameter combination is used as input, and the coupling matrix is used as the label vector to form training samples, thereby constructing a training dataset for neural network training; the specific form of the training samples can be adaptively adjusted according to actual needs, including the following two forms:
[0085] First type of training sample: S 11 The real part, S 11 The imaginary part, S 21 The real part and S 21 The virtual part is used as a channel to form four channels of input data;
[0086] The second type of training sample: S 11 S 21 Each of the input channels TD and TD is used as a channel, and the data is expressed in dB to form a three-channel input data.
[0087] Furthermore, in step S3, the convolutional neural network model includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer connected in sequence, as follows: Figure 3 As shown; more specifically: the input layer is four-channel or three-channel input data, corresponding to 128×4 or 128×3 channels; the first and second convolutional layers use the ELU function as the activation function, with a kernel size of 3×1, and the number of channels in the first and second convolutional layers is 128×32 and 64×64, respectively; the first, second, and third pooling layers all use max pooling, with a filter size of 2×1, and the number of channels in the first, second, and third pooling layers is 64×32, 32×64, and 16×64, respectively; the first fully connected layer has 1024 neurons, the second fully connected layer has 256 neurons, and the third fully connected layer has 64 neurons, all using the ELU function as the activation function; the output layer is a coupling matrix, outputting in the form of a one-dimensional vector, with 10×1 channels.
[0088] Furthermore, in step S4, the neural network model from step S3 is trained using the dataset from step S2. The specific training process is as follows:
[0089] S4.1: Use the input data from the training samples as the input to the neural network model, and the neural network model outputs the prediction vector;
[0090] S4.2: Set the SmoothL1 loss function to the loss function of the neural network, with the following form:
[0091]
[0092] Where x is the average of the absolute values of the element differences between the prediction vector and the label vector;
[0093] S4.3: Set the target for terminating neural network training: Stop training when the number of training rounds reaches a preset threshold or the loss function reaches a preset threshold. If the number of training rounds reaches the preset threshold but the loss function has not yet reached the preset threshold, repeat step S2, collect more sample data, and retrain until the loss function reaches the preset threshold.
[0094] Furthermore, the specific process of step S5 is as follows: the S-parameters and TD-parameters of the microwave filter to be extracted are obtained through testing or simulation, and the input data of the neural network model is formed. The corresponding coupling matrix is output by the trained neural network model.
[0095] The effects of the present invention will be further described below with reference to the accompanying drawings and specific practical examples.
[0096] To demonstrate the effectiveness of this invention in extracting the coupling matrix of a microwave filter, this embodiment extracts the coupling matrix parameters of an equivalent circuit model of a fourth-order cavity filter, and compares the response curve of the microwave filter after parameter extraction with the response curve before parameter extraction. This demonstrates that this invention can accurately extract the coupling matrix of the microwave filter. Specifically, the test environment is: MATLAB 2020b, Python 3.8, PyTorch 1.10.0, PyCharm 2021.2.3, Anaconda Navigator 2.1.0, ADS 2020; the test equipment is: CPU: i7-6700HQ, GPU: GTX 1060, memory: 16GB; the equivalent circuit model of the fourth-order cavity filter is as follows: Figure 5 As shown, its ideal S-parameters are as follows: Figure 6 As shown, for this fourth-order cavity filter, with a center frequency of 1.95 GHz, a bandwidth of 140 MHz, and containing two transmission zeros, its ideal coupling matrix M is shown in Table 3:
[0097] Table 3
[0098]
[0099]
[0100] Based on the ideal coupling matrix of the filter, it is detuned in ADS2020. The detuned coupling matrix M is shown in Table 4.
[0101] Table 4
[0102] Matrix S 1 2 3 4 L S 0 1.2335 0 0 0 0 1 1.2335 0 1.1325 0 0.4213 0 2 0 1.1325 0 0.5825 0 0 3 0 0 0.5825 0 0.9862 0 4 0 0.4213 0 0.9862 0 1.0892 L 0 0 0 0 1.0892 0
[0103] The S-parameters of the filter in the detuned state are as follows: Figure 7 As shown, the S-parameters of the filter in the detuned state were extracted, and its coupling matrix was extracted using this invention. The dataset construction took 10 seconds, and the neural network training consisted of 2000 rounds, taking 200 seconds. The mean absolute error (MAELoss) on the test set was 0.0008359. Table 5 shows a comparison between the extracted non-zero terms of the coupling matrix and its true non-zero terms.
[0104] Table 5
[0105] MS1 M12 M23 M34 M4L M14 True value 1.2335 1.1325 0.5825 0.9862 1.0892 0.4213 Extraction value of this invention 1.2326 1.1326 0.5825 0.9873 1.0900 0.4212 Percentage error 0.0730％ 0.0088％ 0.0000％ 0.1115％ 0.07345％ 0.0237％
[0106] At the same time, such as Figure 8 The image shows the application of the present invention to... Figure 5 The S-parameter curves of the fourth-order cavity filter after coupling matrix extraction and their comparison with the true S-parameter curves are shown in Table 5. Figure 8 As shown, the microwave filter coupling matrix parameters extracted by this invention based on the S-parameters have a very small error compared with the actual coupling matrix parameters, and the corresponding S-parameters have good consistency with the real S-parameter curves, achieving high accuracy.
[0107] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All disclosed features, or steps in all methods or processes, may be combined in any way except for mutually exclusive features and / or steps.
Claims
1. A method for extracting the coupling matrix of a microwave filter based on a neural network, characterized in that, Includes the following steps: S1: Establish a forward dataset extraction program for calculating S-parameters using the microwave filter coupling matrix; the specific process is as follows: S1.1: Let the coupling matrix to be solved be... The corresponding order of the microwave filter is Bandwidth is The center frequency is Quality factor is Solve for the initial frequency. The solution termination frequency is Solve for the frequency interval as ; S1.2: Calculate the order : Number of frequency solutions : ; S1.3: Construction 1-th order matrix : ， ； S1.4: Construction 1-th order matrix : ； S1.5: For each solution frequency Normalize: , for The normalization result; S1.6: Calculate the solution frequency impedance matrix at : ,in, If the imaginary unit is used, then the impedance matrix... The inverse of ; S1.7: Let the matrix be... , For matrix Submatrix: ， ； S1.8: Calculate the solution frequency place for ; S1.9: Calculate the solution frequency place for ; S1.10: Calculate the solution frequency Group latency ; S2: Extract the training dataset containing S-parameters and microwave filter coupling matrix through a program; S3: Construct a convolutional neural network model; S4: Train the convolutional neural network model using the training dataset; S5: Use the trained convolutional neural network model to perform inverse extraction of the microwave filter coupling matrix.
2. The microwave filter coupling matrix extraction method based on neural networks according to claim 1, characterized in that, The specific process of step S2 is as follows: S2.1: According to The design goal of the first-order filter is to synthesize its using the generalized Chebyshev method. ideal coupling matrix of order ; S2.2: In its ideal coupling matrix of Randomly generated within the range Group coupling matrix, where, This represents the possible detuning range of the filter; S2.3: Calculate using the positive dataset extraction procedure Group coupling matrix corresponding to , , Parameters, and the coupling matrix of each group , , The parameter combination is used as input, and the coupling matrix is used as the label vector to form training samples, thereby constructing the training dataset.
3. The microwave filter coupling matrix extraction method based on neural networks as described in claim 2, characterized in that, The training samples include the following two types: First type of training sample: The actual part, The imaginary part, The real part and The virtual part is used as a channel to form four channels of input data; The second type of training sample: , and and Each is treated as a channel, using the following method The data is presented in units, forming three channels of input data.
4. The microwave filter coupling matrix extraction method based on neural networks according to claim 1, characterized in that, In step S3, the convolutional neural network model includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third pooling layer, a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer, which are connected in sequence.
5. The microwave filter coupling matrix extraction method based on neural networks according to claim 4, characterized in that, The input layer accepts either four-channel or three-channel input data. The first and second convolutional layers use the ELU function as the activation function, with a kernel size of 3×1. The first, second, and third pooling layers all use max pooling, with a filter size of 2×1. The first fully connected layer has 1024 neurons, the second fully connected layer has 256 neurons, and the third fully connected layer has 64 neurons, all using the ELU function as the activation function. The output layer is a coupling matrix, output as a one-dimensional vector.
6. The microwave filter coupling matrix extraction method based on neural networks according to claim 1, characterized in that, The specific training process for step S4 is as follows: S4.1: Use the input data from the training samples as the input to the neural network model, and the neural network model outputs the prediction vector; S4.2: Set the SmoothL1 loss function to the loss function of the neural network: ； in, It is the average of the absolute values of the element differences between the prediction vector and the label vector; S4.3: Set the target to terminate neural network training and complete model training: Stop training when the number of training rounds reaches the preset threshold or the loss function reaches the preset threshold. If the number of training rounds reaches the preset threshold but the loss function has not yet reached the preset threshold, repeat step S2, collect more sample data, and retrain until the loss function reaches the preset threshold.
7. The microwave filter coupling matrix extraction method based on neural networks according to claim 1, characterized in that, The specific process of step S5 is as follows: obtain the S-parameters of the microwave filter to be extracted through testing or simulation. The parameters constitute the input data of the neural network model, and the trained neural network model outputs the corresponding coupling matrix.