Microwave filter coupling matrix extraction method of deep separable convolutional neural network

Abstract

The invention discloses a microwave filter coupling matrix extraction method of a depth separable convolutional neural network. The method comprises the following steps: step 1, establishing microwave filter coupling matrix calculation parameters; 2, constructing a data set of parameters and a coupling matrix; step 3, constructing a deep separable convolutional neural network model; 4, training the depth separable convolutional neural network model by using the training data set; step 5, establishing a filter model based on electromagnetic simulation software HFSS to extract parameters, and removing phase offset from simulation parameters; step 6, extracting a coupling matrix; the problems of poor model robustness, low extraction precision, poor model generalization caused by phase inconsistency between data generated by a formula method and simulation / actual measurement data and the like existing in coupling matrix extraction of an existing method are solved.

Description

Technical Field

[0001] This invention belongs to the field of filter design technology, and in particular relates to a method for extracting the coupling matrix of a microwave filter using a deep separable convolutional neural network. Background Technology

[0002] Microwave filters play a crucial role in frequency selection in modern wireless communication systems, and their performance directly affects the overall communication quality. With the evolution of communication standards, filter structures are becoming increasingly complex, and performance requirements are constantly rising. Traditional microwave filter design methods heavily rely on engineers' experience and repeated electromagnetic simulations (such as HFSS and CST). Although electromagnetic simulations offer high accuracy, each simulation is time-consuming, and the design optimization process typically requires multiple iterations to adjust geometric parameters and extract coupling matrices to evaluate performance, resulting in a long overall design cycle and high cost.

[0003] The coupling matrix is ​​a key mathematical model characterizing the topology and performance of microwave filters. During the design process, it is often extracted from actual or simulated data. The coupling matrix corresponding to the scattering parameters is compared with the coupling matrix of the ideal target to identify the deviation, thereby guiding the adjustment of geometric parameters. Currently, the mainstream coupling matrix extraction methods are mainly divided into two categories: analytical methods (such as Cauchy's method and vector fitting method) and optimization methods. Analytical methods are fast, but have specific requirements for the filter topology, poor versatility, and low accuracy in the out-of-band region; optimization methods are highly versatile and accurate, but involve large computational loads, slow convergence speed, and are extremely time-consuming.

[0004] In recent years, artificial intelligence technology, especially deep learning neural networks, has shown great potential in solving complex nonlinear mapping problems. Existing research has attempted to apply neural networks to the modeling, synthesis, and optimization of microwave devices. For example, neural networks have been used to establish filter geometric parameters... Parameter mapping is used to replace some simulations; some studies also explore the use of neural networks for coupled matrix synthesis or from The parameter extraction coupling matrix has achieved certain results, proving the feasibility of this approach. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for extracting the coupling matrix of a microwave filter from a deep separable convolutional neural network, so as to solve the problems of poor model robustness, low extraction accuracy, and poor model generalization caused by the phase inconsistency between the data generated by the formula method and the simulation / measured data in the existing methods.

[0006] Technical solution of the present invention:

[0007] A microwave filter coupling matrix extraction method of a deep separable convolutional neural network, the method comprising:

[0008] Step 1, establishing microwave filter coupling matrix calculation parameters;

[0009] Step 2, constructing parameters and the data set of the coupling matrix;

[0010] Step 3, constructing a deep separable convolutional neural network model;

[0011] Step 4, training the deep separable convolutional neural network model using the training data set;

[0012] Step 5, establishing filter model extraction parameters based on electromagnetic simulation software HFSS, and removing phase offset from simulation parameters;

[0013] Step 6, coupling matrix extraction.

[0014] The method for establishing microwave filter coupling matrix calculation parameters in step 1 comprises:

[0015] Step 1.1, setting the order of the microwave filter to be designed as , the bandwidth as , the center frequency as , the starting frequency as , the ending frequency as , the frequency interval as , the coupling matrix as , the fixed frequency point number as , and the input and output ends coupled with resonant cavities and ;

[0016] Step 1.2, for the coupling matrix of type +2, the order of the coupling matrix corresponding to the order filter is +2, and the frequency solving number is : ;

[0017] Step 1.3, solving at the frequency : , representing the th frequency sampling point, is an imaginary unit, indicates the microwave filter The reflection coefficient of port one in the parameters, Represents frequency Normalized complex frequency at that point Representing frequency Place The elements in the first row and first column of the inverse matrix. represent A diagonal matrix of order 1. The rest are 1. represent A zero matrix of order 1. , The rest are 0. express Order coupling matrix;

[0018] Step 1.4: Solve for the frequency place : , Represents microwave filter The transmission coefficient from port 2 to port 1 in the parameters;

[0019] Step 1.5: Solve for the frequency Normalized complex frequency at : ;

[0020] Step 1.6: Solve for the frequency Normalized frequency ω: ;

[0021] Step 1.7: Solve for the frequency place : ;

[0022] Step 1.8, Construction 1-th order matrix :

[0023] ;

[0024] Step 1.9, Construction 1-th order matrix :

[0025] ;

[0026] Step 1.10, for Type coupling matrix calculation parameter, =1, =1.

[0027] The construction described in step 2 Methods for handling datasets of parameters and coupling matrices include:

[0028] Step 2.1, according to First-order filter design specifications, using Software acquisition ideal coupling matrix of order ;

[0029] Step 2.2: For the ideal coupling matrix Perturb the non-zero elements on the off-diagonal and the elements on the main diagonal excluding the two ends. , The error between the value and the ideal value is randomly generated. Group coupling matrix;

[0030] Step 2.3, Calculation Group coupling matrix corresponding to and parameter, Represents microwave filter The reflection coefficient of port one in the parameters, Represents microwave filter The transmission coefficients from port 2 to port 1 in the parameters are stored in array format. There are non-zero elements in the coupling matrix, and the dimension of the coupling matrix is... , Parameters at each frequency point Real part virtual part Real part and The imaginary part is stored in a two-dimensional array format [frequency points, [Parameters] Save, using a fixed frequency point count Perform frequency sampling. The parameter dimension is ;

[0031] Step 2.4, Group coupling matrix and Dimensional The parameters are stored as an array as a training dataset, used to train a depthwise separable convolutional neural network model.

[0032] Methods for constructing deep separable convolutional neural network models include:

[0033] Step 3.1: The input layer of the depthwise separable convolutional neural network model 2D-DSCNN creates a channel for input. Dimensional Two-dimensional parameter data;

[0034] Step 3.2: The depthwise separable convolutional neural network model 2D-DSCNN is connected in the following order: input layer, depthwise convolutional layer 1, pointwise convolutional layer 1, batch normalization layer 1, max pooling layer 1, depthwise convolutional layer 2, pointwise convolutional layer 2, batch normalization layer 2, max pooling layer 2, flattening layer, fully connected layer 1, fully connected layer 2, fully connected layer 3, and output layer.

[0035] Step 3.3, the input layer will The parameters are used as one-channel input data. After batch normalization and fully connected layers, the SELU function is used as the activation function. The first depthwise convolutional layer uses a 3×3 kernel with 1 channel, keeping the channel count constant. The first pointwise convolutional layer uses a 1×1 kernel with 16 output channels. The batch normalization layer does not change the channel count. The max pooling layer has a pooling window size of (2, 2) and a stride of 2 in both the height and width dimensions. The second depthwise convolutional layer uses a 5×3 kernel with 16 channels, keeping the channel count constant. The second pointwise convolutional layer uses a 1×1 kernel with 32 output channels. The flattening layer flattens the 32×150×1 data after dimensionality reduction by the max pooling layer into 4800 neurons for input to the fully connected layers. Fully connected layer one has 128 neurons, fully connected layer two has 64 neurons, and fully connected layer three has 32 neurons. The output layer has... Each neuron represents a vector of elements of the coupling matrix. The vector has 1,000 elements and is in the form of a one-dimensional vector.

[0036] Methods for training deep separable convolutional neural network models using training datasets include:

[0037] Step 4.1: Connect the coupling matrix with... The training dataset of the parameters is used as input to the deep separable convolutional neural network model 2D-DSCNN, and the model outputs the predicted elements of the coupling matrix.

[0038] Step 4.2: Set AdamW+L2 regularization as the optimizer for the convolutional data network. The update process is as follows:

[0039] Calculate the gradient: ;

[0040] Indicates a time step. Represents the gradient. Represents network model parameters, Represents the loss function. Indicates the parameter Find the gradient. Representation function about The gradient;

[0041] Update the first moment: ;

[0042] This represents the first moment estimate at the current time step. This represents the first-order moment decay rate. This represents the first moment estimate of the previous time step;

[0043] Update the second moment: ;

[0044] The second moment estimate represents the current time step. This represents the second-order moment decay rate. This represents the second moment estimate of the previous time step;

[0045] Bias correction: ;

[0046] and These represent the first moment and the second moment after deviation correction, respectively;

[0047] Parameter update:

[0048] ;

[0049] =0.9, =0.999, Control the step size of parameter updates. Represents the L2 regularization strength. and All , It is a small constant that prevents division by zero;

[0050] Step 4.3: Set the MSE loss function to the loss function of the convolutional data network:

[0051]

[0052] Represents the number of samples. Representing the One predicted value, Representing the One true value;

[0053] Step 4.4: The training process involves initializing configuration parameters, including the learning rate. Number of training samples The maximum number of training epochs, the shape of the model input data, and the dimensions of the model output layer;

[0054] Step 4.5: Create a save directory to save the trained model files, load the training dataset, and randomly divide it into training and test sets in an 8:2 ratio;

[0055] Step 4.6: Load the deepest separable convolutional neural network model onto the GPU, initialize the optimizer, select the AdamW optimizer and add regularization, and set the initial learning rate to [value missing]. Regularization strength for ;

[0056] Step 4.7: Set up the learning rate scheduler to dynamically adjust the learning rate based on the validation loss, set up the data loader to support batch training, and initialize the training monitor to record the loss and accuracy metrics during the training process.

[0057] Step 4.8: Start the training loop. Input the S-parameter data of batch size × number of channels × height × width into the model. Calculate the MSE loss between the predicted and true values ​​using forward propagation. Calculate the gradient using backpropagation and update the model parameters through the AdamW+L2 optimizer. Record the training loss and coefficient of determination. After each training round, the model performance is evaluated on the validation set. If the validation loss does not improve for five consecutive rounds, the learning rate is adjusted.

[0058] Step 4.9, Validation Phase: Input batch number × number of channels × height × width into the model. Using the parameter data, forward propagation is used to calculate the MSE loss between the predicted and actual values, and the training loss and coefficient of determination are recorded. ;

[0059] Step 4.10: Save the model to the specified directory, ensuring that the model file name includes the training epochs and validation loss values ​​for subsequent analysis;

[0060] Step 4.11: Print the progress information and update the loss curve after each training round.

[0061] Step 4.12: Determine if the training has reached the maximum number of rounds. If not, restart the training loop. If the maximum number of rounds has been reached, end the training, save the last trained model, and display the training loss curve.

[0062] Step 5 involves establishing a filter model based on the electromagnetic simulation software HFSS for extraction. Parameters, and for simulation Methods for removing phase shift from parameters include:

[0063] Step 5.1: Establish a filter model with a topology corresponding to the ideal coupling matrix;

[0064] Step 5.2: Extract the filter. Parameter data;

[0065] Step 5.3, Filter The parameters can be represented by rational functions, and here we use vector fitting to approximate them. Considering the phase shift and transmission lines, the entire model's... port The parameters should be:

[0066]

[0067] in, It is the first Approximate reflection coefficient of the port, , Represents the normalized complex frequency. It is the imaginary unit. It is the normalized frequency. It is the filter order. It is an additional order, representing the zeros and poles introduced by phase shift and transmission line. As the extreme point, To leave a residue For constant terms, Assuming zeros are the origin, the zeros and poles are divided into two groups, separating the zeros and poles that are far from the origin. and These are the number of zeros and poles introduced by phase shift and transmission line, respectively;

[0068] Step 5.4, the normalized frequency is expressed as:

[0069] For removal The frequency corresponding to the parameter phase shift effect;

[0070] Step 5.4 The zeros are the eigenvalues ​​of the following matrices:

[0071] yes( A diagonal matrix of order 1, wherein the diagonal elements are poles. , It is a completely one vector. It is a collection of residues The row vector, Indicates matrix transpose;

[0072] Step 5.5: Extract the zeros and poles far from the origin. :

[0073] in Called the phase factor, it is evaluated The phase can be estimated by the phase shift introduced by the transmission line;

[0074] Step 5.5: Remove phase shift The parameters are:

[0075] It is after removing the phase shift parameter, Simulation / Actual Measurement parameter, for

[0076]

[0077] and Related:

[0078] This indicates a phase calculation; pay attention to the calculations. The phase value is between Between these, some phase values ​​may be discontinuous, requiring an "unpacking" operation to make the phase continuous;

[0079] Step 5.6: Remove phase shift The parameters are stored as an array.

[0080] Coupling matrix extraction methods include:

[0081] Step 6.1: Import the trained deep separable convolutional neural network model;

[0082] Step 6.2: Load the test data and input the S-parameters after removing the phase shift into the depthwise separable convolutional neural network model;

[0083] Step 6.3: Preprocess the data by converting the array data into tensors, expanding the dimensions, adding channel dimensions, using 32-bit floating-point data type, and transferring the data processing to the GPU for accelerated computation.

[0084] Step 6.4: Use a deep separable convolutional neural network to predict the input S-parameters and obtain the corresponding coupling matrix;

[0085] Step 6.5: Store the vectors predicted by the model in an array, rearrange the array elements according to their indices, and construct a symmetric matrix;

[0086] Step 6.6: Compare the coupling matrix predicted by the model with the theoretical value, calculate the difference between each element in the matrix, and adjust the geometric parameters according to the correspondence between the elements of the coupling matrix and the filter geometric parameters.

[0087] Step 6.7: Calculate the prediction coupling matrix corresponding to... Parameters, and plot The parameter curves and the un-phase-shifted data extracted by HFSS Compare the original parameter data curves to observe the calculation of the prediction coupling matrix. Can the parameter curve fit the simulation? Parameter curve.

[0088] The beneficial effects of this invention are:

[0089] In the data preparation stage, this invention introduces data constraints by referencing the distribution of simulation data, ensuring that the training dataset conforms to the simulation data distribution. In the model construction stage, a two-dimensional deep separable convolutional neural network is used, enabling more effective feature extraction; the addition of batch normalization layers makes the data distribution in the network model more even and stable; and the use of SELU as the activation function ensures that the outputs of each layer of the neural network maintain a distribution with a mean of 0 and a variance of 1. In the model training stage, the traditional Adam optimizer is replaced with the AdamW+L2 optimizer, improving the model's generalization performance and accuracy.

[0090] To address the issue that the results obtained after training a model using a dataset generated by the formula method differ significantly from the results of simulation / measured data (including phase shift), a method is adopted to eliminate the phase shift effect in the simulation / measured data before inputting it into the model for coupling matrix prediction.

[0091] This invention is based on the traditional method Parameter conversion Based on the parameters and the extraction of the coupling matrix through Cauchy method or vector fitting method, a two-dimensional deep separable convolutional neural network is introduced. Only data is needed to train the network model to predict the coupling matrix of the microwave filter, avoiding complex calculations. It is not only convenient but also highly accurate.

[0092] After training is complete and the model file is obtained, the simulation will be performed after removing the phase shift. The parameters are input into the network to load the model file for prediction, and the predicted coupling matrix is ​​obtained. By comparing it with the ideal coupling matrix, the filter geometric parameters corresponding to the elements in the coupling matrix are adjusted. The geometric parameters are continuously adjusted in a loop until the filter geometric parameters that meet the design specifications are finally obtained.

[0093] This method addresses the problems of poor model robustness and low extraction accuracy in existing methods for coupling matrix extraction, as well as poor model generalization caused by phase inconsistency between the data generated by the formula method and the simulation / measured data. Attached Figure Description

[0094] Figure 1 The ideal coupling matrix corresponding to this invention Parameter indicators;

[0095] Figure 2 This is the simulation / measurement model of the two-port filter of this invention;

[0096] Figure 3 This is an example simulation model of the fourth-order filter of the present invention;

[0097] Figure 4 This is a test diagram of the model of the present invention against simulation data;

[0098] Figure 5 This is the final optimized result of the present invention. Detailed Implementation

[0099] A method for extracting the coupling matrix of a microwave filter from a deep separable convolutional neural network includes:

[0100] S1: Establishing the microwave filter coupling matrix calculation The procedure for parameters;

[0101] S2: Construction Dataset of parameters and coupling matrices;

[0102] S3: Construct a deep separable convolutional neural network model (2D-DSCNN);

[0103] S4: Train the deep separable convolutional neural network model using the training dataset;

[0104] S5: Filter model extraction based on electromagnetic simulation software HFSS Parameters, and for simulation The parameters are used to remove the phase shift;

[0105] S6: Coupling matrix extraction;

[0106] The specific process of step S1 is as follows:

[0107] S1.1: The order of the microwave filter to be designed is Bandwidth is The center frequency is Solve for the initial frequency as The solution termination frequency is The frequency interval is The coupling matrix is The fixed number of solution frequency points is The input and output terminals are coupled to the resonant cavity. , ;

[0108] S1.2: For +2 type coupling matrix, The order of the coupling matrix corresponding to the first-order filter is = +2, frequency solver : ;

[0109] S1.3: Solving for frequencies place : , Representing the At each frequency sampling point, It is the imaginary unit. Represents microwave filter The reflection coefficient of port one in the scattering parameters. Represents frequency Normalized complex frequency at that point Representing frequency Place The elements in the first row and first column of the inverse matrix. represent A diagonal matrix of order 1. The rest are one. represent A zero matrix of order 1. , The rest are 0. express Order coupling matrix;

[0110] S1.4: Solving for frequencies place : , Represents microwave filter The transmission coefficient from port 2 to port 1 in the parameters (scattering parameters);

[0111] S1.5: Solving for frequencies Normalized complex frequency at : ;

[0112] S1.6: Solving for frequencies Normalized frequency ω: ;

[0113] S1.7: Solving for frequencies place : ;

[0114] S1.8: Construction 1-th order matrix :

[0115] ;

[0116] S1.9: Construction 1-th order matrix :

[0117] ;

[0118] S1.10: For Type coupling matrix calculation parameter, =1, =1;

[0119] The specific process of step S2 is as follows:

[0120] S2.1: According to Design specifications for order filters, such as the order of a microwave filter. Bandwidth is The center frequency is etc., use Software acquisition ideal coupling matrix of order In a typical coupling matrix, the non-zero elements off-diagonal represent the coupling strength of each resonant cavity, the elements on the main diagonal represent the resonance of each resonant cavity, and deviations from the ideal value represent deviations from the resonant frequency.

[0121] S2.2: For the ideal coupling matrix Perturb the non-zero elements on the off-diagonal and the elements on the main diagonal excluding the two ends. , The error between the value and the ideal value is randomly generated. Group coupling matrix, This represents the number of coupling matrices generated;

[0122] S2.3: Calculate using the coupling matrix Parameter calculation program Group coupling matrix corresponding to , parameter, Represents microwave filter The reflection coefficient of port one in the scattering parameters. Represents microwave filter The transmission coefficients from port 2 to port 1 in the scattering parameters are stored in an array. The coupling matrix has non-zero elements, so the dimension of the coupling matrix is... , Parameters (scattering parameters) at each frequency point Real part virtual part Real part The imaginary part is stored in a two-dimensional array format [frequency points, [Parameters] Save, using a fixed frequency point count Frequency sampling is performed, so The parameter dimension is ;

[0123] S2.4: Will Group coupling matrix and Dimensional The parameters are stored as an array as the training dataset, which is used to train the depthwise separable convolutional neural network model (2D-DSCNN).

[0124] The specific process of step S3 is as follows:

[0125] S3.1: As can be seen from S2.3, the generated... The parameters are in two-dimensional array format [frequency points, [Parameters] Saved, dimension is In total Group, each group The parameter data corresponds to a set of coupling matrix arrays, so the input layer of a depthwise separable convolutional neural network model (2D-DSCNN) creates one channel for input. Dimensional Two-dimensional parameter data;

[0126] S3.2: The depthwise separable convolutional neural network model (2D-DSCNN) is connected in the following order: input layer, depthwise convolutional layer 1, pointwise convolutional layer 1, batch normalization layer 1, max pooling layer 1, depthwise convolutional layer 2, pointwise convolutional layer 2, batch normalization layer 2, max pooling layer 2, flattening layer, fully connected layer 1, fully connected layer 2, fully connected layer 3, and output layer;

[0127] S3.3: The input layer will The parameters are used as one-channel input data. After the batch normalization layer and the fully connected layer, the SELU function is used as the activation function. The depthwise convolutional layer uses a 3×3 kernel size and 1 channel, keeping the total number of channels unchanged. The pointwise convolutional layer uses a 1×1 kernel size and 16 output channels. The batch normalization layer does not change the number of channels but can stabilize the distribution of the output data, alleviating the vanishing and exploding gradient problems and reducing the risk of overfitting. The max pooling layer has a pooling window size of (2,2...). The stride in both height and width dimensions is 2. The second depthwise convolutional layer uses a kernel size of 5×3 and 16 channels, without changing the number of channels. The second pointwise convolutional layer uses a kernel size of 1×1 and 32 output channels. The flattening layer flattens the data after dimensionality reduction by the max pooling layer (32×150×1) into 4800 neurons, which are used as input to the fully connected layers. The first fully connected layer has 128 neurons, the second has 64 neurons, and the third has 32 neurons. The output layer has... Each neuron represents a vector of elements of the coupling matrix. Each element is a one-dimensional vector;

[0128] The specific process of step S4 is as follows:

[0129] S4.1: Connect the coupling matrix with The training dataset of parameters is used as input to a deep separable convolutional neural network model (2D-DSCNN), and the model outputs the predicted elements of the coupling matrix.

[0130] S4.2: Set AdamW+L2 regularization as the optimizer for the convolutional data network. Its update process is as follows:

[0131] Calculate the gradient:

[0132] ;

[0133] Represents a time step. Represents gradient, Represents network model parameters, Represents the loss function. Indicates the parameter Find the gradient. Representation function about The gradient;

[0134] Update the first moment:

[0135] ;

[0136] Let represent the first moment estimate at the current time step, and represent the exponential moving average of the gradient. This represents the first-order moment decay rate. This represents the first moment estimate of the previous time step;

[0137] Update the second moment:

[0138] ;

[0139] The second moment estimate represents the current time step, and the exponential moving average of the squared gradient is denoted as . This represents the second-order moment decay rate. This represents the second moment estimate of the previous time step;

[0140] Bias correction:

[0141] ;

[0142] and These represent the first moment and the second moment after deviation correction, respectively;

[0143] Parameter update (including decoupling L2):

[0144] ;

[0145] =0.9, =0.999, Control the step size of parameter updates. Represents the L2 regularization strength. and All , It is a small constant that prevents division by zero, and its value is ;

[0146] S4.3: Set the MSE loss function to the loss function of the convolutional data network:

[0147]

[0148] in, Represents the number of samples. Representing the One predicted value, Representing the One true value;

[0149] S4.4: The training process initializes configuration parameters, including the learning rate. Number of training samples The maximum number of training epochs, the shape of the model input data, and the dimensionality of the model output layer, etc.

[0150] S4.5: Create a save directory to save the trained model files, load the training dataset generated by S2, and divide it into two parts, training set and test set, randomly divided in an 8:2 ratio;

[0151] S4.6: Construct the convolutional neural network model structure described in S3 and load the model onto the GPU. Initialize the optimizer, select the AdamW optimizer and add regularization, and set the initial learning rate to... Regularization strength for ;

[0152] S4.7: Set up the learning rate scheduler to dynamically adjust the learning rate based on the validation loss, set up the data loader to support batch training, and initialize the training monitor to record the loss and accuracy metrics during the training process.

[0153] S4.8: Start the training loop, enter the training phase, input the S-parameter data of batch × number of channels × height × width into the model, calculate the MSE (mean squared error) loss between the predicted and true values ​​through forward propagation, calculate the gradient through backpropagation and update the model parameters through the AdamW+L2 optimizer, and record the training loss and... , The coefficient of determination is 1. The closer it is to 1, the better the model's performance. The model's performance is evaluated on the validation set after each training round. If the validation loss does not improve for five consecutive rounds, the learning rate is adjusted.

[0154] S4.9: In the validation phase, input batch × number of channels × height × width into the model. Parameter data, forward propagation calculates the MSE (mean squared error) loss between predicted and true values, and records the training loss and... (Determination coefficient);

[0155] S4.10: Save the model. Save the model to the specified directory, ensuring that the model file name includes the training epochs and validation loss values ​​for subsequent analysis.

[0156] S4.11: Update the loss curve and print progress information. Update the loss curve after each training round.

[0157] S4.12: Determine if the training has reached the maximum number of rounds. If not, restart the training loop and return to S4.8 to continue execution. If the maximum number of rounds has been reached, end the training, save the last trained model, and display the training loss curve.

[0158] The specific process of step S5 is as follows:

[0159] S5.1: Establish a filter model with a topology corresponding to the ideal coupling matrix from HFSS; S5.2: Extract the filter's... Parameter data;

[0160] S5.3, Filter The parameters can be represented by rational functions, and here we use vector fitting to approximate them. Considering the phase shift and transmission lines, the entire model's... port The parameters should be:

[0161]

[0162] in, It is the first Approximate reflection coefficient of the port, , Represents the normalized complex frequency. It is the imaginary unit. It is the normalized frequency. It is the filter order. It is an additional order, representing the zeros and poles introduced by phase shift and transmission line. As the extreme point, To leave a residue For constant terms, Assuming zeros are the origin, the zeros and poles are divided into two groups, separating the zeros and poles that are far from the origin. and These are the number of zeros and poles introduced by phase shift and transmission line, respectively;

[0163] S5.4, Normalized frequency is expressed as:

[0164] For removal The frequency corresponding to the parameter phase shift effect;

[0165] S5.4 The zeros are the eigenvalues ​​of the following matrices:

[0166] yes( A diagonal matrix of order 1, wherein the diagonal elements are poles. , It is a completely one vector. It is a collection of residues The row vector, Indicates matrix transpose;

[0167] S5.5 Extracting zeros and poles far from the origin :

[0168] in Called the phase factor, it is evaluated The phase can be estimated by the phase shift introduced by the transmission line;

[0169] S5.5, after removing phase shift The parameters are:

[0170] It is after removing the phase shift parameter, Simulation / Actual Measurement parameter, for

[0171]

[0172] and Related:

[0173] This indicates a phase calculation; pay attention to the calculations. The phase value is between Between these, some phase values ​​may be discontinuous, requiring an "unpacking" operation to make the phase continuous;

[0174] S5.6, Remove phase shift The parameters are stored as an array.

[0175] Furthermore, the specific process of step S6 is as follows:

[0176] S6.1: Import the trained deep separable convolutional neural network model (2D-DSCNN).

[0177] S6.2: Load the test data and input the S-parameters after removing the phase shift into the depthwise separable convolutional neural network model (2D-DSCNN);

[0178] S6.3: Preprocess the data, convert array data into tensors, expand dimensions, add channel dimensions, use 32-bit floating-point data type, and transfer data processing to the GPU for accelerated computation;

[0179] S6.4: The input S-parameters are predicted using a deep separable convolutional neural network to obtain the corresponding coupling matrix;

[0180] S6.5: Reconstruct the coupling matrix by storing the vectors predicted by the model in an array, rearranging the array elements according to their indices, and constructing a symmetric matrix.

[0181] S6.6: Coupling matrix comparison analysis compares the coupling matrix predicted by the model with the theoretical value, calculates the difference between each element in the matrix, and adjusts the geometric parameters according to the correspondence between the elements of the coupling matrix and the filter geometric parameters, giving priority to adjusting the geometric parameters corresponding to the elements of the coupling matrix with larger differences.

[0182] S6.7: Parameter visualization, calculated using the S1 coupling matrix. The program calculates the parameters corresponding to the prediction coupling matrix. Parameters, and plot The parameter curves and the un-phase-shifted data extracted by HFSS Compare the original parameter data curves to observe the calculation of the coupling matrix predicted by the model. Can the parameter curve fit the simulation? Parameter curve.

[0183] A practical example illustrates the design of a fourth-order cavity filter, which includes:

[0184] S1: Establishing the microwave filter coupling matrix calculation The procedure for parameters (scattering parameters);

[0185] S1.1: This example uses the design of a microwave filter of order [order missing]. =4, bandwidth is =110MHz, center frequency is =2069.3MHz, the starting frequency is to be determined. =1869.3MHz, the solution termination frequency is =2269.3MHz, frequency interval is The coupling matrix is The fixed solution frequency point is =600, input and output terminals are coupled to the resonant cavity. =1、 =1, return loss The effectiveness of the proposed method was verified using a fourth-order cavity filter with a voltage ≤-20dB. Key geometric parameters included the heights of the four resonant cavities (…). , , , ) and the height of the coupling window between the three cavities ( , , ).

[0186] The fourth-order filter corresponds to +2 type ideal coupling matrix As shown in Table 1, the corresponding Parameter indicators such asFigure 1 As shown;

[0187] Table 1

[0188] Matrix 0 1 2 3 4 5 0 0 1.0352 0 0 0 0 1 1.0352 0 0.91058 0 0 0 2 0 0.91058 0 0.69992 0 0 3 0 0 0.69992 0 0.91058 0 4 0 0 0 0.91058 0 1.0352 5 0 0 0 0 1.0352 0

[0189] S1.2: Solving for frequencies place : ;

[0190] S1.3: Solving for frequencies place : ,

[0191] For detailed derivation process, please refer to step S1 of the instruction manual;

[0192] S1.4: Coupling Matrix Calculation The parameter program has been built.

[0193] S2: Built via program Training dataset for parameters and coupling matrices:

[0194] S2.1: Based on the design specifications of the 4th-order filter in step S1.1, use... Software acquisition +2nd order ideal coupling matrix As shown in Table 1;

[0195] S2.2: For the ideal coupling matrix The non-zero elements on the off-diagonal and the elements on the main diagonal excluding the two ends are perturbed by ±5%, and then randomly generated using a uniform distribution. =100,000 sets of coupling matrices;

[0196] S2.3: Calculate using the coupling matrix Parameter calculation program Group coupling matrix corresponding to , Parameters are stored as an array. =9 coupling matrix elements (as shown in Table 1) , , , , , , , and Therefore, the dimension of the coupling matrix is , Parameters at each frequency point Real part virtual part Real part The imaginary part is stored in a two-dimensional array format [frequency points, [Parameters] Save, use fixed frequency points Frequency sampling is performed, so The parameter dimension is ;

[0197] S2.4: Connect the coupling matrix with The parameters are stored as an array as a training dataset for training a depthwise separable convolutional neural network model (2D-DSCNN).

[0198] S3: Construct a deep separable convolutional neural network model (2D-DSCNN);

[0199] S3.1: The parameters are data stored in a two-dimensional array format. They are treated as a set of data and correspond to a set of coupling matrix arrays. Therefore, the input layer of the depthwise separable convolutional neural network model (2D-DSCNN) creates a channel for inputting two-dimensional data. So the data input to the network is [batch, number of channels, height, width]. The batch is set to 64, the number of channels is 1, the height is 600, and the width is 4.

[0200] S3.2: The detailed structure of the Deeply Separable Convolutional Neural Network (2D-DSCNN) model is shown in Table 2. This model can achieve... Mapping of parameters (scattering parameters) to coupling matrix;

[0201] Table 2

[0202] Layer Type Size / Parameter Stride Output Shape Activation Function Input - - 1 x 600 x 4 - Depthwise Convolution Kernel = (3, 3), Groups = 1 1 1 x 600 x 4 - Pointwise Convolution 1 x 1, Output = 16 1 16 x 600 x 4 - Batch Normalization 16 - 16 x 600 x 4 SELU Max Pooling Kernel = (2, 2) (2,2) 16 x 300 x 2 - Depthwise Convolution Kernel = (5, 3), Groups = 16 1 16 x 300 x 2 - Pointwise Convolution 1 x 1, Output = 32 1 32 x 300 x 2 - Batch Normalization 32 - 32 x 300 x 2 SELU Max Pooling Kernel = (2, 2) (2,2) 32 x 150 x 1 - Flatten - - 4800 - Fully Connected Layer 4800 -> 1 - 128 SELU Fully Connected Layer 128 -> 64 - 64 SELU Fully Connected Layer 6-> 32 - -> SELU Output Layer 32 -> 9 - 9 -

[0203] S4: Train the deep separable convolutional neural network model using the training dataset;

[0204] S4.1: Connect the coupling matrix with The training dataset of parameters is used as input to a deep separable convolutional neural network model (2D-DSCNN), and the model outputs the predicted elements of the coupling matrix.

[0205] S4.2: Set AdamW+L2 regularization as the optimizer for convolutional data networks;

[0206] S4.3: Set the MSE loss function (mean squared loss) to the loss function of the convolutional data network;

[0207] S4.4: Initialize configuration parameters;

[0208] S4.5: Create a save directory to save the trained model files, add the training dataset, and divide it into training set and test set, randomly divided in an 8:2 ratio;

[0209] S4.6: Construct the convolutional neural network model structure described in S3 and load the model onto the GPU, and initialize the optimizer;

[0210] S4.7: Set up the learning rate scheduler to dynamically adjust the learning rate based on the validation loss, set up the data loader to support batch training, and initialize the training monitor to record the loss and accuracy metrics during the training process.

[0211] S4.8: Start the training loop, enter the training phase, and input batch × number of channels × height × width into the model. The parameters are used to calculate the MSE loss between the predicted and actual values ​​via forward propagation, and the gradient is calculated via backpropagation. The model parameters are then updated using the AdamW optimizer. The training loss is recorded. The coefficient of determination is a metric used to evaluate model performance on the validation set after each training round. If the validation loss does not improve for five consecutive rounds, the learning rate is adjusted.

[0212] S4.9: In the validation phase, input batch × number of channels × height × width into the model. Parameter data, forward propagation calculates the MSE loss between predicted and true values, and records the training loss and... (Determination coefficient);

[0213] S4.10: Save the model. Save the model to the specified directory, ensuring that the model file name includes the training epochs and validation loss values ​​for subsequent analysis.

[0214] S4.11: Update the loss curve and print progress information. Update the loss curve after each training round.

[0215] S4.12: Determine if the training has reached the maximum number of rounds. If not, restart the training loop and return to S4.8 to continue execution. If the maximum number of rounds has been reached, end the training, save the last trained model, and display the training loss curve.

[0216] S5: Extracting filter models from HFSS Parameters, and for simulation The parameters are used to remove the phase shift;

[0217] S5.1: Establish a filter model with a topology corresponding to the ideal coupling matrix from HFSS, as follows: Figure 3 As shown;

[0218] S5.2: Extracting the filter Parameter data;

[0219] S5.3: In microwave filter measurements, the measured values ​​may vary due to the influence of the test fixture or transmission line. The parameters may contain additional phase shifts. These phase shifts are not inherent characteristics of the filter itself and need to be removed for accurate filter parameter extraction (such as the coupling matrix).

[0220] S5.4: If the filter phase shift is as follows Figure 2 As shown, the simulation Parameters and removal of phase offset The parameters have the following relationship:

[0221]

[0222] The detailed derivation process is explained in step S5 of the instruction manual;

[0223] S5.5: Remove phase shift The parameters are stored in array format;

[0224] S6: Coupling matrix extraction;

[0225] S6.1: Import the trained deep separable convolutional neural network model (2D-DSCNN).

[0226] S6.2: Load data, removing phase shifts. The parameters are input into a depthwise separable convolutional neural network model (2D-DSCNN);

[0227] S6.3: Preprocess the data, convert the array data to a tensor, expand the dimensions, add channel dimensions, use 32-bit floating-point data type, and convert according to [batch, number of channels, height, width]. The parameter format is as follows: since there is only one set of data, the batch size is 1, the number of channels is 1, the height is 600, and the width is 4. The data processing is transferred to the GPU for accelerated calculation.

[0228] S6.4: Using a deep separable convolutional neural network to process the input... The parameters are used for model prediction to obtain the corresponding coupling matrix;

[0229] S6.5: Reconstruct the coupling matrix by storing the vectors predicted by the model in an array, rearranging the array elements according to their indices, and constructing a symmetric matrix.

[0230] S6.6: Coupling matrix comparison analysis compares the coupling matrix predicted by the model with the theoretical value, calculates the difference between each element in the matrix, and adjusts the geometric parameters according to the correspondence between the elements of the coupling matrix and the filter geometric parameters, giving priority to adjusting the geometric parameters corresponding to the elements of the coupling matrix with larger differences.

[0231] S6.7: Parameter visualization, calculated using the S1 coupling matrix. The program calculates the parameters corresponding to the prediction coupling matrix. Parameters, and plot The parameter curves and the un-phase-shifted data extracted by HFSS Compare the original parameter data curves to observe the calculation of the coupling matrix predicted by the model. Can the parameter curve fit the simulation? The parameter curves and their fitting results are as follows: Figure 4 As shown.

[0232] Continuously adjust geometric parameters and extract Parameters, phase shift removal, model prediction coupling matrix, comparison of the predicted coupling matrix with the ideal coupling matrix, and then continued adjustment of geometric parameters until... The parameter indicators meet the design requirements, such as Figure 5 As shown.

Claims

1. A method for extracting the coupling matrix of a microwave filter from a deep separable convolutional neural network, characterized in that: The method comprises: Step 1, Establishing the microwave filter coupling matrix calculation Parameters; Step 2, construction Parameters and dataset of the coupling matrix; Step 3, constructing a deep separable convolutional neural network model; Step 4, training the deep separable convolutional neural network model using a training data set; Step 5, Extracting parameters based on electromagnetic simulation software HFSS to establish filter model and removing phase offset from simulation parameters; Step 6, coupling matrix extraction.

2. The method of claim 1, wherein: The calculation of the microwave filter coupling matrix build-up described in step 1 The method of claim 1, wherein the parameters comprise: Step 1.1, let the order of the microwave filter to be designed be , the bandwidth be , the center frequency be , the start frequency be solved as , the end frequency be solved as , the frequency interval be , the coupling matrix be , the fixed number of frequency points to be solved be , the input and output ends be coupled with the resonant cavities and ; Step 1.2, for +2 type coupling matrix, The order of the coupling matrix corresponding to the order of the filter is = +2, the number of frequency solutions : ​ Step 1.3, solving for frequency at the : , represents the th frequency sample point, is the imaginary unit, denotes the reflection coefficient of port one in the microwave filter parameters, denotes the normalized complex frequency at the frequency represents the inverse matrix of the matrix, the element in the first row and first column, represents a diagonal matrix of order , the rest being 1, represents a zero matrix of order , , the rest being 0, denotes a coupling matrix of order Step 1.4, solving for frequency at the : , representing the transmission coefficient of port two to port one in the microwave filter parameters; Step 1.5, solving for frequency normalized complex frequency at the : ​ Step 1.6, solving for frequency Normalized frequency ω: ​ Step 1.7, solving for frequency at : ; Step 1.8, construction : matrix : ； Step 1.9, construction rank matrix : ； Step 1.10, for Type coupling matrix calculation Parameters, = 1, = 1.

3. The method of claim 1, wherein: the construction described in step 2 The method of parameterizing a dataset of coupling matrices comprises: Step 2.1, according to Filter design criteria, using Software to obtain Ideal coupling matrix ; Step 2.2, perturbing the ideal coupling matrix The non-zero elements off-diagonal are perturbed with the elements of the main diagonal except the two ends , representing the error from the ideal value, randomly generated group coupling matrix; Step 2.3, calculation The group coupling matrix corresponds to And Parameters, The microwave filter The reflection coefficient of port one in the parameter, The microwave filter The transmission coefficient from port two to port one in the parameter, saved in array form The number of non-zero elements of the coupling matrix, the dimension of the coupling matrix is , The parameter saves the real part and the imaginary part of the coupling matrix at each frequency point Real part, Imaginary part, Real part and Imaginary part, saved in two-dimensional array format [frequency point, Parameter] saved, with a fixed frequency point number Frequency sampling, The parameter dimension is ; Step 2.4, to The group coupling matrix is coupled with The dimension of The parameters are saved in an array as a training dataset for training the deep separable convolutional neural network model.

4. The method of claim 1, wherein: The method for constructing a deep separable convolutional neural network model comprises: Step 3.1, Input layer of the deep separable convolutional neural network model 2D-DSCNN creates one channel for input dimensional parametric two-dimensional data; Step 3.2, the deep separable convolutional neural network model 2D-DSCNN is sequentially connected in order as an input layer, a depth-wise convolutional layer one, a point-wise convolutional layer one, a batch normalization layer one, a maximum pooling layer one, a depth-wise convolutional layer two, a point-wise convolutional layer two, a batch normalization layer two, a maximum pooling layer two, a flattening layer, a fully connected layer one, a fully connected layer two, a fully connected layer three, and an output layer; Step 3.3, the input layer will The parameters are input as channel data, the SELU function is used as the activation function after the batch normalization layer and the full connection layer, the first depth convolution layer uses a convolution kernel size of 3x3, the channel number is 1, and the channel number is not changed, the first point convolution layer uses a convolution kernel size of 1x1, the output channel number is 16, the batch normalization layer does not change the channel number, the maximum pooling layer has a pooling window size of (2, 2), the step size in the height and width dimensions is 2, the second depth convolution layer uses a convolution kernel size of 5x3, the channel number is 16, and the channel number is not changed, the second point convolution layer uses a convolution kernel size of 1x1, the output channel number is 32, the flattening layer flattens the data 32x150x1 after dimensionality reduction by the maximum pooling layer into 4800 neurons, which is used as the input of the full connection layer, the first full connection layer has 128 neurons, the second full connection layer has 64 neurons, the third full connection layer has 32 neurons, and the output layer has neurons representing the elements of the coupling matrix element vector in the form of a one-dimensional vector.

5. The method of claim 1, wherein: The method for training the deep separable convolutional neural network model using a training data set comprises: Step 4.1, coupling matrix is coupled with The training dataset of parameters as input of the deep separable convolutional neural network model 2D-DSCNN, the predicted coupling matrix elements are output by the model; Step 4.2, setting AdamW+L2 regularization as the optimizer of the convolutional data network, and the updating process is as follows: Compute gradient: ​ denotes a time step, denotes a gradient, denotes a network model parameter, denotes a loss function, denotes a gradient of a function with respect to a parameter denotes a function with respect to a parameter; Updating the first moment: ; a first moment estimate representing a current time step, a first moment decay rate, a first moment estimate representing a previous time step; Updating the second moment: ; a second moment estimate representative of a current time step, denotes a second moment decay rate, denotes a second moment estimate of a previous time step; Bias correction: ​ with respectively represent the first moment after bias correction and the second moment after bias correction; Parameter updating: ； = 0.9, = 0.999, a step size for updating the control parameter, represents an L2 regularization strength, and are both , is a small constant to prevent division by zero; Step 4.3, setting the MSE loss function as the loss function of the convolutional data network: ； representative sample number, representative first forecast value, representative first true value; Step 4.4, the training process is configured with parameters including learning rate , the number of samples for batch training , the maximum number of rounds of training, the shape of the model input data, and the dimension of the model output layer Step 4.5, creating a saving directory for saving the trained model file, loading the training data set, and randomly dividing it into a training set and a test set in a ratio of 8:2; Step 4.6, Load the deep separable convolutional neural network model to the GPU, initialize the optimizer, select the AdamW optimizer and add regularization, set the initial learning rate to , the regularization strength is ; Step 4.7, setting a learning rate scheduler to dynamically adjust the learning rate according to the validation loss, setting a data loader to support batch training, and initializing a training monitor to record the loss and accuracy indicators during training; Step 4.8, Start training loop, input batch x channel number x height x width S parameter data to the model, forward propagation to calculate the MSE loss between the predicted value and the true value, backward propagation to calculate the gradient and update the model parameters by AdamW+L2 optimizer, record the training loss and the determination coefficient Evaluate the model performance on the validation set after each round of training, and adjust the learning rate if the validation loss does not improve for five consecutive rounds. Step 4.9, Validation Phase: Input batch number × number of channels × height × width into the model. Using the parameter data, forward propagation is used to calculate the MSE loss between the predicted and actual values, and the training loss and coefficient of determination are recorded. ; Step 4.10, saving the model to a specified directory, ensuring that the model file name contains the training round and the validation loss value for subsequent analysis; Step 4.11, printing progress information and updating the loss curve graph after each training round; Step 4.12, determining whether the training has reached the maximum number of rounds, and if not, restarting the training loop, if the maximum number of rounds is reached, ending the training, saving the last trained model, and displaying the training loss curve graph.

6. The method of claim 1, wherein: The filter model extraction based on the electromagnetic simulation software HFSS described in step 5 parameters, and a method for removing the phase offset from the simulation parameters includes: Step 5.1, establishing a filter model corresponding to the topology structure of the ideal coupling matrix; Step 5.2, extracting filter parameters from the parameter data; Step 5.3, Filter's Parameters can be represented by rational functions, here approximated using vector fitting, taking into account phase shifts and transmission lines, the whole model being Parameters of the port should be: Parameters should be: ； wherein, is the first approximate reflection coefficient of the port, , denotes the normalized complex frequency, is the imaginary unit, is the normalized frequency, is the filter order, is the additional added order, indicating the zero-pole introduced by phase shift and transmission line, is the pole, is the residue, is the constant term, is the zero, separating the zero-pole into two groups, separating the zero-pole far from the origin, and are the number of zeros and poles introduced by phase shift and transmission line, respectively; Step 5.4, Normalizing the frequency representation as: ​ for removal of a frequency corresponding to the parametric phase offset effect; Step 5.4、 The zero point for the matrix 7. is a diagonal matrix of order with the poles , is a unit vector, is a row vector collecting the residues , denotes the matrix transpose; Step 5.5, extracting the zero-pole pairs away from the origin : ; wherein referred to as the phase factor, the phase of the phase offset introduced by the phase offset and the transmission line; Step 5.5, removing phase offset after Parameters are:

8. is the phase offset removed parameters, are simulated / measured parameters, for ； With respect to ; denotes the phase, note that the calculation of the phase values of between and can lead to discontinuities in certain phase values, requiring a "unwrap" operation to make the phase continuous; Step 5.6, remove phase offset from parameters are saved in an array.

9. The method of claim 1, wherein: The coupling matrix extraction method comprises: Step 6.1, importing the trained deep separable convolutional neural network model; Step 6.2, loading test data and inputting the S parameters after removing the phase shift into the deep separable convolutional neural network model; Step 6.3, preprocessing the data, converting the array data to a tensor, expanding the dimensions, adding a channel dimension, using 32 as the floating-point data type, and transferring the data processing to the GPU for accelerated calculation; Step 6.4, predicting the input S parameters using the deep separable convolutional neural network to obtain the corresponding coupling matrix; Step 6.5, saving the model prediction vector as an array, rearranging the array elements according to the index, and constructing a symmetric matrix; Step 6.6, comparing the coupling matrix obtained by the model prediction with the theoretical value, calculating the difference of each element in the matrix, and adjusting the geometric parameters according to the correspondence between the coupling matrix elements and the filter geometric parameters; Step 6.7, calculate the predicted coupling matrix corresponding parameters, and plot the parameter curve against the HFSS extracted parameter raw data curve without de-aliasing phase shift, and observe whether the predicted coupling matrix calculated parameter curve can fit the simulation parameter curve.

Images