A network parameter representation method based on a multi-path convolutional autoencoder
By using the multipath convolutional autoencoder (MCAE) method, the limitations of existing network parameter representation methods and the low training efficiency are solved. This method achieves high-precision network parameter representation and rapid construction of ANN surrogate models, which are applicable to the design of various microwave devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-02-10
- Publication Date
- 2026-05-29
AI Technical Summary
Existing network parameter characterization methods suffer from limited application scope and low training efficiency in microwave device design. In particular, the coupling matrix method has limited application scope, and the vector fitting method suffers from residue-pole mismatch, which leads to low training efficiency or failure to convergence of ANN surrogate models.
The multipath convolutional autoencoder (MCAE) method is adopted to divide the network parameters into real and imaginary parts, process them through multiple paths, and use discrete Fourier transform to determine the convolution kernel length and step size to construct an MCAE model, thereby achieving efficient representation of network parameters. Microwave devices can be quickly designed through an ANN surrogate model.
It achieves higher accuracy in network parameter characterization, avoids residue-pole mismatch, improves the construction efficiency and accuracy of ANN surrogate models, and is suitable for the design of various microwave devices.
Smart Images
Figure CN122114038A_ABST
Abstract
Description
Technical Field
[0002] This invention belongs to the field of communication technology, specifically relating to a network parameter characterization method based on a multi-path convolutional autoencoder, which can be used to establish a neural network surrogate model and accelerate the design of microwave devices. Background Technology
[0004] Traditional microwave device design methods rely heavily on iterative electromagnetic simulations and optimizations, resulting in excessively long design times. The development of artificial intelligence is highly likely to be integrated into microwave device design, potentially accelerating the design process and shortening design time.
[0005] By constructing neural network surrogate models for microwave devices (hereinafter referred to as ANN surrogate models), it is possible to quickly reuse microwave devices without the need for electromagnetic simulation iterations, thereby enabling rapid design of microwave devices. In the design of ANN surrogate models, the representation of network parameters is a crucial step, directly affecting the neural network size and final prediction accuracy of the ANN surrogate model. Summary of the Invention
[0007] Existing methods for representing network parameters mainly include the coupling matrix method and the vector fitting method. The coupling matrix method can only be used for specific filter topologies, limiting its application scope. The vector fitting method suffers from residue-pole mismatch, leading to low training efficiency of subsequent ANN surrogate models, and even failure to converge.
[0008] This invention provides a network parameter representation method based on a multi-path convolutional autoencoder (MCAE), hereinafter referred to as the MCAE method. This MCAE method avoids the residue-pole mismatch problem of existing vector fitting methods, achieving a higher accuracy network parameter representation method, and thus enabling efficient construction of ANN surrogate models.
[0009] The architecture of the MCAE described in this invention is as follows: Figure 1 As shown, the feature is that the network parameters are divided into two parts: the real part and the imaginary part. Taking the real part of the network parameters as an example, they are passed through N+1 paths, the first of which is a low-frequency path, and the remaining N are high-frequency compensation paths. Each path consists of an encoder and a decoder. The low-frequency path includes a low-frequency encoder and a low-frequency decoder, and the high-frequency compensation path includes a high-frequency compensation encoder k and a high-frequency compensation decoder k, where k = 1, 2, ..., N. After passing through N+1 encoders, the real part of the network parameters generates the corresponding compressed network parameter vector, i.e., the real part S. code (Low frequency) and real part S code(High frequency k). Next, it is restored to a quantity with the same length as the network parameters through the corresponding N+1 decoders. Then, the real part S is... code (Low frequency) and real part S code (High-frequency k) is combined through linear layers to restore the real part of the network parameters. The processing of the imaginary part of the network parameters is the same as that of the real part. The real and imaginary parts of the network parameters are then recombine to obtain the original network parameters.
[0010] The implementation of the MCAE method described in this invention is as follows: Figure 2 As shown, it includes the following steps:
[0011] Step 1: Based on the microwave device to be designed, perform parametric modeling of key geometric parameters to generate an adjustable microwave device model. Determine the scanning range of key geometric parameters and generate key geometric parameter data. Use electromagnetic simulation software to generate multiple sets of network parameter data, and combine them with the key geometric parameter data to form a key geometric parameter-network parameter dataset.
[0012] Step 2: Perform a Discrete Fourier Transform on the network parameter data of the geometric parameter-network parameter dataset. The low-frequency path frequencies are calculated using energy weighting, i.e. , where f i To obtain all frequencies using the Discrete Fourier Transform, where K is the number of frequencies, E i This corresponds to its energy. The real and imaginary parts of the network parameters are not centered and may contain extremely low frequency components, so calculations need to start from the *a*-th frequency. Similarly, the formula for calculating the frequency of the *k*-th high-frequency compensation path is: , where m k The frequency starting index of the k-th high-frequency compensation path obtained by the discrete Fourier transform is denoted as .
[0013] Step 3, obtain the low-frequency path frequency. and high-frequency compensation path frequency Afterwards, through Determine the kernel length of the low-frequency path convolution. ,pass Determine the kernel length of the high-frequency compensation path convolution. Among them, f s β is the sampling frequency of the Discrete Fourier Transform, and β is the shape coefficient. The kernel step size S... s The calculation formula is Where η is the overlap factor, and the kernel length is M. or .
[0014] Step 4: Based on the above convolutional kernel length and step size, construct the MCAE. Determine the number of neurons in the linear layer according to the number of frequency points in the network parameters. Train the MCAE to obtain an MCAE model that matches the network parameters of the microwave device.
[0015] The steps to build an ANN proxy model based on MCAE are as follows:
[0016] The network parameters are input into the encoder of MCAE to obtain a compressed network parameter vector, which, together with the key geometric parameters, forms the key geometric parameter-compressed network parameter vector dataset. The MCAE encoder can be invoked independently by modifying the forward function, changing the original data flow through the encoder, decoder, and linear layers to output solely through the encoder, thus achieving the effect of calling the encoder independently.
[0017] An ANN model was built and trained based on this dataset. This ANN model represents the mapping relationship between key geometric parameters and compressed network parameter vectors.
[0018] The decoder part of MCAE is cascaded after the trained ANN model, and the output is the network parameters. The decoder of MCAE is called separately by modifying the forward function, changing the original data flow through encoder, decoder and linear layers to output only through decoder, thus achieving the effect of calling decoder separately.
[0019] ANN proxy model, such as Figure 3 As shown, it includes the ANN model input layer, ANN model hidden layer, ANN model output layer, and decoder from MCAE.
[0020] When using the ANN proxy model, the input is the geometric parameters of the microwave device, and the output is the network parameters of the microwave device.
[0021] Compared with the prior art, the MCAE method described in this invention has the following advantages:
[0022] (1) Compared with the vector fitting method, the output vector order of the MCAE method is completely fixed and given by the structure of the autoencoder, which can completely avoid the data chaos caused by residue-pole mismatch and the problem of training failure.
[0023] (2) Compared with the fully connected layer autoencoder method, since the initial parameters of the MCAE method are determined by the physical meaning in the discrete Fourier transform, it can quickly determine the parameters of its own neural network and provide higher compression efficiency and higher accuracy, which greatly improves the construction efficiency of ANN proxy model.
[0024] (3) Compared with the coupling matrix method, the MCAE method can quickly and accurately characterize the network parameters of filter topologies that are difficult to characterize by the coupling matrix method without going through the coupling matrix extraction step. Attached Figure Description
[0026] Figure 1 This is a schematic diagram of the MCAE architecture;
[0027] Figure 2 This is the MCAE method flowchart;
[0028] Figure 3 This is a schematic diagram of an ANN proxy model based on MCAE;
[0029] Figure 4 This is a schematic diagram of the model in Example 1;
[0030] Figure 5 S is Example 1 11 Discrete Fourier transform energy spectrum of the parameters;
[0031] Figure 6 S is Example 1 21 Discrete Fourier transform energy spectrum of the parameters;
[0032] Figure 7 This is a schematic diagram of the MCAE architecture in Example 1;
[0033] Figure 8 S is Example 1 11 A diagram illustrating the comparison of parameter characterization effects;
[0034] Figure 9 S is Example 1 21 A diagram illustrating the comparison of parameter characterization effects;
[0035] Figure 10 This is a schematic diagram of the physical structure of Example 2;
[0036] Figure 11 This is a schematic diagram of the circuit topology of Example 2;
[0037] Figure 12 S is Example 2 21 A diagram illustrating the comparison of parameter characterization effects;
[0038] Figure 13 S is Example 2 31 A diagram showing the comparison of parameter characterization effects. Detailed Implementation
[0040] To more clearly illustrate the purpose, technical solution, and advantages of this invention, the technical solution of this invention is described clearly and completely below with reference to embodiments. It should be understood that the specific embodiments described herein are only for explaining this invention and are not intended to limit this invention. Obviously, the embodiments used herein are some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0041] Example 1 is a microstrip line filter, such as Figure 3 As shown. According to the implementation steps of the MCAE method, step 1 is to perform parametric modeling of key geometric parameters to generate a highly tunable filter model. It is understood that, for simplicity, Example 1 selects resonator lengths X1, X2, and X3 as key geometric parameters. Other geometric parameters of Example 1, such as spacing, feed position, etc., can also be used in parametric modeling. After completing the parametric modeling, changing these key geometric parameters can quickly adjust the model of Example 1, thus Example 1 can be used to achieve different frequency responses.
[0042] Step 2: Set the variation range for each key geometric parameter and scan to obtain multiple sets of key geometric parameter combinations. Then, for each key geometric parameter combination, perform electromagnetic simulation using parametric electromagnetic simulation software to generate network parameter data, forming a key geometric parameter-network parameter dataset. This dataset is then exported in batches as S2P files using a Python script. A Python script for batch processing S2P files is constructed, using the scikit-rf library to read S2P files in batches and convert and export data samples. Without loss of generality, Example 1 selects the scattering parameter, i.e., the S-parameter, as the network parameter. The S-parameter is a complex number, including both real and imaginary parts, across all sampling points within the target frequency range (800 points are selected in Example 1).
[0043] Step 3: Perform a Discrete Fourier Transform on the network parameter data based on the obtained key geometric parameters-network parameter dataset to obtain the frequencies of the low-frequency path and the high-frequency compensation path, thereby determining the initial lengths of the low-frequency and high-frequency convolution kernels. In Example 1, the network parameter is used as the scattering parameter S. 11 and S 21 S 11 and S 21 Perform Discrete Fourier Transform (DFT) separately, and set the sampling frequency f of the DFT. s =1000Hz and shape coefficient β=4, the frequency characteristic distribution after discrete Fourier transform is obtained, such as Figure 5 and Figure 6As shown. Based on the characteristics of this frequency distribution, Example 1 decides to use one low-frequency path and one high-frequency compensation path, therefore N=1.
[0044] Calculating low-frequency path frequencies At that time, according to Figure 5 and Figure 6 Let a = 4. Also, K = 20. Regarding S... 11 Calculated =9.80 Hz, for S 21 Calculated =13.43 Hz.
[0045] Calculating the high-frequency compensation frequency At that time, according to Figure 5 and Figure 6 Arranged according to the energy magnitude of frequency, and determined Calculations yielded , .
[0046] Due to S 11 and S 21 The calculated low-frequency path frequency and high-frequency compensation frequency are quite close, therefore S 11 and S 21 Using the same kernel length, therefore, , The convolution kernel step size is calculated based on the overlap factor; the step size for the low-frequency path convolution kernel is... The high-frequency compensation path convolution kernel step is .
[0047] Step 4: Based on the initial length and stride of each convolutional kernel, construct the MCAE, as follows: Figure 7 As shown, it consists of an encoder, a decoder, and linear layers. The encoder is composed of convolutional layers constructed from the low-frequency path convolutional kernels and convolutional layers constructed from the high-frequency compensation path convolutional kernels. The decoder is composed of convolutional layers constructed from deconvolutional kernels symmetrical to the low-frequency path convolutional kernels and convolutional layers constructed from deconvolutional kernels symmetrical to the high-frequency compensation path convolutional kernels.
[0048] The convolutional layers described above all use the sigmoid activation function, but other activation functions can be used as needed. The linear layers serve only as learnable weight assignments, so they do not use an activation function and output directly. It is understood that since the output of the linear layer in Example 1 is an S-parameter, the length of the linear layer's output vector is directly determined by the number of discrete frequency points of the S-parameter.
[0049] The MCAE method described in this invention achieves the following restoration effect on Example 1: Figure 8 and Figure 9As shown in the figure. Here, the VF curve refers to the S-parameters restored by the vector fitting method, the MCAE curve refers to the S-parameters restored by the MCAE method described in this invention, and the EM-DATA curve refers to the original S-parameter data obtained from electromagnetic simulation.
[0050] When the vector fitting method uses a sufficient number of residue pole sets, the characterization error can be very small. However, because a uniform compressed S-parameter vector length is required, the vector fitting method must enforce a consistent number of residue pole sets. In Example 1, a number of residue pole sets matching the physical structure is required; therefore, the vector fitting method uses 5 residue pole sets for characterization. The MCAE method, on the other hand, uses a compressed S-parameter vector of length 15 for characterization.
[0051] like Figure 8 and Figure 9 As shown, the root mean square error (RMSE) of the vector fitting method is 1.2%, while the MCAE method has an RMSE of 0.62%, exhibiting higher accuracy. Example 1 strongly demonstrates that the MCAE method described in this invention is applicable to two-port transmission line filters.
[0052] Example 2 Figure 10 The diagram shown is of a microwave device consisting of a three-port LC duplexer, and its topology is as follows. Figure 11 As shown. The effect of restoring S-parameters using vector fitting and the MCAE method described in this invention is as follows. Figure 12 and 13 As shown, the root mean square error (RMSE) of the vector fitting method is 3.0%, while the MCAE method has an RMSE of 0.55%, exhibiting higher accuracy. Example 2 strongly demonstrates that the MCAE method described in this invention is applicable to three-port lumped parameter duplexers.
[0053] As can be seen from the two different types of embodiments above, the MCAE method of the present invention is applicable to two-port filters and three-port duplexers. Furthermore, the MCAE method of the present invention can also be applied to other implementations and microwave devices with different functions.
Claims
1. A method for characterizing network parameters, characterized in that: The network parameters are divided into two parts: the real part and the imaginary part. Taking the real part as an example, the network parameters are processed through N+1 paths. The first path is a low-frequency path, and the remaining N paths are high-frequency compensation paths. Each path consists of an encoder and a decoder. The low-frequency path includes a low-frequency encoder and a low-frequency decoder, and the high-frequency compensation path includes a high-frequency compensation encoder k and a high-frequency compensation decoder k, where k = 1, 2, ..., N. After passing through N+1 encoders, the real part of the network parameters generates the corresponding compressed network parameter vector, i.e., the real part S. code (Low frequency) and real part S code (High frequency k); then, through the corresponding N+1 decoders, it is restored to a quantity with the same length as the network parameters. Then, the real part S... code (Low frequency) and real part S code (High-frequency k) is combined through a linear layer to restore the real part of the network parameters; the processing of the imaginary part of the network parameters is the same as that of the real part of the network parameters. The real and imaginary parts of the network parameters are then recombined to obtain the original network parameters.
2. The network parameter characterization method according to claim 1, comprising the following steps: Step 1: Generate a dataset of key geometric parameters and network parameters for microwave devices; Step 2: Perform Discrete Fourier Transform on the dataset to calculate the frequencies of low-frequency paths and high-frequency compensation paths; Step 3: Calculate the kernel length and stride of low-frequency paths and high-frequency compensation paths; Step 4: Construct and train a multi-path convolutional autoencoder.
3. The network parameter characterization method according to claim 1, low-frequency path frequency Calculated by energy weighting, i.e. ,in, f i To obtain all frequencies using the Discrete Fourier Transform, where K is the number of frequencies, E i This corresponds to its energy. The real and imaginary parts of the network parameters are not centered and may contain ultra-low frequency components, so calculations need to start from the a-th frequency; the frequency of the k-th high-frequency compensation path... The calculation formula is: , where m k The frequency starting index of the k-th high-frequency compensation path obtained by the discrete Fourier transform is denoted as .
4. The network parameter characterization method according to claim 1, yielding the low-frequency path frequency. and high-frequency compensation path frequency Afterwards, through Determine the kernel length of the low-frequency path convolution. ,pass Determine the kernel length of the high-frequency compensation path convolution. ;in, f s β is the sampling frequency of the Discrete Fourier Transform, and β is the shape coefficient. The kernel step size S... s The calculation formula is Where η is the overlap factor, and the kernel length M is... or .
5. The network parameter characterization method according to claim 1 is applicable not only to two-port and three-port microwave passive devices, but also to other types and functions of microwave devices.
6. According to the network parameter representation method of claim 1, the specific implementation of constructing the ANN proxy model is to cascade the decoder part of the multi-path convolutional autoencoder after the trained ANN model.
7. The ANN surrogate model according to claim 6, comprising an ANN model input layer, an ANN model hidden layer, an ANN model output layer, and a decoder of a multipath convolutional autoencoder (MCAE).