A neural network-based three-pin waveguide impedance matching method

By using a three-pin waveguide impedance matching method based on neural networks, a nonlinear mapping relationship between pin positions and scattering parameters is established, which solves the problems of low computational efficiency and insufficient dynamic tracking capability in traditional methods, and achieves efficient impedance matching and energy utilization.

CN123116835APending Publication Date: 2026-11-17CHENGDU FENYU ELECTRONIC TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611358687.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2025-12-25
Filing Date
2026-09-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional three-pin impedance matching methods are computationally inefficient, highly dependent on models, and have limited real-time tracking capabilities for dynamic loads, making it difficult to achieve high-precision matching.

Method used

A neural network-based three-pin waveguide impedance matching method is adopted. The nonlinear mapping relationship between pin position and network scattering parameters is established through data-driven approach, and an impedance matching knowledge base covering the entire parameter space is constructed. The neural network is then used to perform fast and globally optimal impedance matching.

Benefits of technology

It achieves fast and globally optimal impedance matching, overcomes the matching blind zone and response delay problems of traditional methods, improves energy transmission efficiency and system adaptability, and keeps the load's microwave energy utilization rate above 99%.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116835A_ABST
    Figure CN123116835A_ABST
Patent Text Reader

Abstract

This application discloses a neural network-based method for impedance matching in a three-pin waveguide, comprising: inserting three pins into the waveguide at depths of a first dimension; measuring the two-port scattering parameters of the waveguide within a preset depth range; and packaging the pin insertion depths and corresponding two-port scattering parameters into a training dataset; inputting the training dataset into a neural network for training to obtain a neural network model; inputting the pin insertion depth data into the neural network model to generate a scattering parameter dataset with the pin insertion depths of the waveguide at second dimension intervals; obtaining the current load impedance under a preset reflection coefficient; and querying the pin insertion depth of the waveguide with the closest matching impedance to the current load impedance from the scattering parameter dataset. This application significantly improves the utilization efficiency of microwave energy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of microwave technology, and in particular to a three-pin waveguide impedance matching method based on neural networks. Background Technology

[0002] Microwave heating technology, with its advantages of cleanliness, high efficiency, and strong controllability, has been widely used in industrial and scientific research fields such as materials synthesis, food processing, and medical sterilization. This technology utilizes microwave electromagnetic energy to directly act on dielectric materials, achieving bulk heating through molecular polarization and relaxation effects. However, in practical applications, the impedance mismatch between the microwave transmission path and the load to be heated severely restricts the improvement of system energy efficiency and performance. Therefore, three-pin impedance matching devices have been widely introduced to solve this problem.

[0003] Traditional three-pin impedance matching schemes employ matching algorithms based on equivalent circuit models. This involves establishing a mapping relationship between pin insertion depth and equivalent circuit parameters, combined with an enumeration search strategy to achieve impedance matching. The scheme first establishes an equivalent circuit model of the three-pin waveguide, derives port scattering parameters and reverse equivalent impedance, constructs a pin configuration database, and then determines the optimal pin position through a lookup table. However, because this scheme relies on the accuracy of the equivalent circuit model, and the computational cost of the enumeration search increases exponentially with search accuracy, it suffers from low computational efficiency, sensitivity to model errors, and limited real-time tracking capability for dynamic loads when handling high-precision matching requirements. Summary of the Invention

[0004] In view of this, this application provides a three-pin waveguide impedance matching method based on neural networks, which breaks through the dependence of traditional methods on the analytical model of equivalent circuits, significantly improves matching efficiency through data-driven approach, and effectively solves the shortcomings of traditional schemes in terms of computational complexity, model dependence and dynamic tracking performance.

[0005] This application discloses a three-pin waveguide impedance matching method based on neural networks, which includes: Step 1: Insert the three pins into the waveguide at intervals of the first dimension, measure the two-port scattering parameters of the waveguide within the preset depth range, and package the pin insertion depths and corresponding two-port scattering parameters into a training dataset. Step 2: Input the training dataset into the neural network for training to obtain the neural network model; Step 3: Input the depth data of the three pins inserted into the waveguide into the neural network model to generate a scattering parameter dataset with the depth of the three pins inserted into the waveguide at intervals of the second dimension; the second dimension is smaller than the first dimension; Step 4: Based on the preset reflection coefficient, obtain the current load impedance under the preset reflection coefficient, and based on the current load impedance, query the pin insertion depth of the waveguide under the matching impedance that is closest to the current load impedance from the scattering parameter dataset.

[0006] Furthermore, the neural network includes an input layer, a hidden layer, and an output layer connected in sequence; The input data of the input layer has three feature values, namely the insertion depths H1, H2, and H3 of the three pins into the waveguide, which directly determine that the input layer has three neurons. The hidden layer consists of multiple fully connected layers. Each fully connected layer performs a linear transformation on the output of the previous fully connected layer, and then passes it through a non-linear ReLU activation function so that the neural network can learn the complex non-linear relationship between the input and the output. The output layer outputs four predicted scattering parameters: S11, S12, S21, and S22. A linear activation function is used to constrain the range of values ​​for the scattering parameters. S11_Real and S11_Imag are the real and imaginary parts of S11, respectively; S12_Real and S12_Imag are the real and imaginary parts of S12, respectively; S21_Real and S21_Imag are the real and imaginary parts of S21, respectively; and S22_Real and S22_Imag are the real and imaginary parts of S22, respectively.

[0007] Furthermore, in the multiple fully connected layers, the number of neurons in each fully connected layer is less than the number of neurons in the previous fully connected layer, so as to force the neural network to extract and compress information step by step in each fully connected layer and abstract higher-level features from the original input.

[0008] Furthermore, in step 2, the parameters in the neural network are adjusted by the Adam optimizer to minimize the loss function; the loss function is a function that controls the step size in each training round. During neural network training, the loss function is calculated using the training dataset: a fixed number of samples are randomly selected from the training dataset for each batch, and the predicted values ​​are obtained through forward propagation of the neural network. The error between the predicted value and the true label is calculated for each sample using the mean squared error formula. The average error of all samples in the batch is then taken to obtain the loss function for that batch. This loss function is directly used for backpropagation gradient calculation and weight update, and its trend reflects the degree of fit of the neural network model to the training data.

[0009] Furthermore, the loss function uses the mean squared error (MSE): (1) in, The total number of samples in the training dataset, To use real data in the training dataset, It is the predicted data output by the neural network model; For a single sample, the neural network outputs the following value: (2) in, The value output by the neural network has a range of values. , , , , , , , ; It is the real part of S11 predicted by the neural network model. It is the imaginary part of S11 predicted by the neural network model; It is the real part of S12 predicted by the neural network model. It is the imaginary part of S12 predicted by the neural network model; It is the real part of S21 predicted by the neural network model. It is the imaginary part of S21 predicted by the neural network model; It is the real part of S22 predicted by the neural network model. S22 is the imaginary part of the neural network model prediction; S11, S12, S21 and S22 are all scattering parameters. The expression for the real label is: (3) Where y is the actual label, and its value is... , , , , , , , The true labels are the samples in the training sample set; It is the real part of S11 in the training dataset. It is the imaginary part of S11 in the training dataset; It is the real part of S12 in the training dataset. It is the imaginary part of S12 in the training dataset; It is the real part of S21 in the training dataset. It is the imaginary part of S21 in the training dataset; It is the real part of S22 in the training dataset. It is the imaginary part of S22 in the training dataset; The mean squared error (MSE) of a single sample is calculated using the following formula: (4) in, The mean squared error (MSE) for a single sample; In each training round, the training dataset is divided into B batches, and the loss function for each of the B batches is calculated using the following formula. : (5) in, It refers to the j-th scattering parameter in the i-th batch of training data; It refers to the j-th scattering parameter in the i-th batch of training data predicted by the neural network model; Equation (5) can be simplified to: (6) The gradient of the loss function in formula (6) is calculated using the following formula. : (7) in, These are the eight scattering parameters in the training dataset; These are the eight scattering parameters predicted by the neural network model; It is a time step The weights at time; w is the weight matrix connecting the two neurons in each fully connected layer, which is automatically learned from the training data through the backpropagation algorithm. Its purpose is to minimize the training loss between the scattering parameters predicted by the neural network model and the scattering parameters in the training dataset. The time step is obtained using the following formula. Time weight : (8) in, It is the learning rate.

[0010] Furthermore, during the neural network training process, after each round of training, the final training loss between the data predicted by the neural network model and the original pre-training data is calculated, and the final training loss is used to evaluate whether the neural network model obtained in each round of training meets the requirements.

[0011] Further, step 2 includes: After each round of training, the training loss for a single sample is: (9) in, The training loss for a single sample, It is the real part of S11 in the training dataset. It is the imaginary part of S11 in the training dataset; It is the real part of S12 in the training dataset. It is the imaginary part of S12 in the training dataset; It is the real part of S21 in the training dataset. It is the imaginary part of S21 in the training dataset; It is the real part of S22 in the training dataset. It is the imaginary part of S22 in the training dataset; It is the real part of S11 predicted by the neural network model. It is the imaginary part of S11 predicted by the neural network model; It is the real part of S12 predicted by the neural network model. It is the imaginary part of S12 predicted by the neural network model; It is the real part of S21 predicted by the neural network model. It is the imaginary part of S21 predicted by the neural network model; It is the real part of S22 predicted by the neural network model. S22 is the imaginary part of the neural network model prediction; S11, S12, S21 and S22 are all scattering parameters. After each training round, the training loss for batch B is: (10) in, The training loss of batch B after each round of training. This represents the training loss for a single sample in the i-th batch after each round of training. After each round of training, the final training loss is: (11) in, This represents the final training loss at the end of each training round. This represents the total number of training data in batch B during one round of training.

[0012] Further, step 2 includes: In each training round, a predetermined proportion of data is extracted from the original measured data as a validation dataset to validate the training loss, and the validation loss formula is defined as follows: (12) in, To verify the loss, To verify the total number of samples in the dataset, To validate the dataset, To verify the first in the dataset One sample, For the first The loss value for each sample.

[0013] Further, obtaining the current load impedance under the preset reflection coefficient includes: The load impedance is obtained using the following formula: (13) in, Given the current load impedance, , , , These are all two-port scattering parameters of the waveguide. For a fixed reflection coefficient, This is the characteristic impedance.

[0014] Furthermore, characteristic impedance The expression is: (14) in, Let b be the length of the waveguide, c be the height of the waveguide, c be the speed of light in a vacuum, and f be the frequency of the electromagnetic wave. and These are the permeability constant and dielectric constant of the waveguide filler, respectively.

[0015] Furthermore, during the acquisition of the training dataset, a three-pin connector was used as an impedance matching device, and the varying load was obtained by adjusting the depth of the three-pin connector inserted into the waveguide.

[0016] Due to the adoption of the above technical solution, this application has the following advantages: 1. This application establishes a nonlinear mapping relationship between pin position and network scattering parameters through a deep learning model. Using this model, an impedance matching knowledge base (scattering parameter dataset) covering the entire parameter space is constructed. In actual matching, the optimal pin configuration is directly queried based on the real-time impedance, realizing fast and globally optimal impedance matching. This effectively overcomes the matching blind zone and response delay problems of traditional methods, and improves energy transmission efficiency and system adaptability.

[0017] 2. This application employs a distance-based fast matching query mechanism to achieve real-time and accurate matching. This application overcomes the dependence of traditional methods on the analytical model of equivalent circuits, significantly improving matching efficiency through a data-driven approach, and effectively addressing the shortcomings of traditional schemes in terms of computational complexity, model dependence, and dynamic tracking performance.

[0018] 3. This application can ensure that microwave heating is within the effective matching range, and the load can still maintain a microwave energy utilization rate of over 99%. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0020] Figure 1 This is a schematic diagram of a three-pin waveguide according to an embodiment of this application; Figure 2 This is a schematic diagram of the neural network architecture of an embodiment of this application; Figure 3 This is a schematic diagram of the impedance matching process for a three-pin waveguide according to an embodiment of this application; Figure 4 This is a schematic diagram of the optimized matching result of an embodiment of this application. Detailed Implementation

[0021] The present application will be further described in conjunction with the accompanying drawings and embodiments. The described embodiments are only some, not all, of the embodiments of the present application. All other embodiments obtained by those skilled in the art should fall within the protection scope of the embodiments of the present application.

[0022] See Figure 3 This application provides an embodiment of a three-pin waveguide impedance matching method based on neural networks, which includes: Step 1: Insert the three pins into the waveguide at intervals of the first dimension, measure the two-port scattering parameters of the waveguide within the preset depth range, and package the pin insertion depths and corresponding two-port scattering parameters into a training dataset. Step 2: Input the training dataset into the neural network for training to obtain the neural network model; Step 3: Input the depth data of the three pins inserted into the waveguide into the neural network model to generate a scattering parameter dataset with the depth of the three pins inserted into the waveguide at intervals of the second dimension; the second dimension is smaller than the first dimension; Step 4: Based on the preset reflection coefficient, obtain the current load impedance under the preset reflection coefficient, and based on the current load impedance, query the pin insertion depth of the waveguide under the matching impedance that is closest to the current load impedance from the scattering parameter dataset.

[0023] like Figure 1 As shown, the three-pin waveguide consists of three pins, a waveguide, and a lossy load.

[0024] Optionally, see Figure 2 The neural network comprises an input layer, a hidden layer, and an output layer connected in sequence. The input data of the input layer has three feature values, namely the insertion depths H1, H2, and H3 of the three pins into the waveguide, which directly determine that the input layer has three neurons. The hidden layer consists of multiple fully connected layers. Each fully connected layer performs a linear transformation on the output of the previous fully connected layer, and then passes it through a non-linear ReLU activation function so that the neural network can learn the complex non-linear relationship between the input and the output. The output layer outputs four predicted scattering parameters: S11, S12, S21, and S22. A linear activation function is used to constrain the range of values ​​for the scattering parameters. S11_Real and S11_Imag are the real and imaginary parts of S11, respectively; S12_Real and S12_Imag are the real and imaginary parts of S12, respectively; S21_Real and S21_Imag are the real and imaginary parts of S21, respectively; and S22_Real and S22_Imag are the real and imaginary parts of S22, respectively.

[0025] Optionally, in the plurality of fully connected layers, the number of neurons in each fully connected layer is less than the number of neurons in the previous fully connected layer, so as to force the neural network to extract and compress information step by step in each fully connected layer and abstract higher-level features from the original input.

[0026] Optionally, in step 2, the parameters in the neural network are adjusted by the Adam optimizer to minimize the loss function; the loss function is a function that controls the step size in each training round. During neural network training, the loss function is calculated using the training dataset: a fixed number of samples are randomly selected from the training dataset for each batch, and the predicted values ​​are obtained through forward propagation of the neural network. The error between the predicted value and the true label is calculated for each sample using the mean squared error formula. The average error of all samples in the batch is then taken to obtain the loss function for that batch. This loss function is directly used for backpropagation gradient calculation and weight update, and its trend reflects the degree of fit of the neural network model to the training data.

[0027] Optionally, the loss function uses the mean squared error (MSE): (1) in, The total number of samples in the training dataset, To use real data in the training dataset, It is the predicted data output by the neural network model; For a single sample, the neural network outputs the following value: (2) in, The value output by the neural network has a range of values. , , , , , , , ; It is the real part of S11 predicted by the neural network model. It is the imaginary part of S11 predicted by the neural network model; It is the real part of S12 predicted by the neural network model. It is the imaginary part of S12 predicted by the neural network model; It is the real part of S21 predicted by the neural network model. It is the imaginary part of S21 predicted by the neural network model; It is the real part of S22 predicted by the neural network model. S22 is the imaginary part of the neural network model prediction; S11, S12, S21 and S22 are all scattering parameters. The expression for the real label is: (3) Where y is the actual label, and its value is... , , , , , , , The true labels are the samples in the training sample set; It is the real part of S11 in the training dataset. It is the imaginary part of S11 in the training dataset; It is the real part of S12 in the training dataset. It is the imaginary part of S12 in the training dataset; It is the real part of S21 in the training dataset. It is the imaginary part of S21 in the training dataset; It is the real part of S22 in the training dataset. It is the imaginary part of S22 in the training dataset; The mean squared error (MSE) of a single sample is calculated using the following formula: (4) in, The mean squared error (MSE) for a single sample; In each training round, the training dataset is divided into B batches, and the loss function for each of the B batches is calculated using the following formula. : (5) in, It refers to the j-th scattering parameter in the i-th batch of training data; It refers to the j-th scattering parameter in the i-th batch of training data predicted by the neural network model; Equation (5) can be simplified to: (6) The gradient of the loss function in formula (6) is calculated using the following formula. : (7) in, These are the eight scattering parameters in the training dataset; These are the eight scattering parameters predicted by the neural network model; It is a time step The weights at time; w is the weight matrix connecting the two neurons in each fully connected layer, which is automatically learned from the training data through the backpropagation algorithm. Its purpose is to minimize the training loss between the scattering parameters predicted by the neural network model and the scattering parameters in the training dataset. The time step is obtained using the following formula. Time weight : (8) in, It is the learning rate.

[0028] The partial derivatives (i.e., gradients) of the loss function with respect to the weights in the embodiments of this application are as follows. J(w_t) has core physical significance and application value. By calculating the partial derivatives of the loss function (such as mean square error MSE) with respect to each weight parameter in the neural network, the "contribution" and "direction of influence" of each weight on the error of the predicted scattering parameters can be accurately quantified. Gradient information not only reveals the local sensitivity of the complex nonlinear mapping relationship between the pin insertion depth (H1, H2, H3) and the waveguide scattering parameters (S11, S12, S21, S22), but also indicates the direction of weight adjustment required to minimize the prediction error. Based on the gradient information, the weight update formula of formula (8) is used to iteratively optimize the network parameters, where the learning rate η controls the update step size. The weight update process continues until the network can quickly and accurately output the corresponding scattering parameters according to any combination of input pin depths.

[0029] Optionally, during the neural network training process, after each round of training, the final training loss between the data predicted by the neural network model and the original pre-training data is calculated, and the neural network model obtained in each round of training is evaluated to determine whether it meets the requirements.

[0030] Optionally, step 2 includes: After each round of training, the training loss for a single sample is: (9) in, The training loss for a single sample, It is the real part of S11 in the training dataset. It is the imaginary part of S11 in the training dataset; It is the real part of S12 in the training dataset. It is the imaginary part of S12 in the training dataset; It is the real part of S21 in the training dataset. It is the imaginary part of S21 in the training dataset; It is the real part of S22 in the training dataset. It is the imaginary part of S22 in the training dataset; It is the real part of S11 predicted by the neural network model. It is the imaginary part of S11 predicted by the neural network model; It is the real part of S12 predicted by the neural network model. It is the imaginary part of S12 predicted by the neural network model; It is the real part of S21 predicted by the neural network model. It is the imaginary part of S21 predicted by the neural network model; It is the real part of S22 predicted by the neural network model. S22 is the imaginary part of the neural network model prediction; S11, S12, S21 and S22 are all scattering parameters. After each training round, the training loss for batch B is: (10) in, The training loss of batch B after each round of training. This represents the training loss for a single sample in the i-th batch after each round of training. After each round of training, the final training loss is: (11) in, This represents the final training loss at the end of each training round. This represents the total number of training data in batch B during one round of training.

[0031] Optionally, step 2 includes: In each training round, a predetermined proportion of data is extracted from the original measured data as a validation dataset to validate the training loss, and the validation loss formula is defined as follows: (12) in, To verify the loss, To verify the total number of samples in the dataset, To validate the dataset, To verify the first in the dataset One sample, For the first The loss value for each sample.

[0032] Optionally, obtaining the current load impedance under a preset reflection coefficient includes: The load impedance is obtained using the following formula: (13) in, Given the current load impedance, , , , These are all two-port scattering parameters of the waveguide. For a fixed reflection coefficient, This is the characteristic impedance.

[0033] Optionally, characteristic impedance The expression is: (14) in, Let b be the length of the waveguide, c be the height of the waveguide, c be the speed of light in a vacuum, and f be the frequency of the electromagnetic wave. and These are the permeability constant and dielectric constant of the waveguide filler, respectively.

[0034] Optionally, during the acquisition of the training dataset, a three-pin pin is used as an impedance matching device, and the varying load is obtained by adjusting the depth of the three-pin pin inserted into the waveguide.

[0035] In this embodiment, 60% of the original measured data can be used as the training dataset, 20% as the validation dataset, and 20% as the test dataset. In each training batch, backpropagation and weight updates are directly driven based on the training set, enabling the model to learn the physical mapping between pin depth and scattering parameters. After each training round, the model's performance on data not used in training is monitored based on the validation dataset, guiding hyperparameter adjustment and implementing an early stopping strategy to effectively prevent overfitting. Finally, after complete model training and validation, a completely independent test set is used for final evaluation. This test set simulates new samples in real-world application scenarios, objectively reflecting the model's generalization ability and practical deployment potential. The aforementioned technical solution ensures the effectiveness, stability, and generalization ability of model training, guaranteeing the integrity and reliability from model learning to final application.

[0036] For ease of understanding, this application provides a more specific embodiment: (1) Prepare the dataset: Based on the existing BJ26 three-pin waveguide, the pin insertion depth is at 1mm intervals. The actual waveguide two-port scattering parameters within 0mm to 20mm are measured and packaged into a training dataset.

[0037] (2) Building a neural network: Determine the network hierarchical architecture, the number of neurons in each layer, and the activation function.

[0038] (3) Selecting a training optimizer: Since the three-pin matching is a typical non-linear structure, the Adam algorithm can be selected as the optimization trainer, and a suitable loss function can be built.

[0039] (4) Network training: Set an appropriate training loss function, monitor the optimization effect during training, and select the model with the smallest loss to save.

[0040] (5) Generate matching dataset: Use the model to generate a scattering parameter dataset with a pin spacing of 0.1 mm, and then assume a fixed reflection coefficient. The load impedance under this reflection coefficient is calculated using formula (13).

[0041] (6) Search for matching: Calculate the current load impedance using formula (13), and then search the matching dataset for the pin insertion depth with the closest matching impedance.

[0042] The following are examples of verification of this method: Select BJ26 rectangular waveguide, i.e., waveguide dimensions , Operating frequency .

[0043] To ensure the reliability of the obtained matching network, it is necessary to verify the correctness of the calculation result of formula (13). The formula is verified using simulation. When the pin is zeroed, the relationship between the impedance of the load port and the load impedance calculated by formula (13) is set, as shown in Table 1. As can be seen from Table 1, the calculation formula can effectively calculate the load impedance.

[0044] Table 1. Impedance of the load port set during pin reset and load impedance calculated by formula.

[0045] Zero the pins, set a manually adjusted three-pin simulated load, measure the current emission coefficient, and determine whether the current reflection coefficient amplitude is less than or equal to 0.1. If not, use formula (13) to calculate the current load impedance, find the pin insertion depth with similar impedance in the target matching impedance database (scattering parameter dataset), modify the pin insertion depth in the model, and simulate again to check the current reflection coefficient until the reflection coefficient amplitude is less than 0.1. The pins need to be zeroed each time the initial reflection coefficient is simulated.

[0046] Simulation modeling verifies the matching effect, using the reflection coefficient Γ for evaluation. A smaller Γ indicates better energy absorption. When, it means that the energy utilization efficiency reaches 90% or above, when When the energy utilization efficiency reaches 99% or higher, the simulation verification results are as follows: Figure 4 As shown.

[0047] From Figure 4 The verification results show that the simulation results and theoretical analysis results are in good agreement. The three-pin waveguide impedance matching scheme proposed in this application has a very good application effect. Within the effective impedance matching range, the reflection coefficient Γ is less than -20dB (i.e., the energy absorption efficiency is greater than 99%).

[0048] The three-pin waveguide impedance matching method proposed in this application can perform impedance matching for a wide range of load impedance variations, greatly improving the utilization efficiency of microwave energy.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of this application. Any modifications or equivalent substitutions that do not depart from the spirit and scope of this application should be covered within the protection scope of the claims of this application.

Claims

1. A three-pin waveguide impedance matching method based on neural networks, characterized in that, include: Step 1: Insert the three pins into the waveguide at intervals of the first dimension, measure the two-port scattering parameters of the waveguide within the preset depth range, and package the pin insertion depths and corresponding two-port scattering parameters into a training dataset. Step 2: Input the training dataset into the neural network for training to obtain the neural network model; Step 3: Input the depth data of the three pins inserted into the waveguide into the neural network model to generate a scattering parameter dataset with the depth of the three pins inserted into the waveguide at intervals of the second dimension; The second dimension is smaller than the first dimension; Step 4: Based on the preset reflection coefficient, obtain the current load impedance under the preset reflection coefficient, and based on the current load impedance, query the pin insertion depth of the waveguide under the matching impedance that is closest to the current load impedance from the scattering parameter dataset.

2. The method according to claim 1, characterized in that, The neural network comprises an input layer, a hidden layer, and an output layer connected in sequence. The input data of the input layer has three feature values, namely the insertion depths H1, H2, and H3 of the three pins into the waveguide, which directly determine that the input layer has three neurons. The hidden layer consists of multiple fully connected layers. Each fully connected layer performs a linear transformation on the output of the previous fully connected layer, and then passes it through a non-linear ReLU activation function so that the neural network can learn the complex non-linear relationship between the input and the output. The output layer outputs four predicted scattering parameters: S11, S12, S21, and S22. A linear activation function is used to constrain the range of values ​​for the scattering parameters. S11_Real and S11_Imag are the real and imaginary parts of S11, respectively; S12_Real and S12_Imag are the real and imaginary parts of S12, respectively; S21_Real and S21_Imag are the real and imaginary parts of S21, respectively; and S22_Real and S22_Imag are the real and imaginary parts of S22, respectively.

3. The method according to claim 2, characterized in that, In each of the multiple fully connected layers, the number of neurons in each fully connected layer is less than the number of neurons in the previous fully connected layer, so as to force the neural network to extract and compress information step by step in each fully connected layer and abstract higher-level features from the original input.

4. The method according to claim 1, characterized in that, In step 2, the parameters in the neural network are adjusted by the Adam optimizer to minimize the loss function; the loss function is a function that controls the step size in each training round. During neural network training, the loss function is calculated using the training dataset: a fixed number of samples are randomly selected from the training dataset for each batch, and the predicted values ​​are obtained through forward propagation of the neural network. The error between the predicted value and the true label is calculated for each sample using the mean squared error formula. The average error of all samples in the batch is then taken to obtain the loss function for that batch. This loss function is directly used for backpropagation gradient calculation and weight update, and its trend reflects the degree of fit of the neural network model to the training data.

5. The method according to claim 4, characterized in that, The loss function uses mean squared error (MSE): (1) in, The total number of samples in the training dataset, To use real data in the training dataset, It is the predicted data output by the neural network model; For a single sample, the neural network outputs the following value: (2) in, The value output by the neural network has a range of values. , , , , , , , ; It is the real part of S11 predicted by the neural network model. It is the imaginary part of S11 predicted by the neural network model; It is the real part of S12 predicted by the neural network model. It is the imaginary part of S12 predicted by the neural network model; It is the real part of S21 predicted by the neural network model. It is the imaginary part of S21 predicted by the neural network model; It is the real part of S22 predicted by the neural network model. S22 is the imaginary part of the neural network model prediction; S11, S12, S21 and S22 are all scattering parameters. The expression for the real label is: (3) Where y is the actual label, and its value is... , , , , , , , The true labels are the samples in the training sample set; It is the real part of S11 in the training dataset. It is the imaginary part of S11 in the training dataset; It is the real part of S12 in the training dataset. It is the imaginary part of S12 in the training dataset; It is the real part of S21 in the training dataset. It is the imaginary part of S21 in the training dataset; It is the real part of S22 in the training dataset. It is the imaginary part of S22 in the training dataset; The mean squared error (MSE) of a single sample is calculated using the following formula: (4) in, The mean squared error (MSE) for a single sample; In each training round, the training dataset is divided into B batches, and the loss function for each of the B batches is calculated using the following formula. : (5) in, It refers to the j-th scattering parameter in the i-th batch of training data; It refers to the j-th scattering parameter in the i-th batch of training data predicted by the neural network model; Equation (5) can be simplified to: (6) The gradient of the loss function in formula (6) is calculated using the following formula. : )(7) in, These are the eight scattering parameters in the training dataset; These are the eight scattering parameters predicted by the neural network model; It is a time step The weights at time; w is the weight matrix connecting the two neurons in each fully connected layer, which is automatically learned from the training data through the backpropagation algorithm. Its purpose is to minimize the training loss between the scattering parameters predicted by the neural network model and the scattering parameters in the training dataset. The time step is obtained using the following formula. Time weight : (8) in, It is the learning rate.

6. The method according to claim 4, characterized in that, During neural network training, after each round of training, the final training loss between the data predicted by the neural network model and the original pre-training data is calculated. The final training loss is used to evaluate whether the neural network model obtained in each round of training meets the requirements.

7. The method according to claim 6, characterized in that, Step 2 includes: After each round of training, the training loss for a single sample is: (9) in, The training loss for a single sample, It is the real part of S11 in the training dataset. It is the imaginary part of S11 in the training dataset; It is the real part of S12 in the training dataset. It is the imaginary part of S12 in the training dataset; It is the real part of S21 in the training dataset. It is the imaginary part of S21 in the training dataset; It is the real part of S22 in the training dataset. It is the imaginary part of S22 in the training dataset; It is the real part of S11 predicted by the neural network model. It is the imaginary part of S11 predicted by the neural network model; It is the real part of S12 predicted by the neural network model. It is the imaginary part of S12 predicted by the neural network model; It is the real part of S21 predicted by the neural network model. It is the imaginary part of S21 predicted by the neural network model; It is the real part of S22 predicted by the neural network model. S22 is the imaginary part of the neural network model prediction; S11, S12, S21 and S22 are all scattering parameters. After each training round, the training loss for batch B is: (10) in, The training loss of batch B after each round of training. This represents the training loss for a single sample in the i-th batch after each round of training. After each round of training, the final training loss is: (11) in, This represents the final training loss at the end of each training round. This represents the total number of training data in batch B during one round of training.

8. The method according to claim 1, characterized in that, Step 2 includes: In each training round, a predetermined proportion of data is extracted from the original measured data as a validation dataset to validate the training loss, and the validation loss formula is defined as follows: (12) in, To verify the loss, To verify the total number of samples in the dataset, To validate the dataset, To verify the first in the dataset One sample, For the first The loss value for each sample.

9. The method according to claim 1, characterized in that, The step of obtaining the current load impedance under the preset reflection coefficient includes: The load impedance is obtained using the following formula: (13) in, Given the current load impedance, , , , These are all two-port scattering parameters of the waveguide. For a fixed reflection coefficient, This is the characteristic impedance.

10. The method according to claim 9, characterized in that, Characteristic impedance The expression is: (14) in, Let b be the length of the waveguide, c be the height of the waveguide, c be the speed of light in a vacuum, and f be the frequency of the electromagnetic wave. and These are the permeability constant and dielectric constant of the waveguide filler, respectively.