Filter Design Method and Device for Neural Network-Assisted Implicit Space Mapping

Abstract

A filter design method and device based on neural network-assisted implicit space mapping. The method includes: establishing a rough model; feeding design parameters and auxiliary parameters into the rough model to obtain a response; using each set of design parameters and their corresponding responses as input vectors of the neural network, and using the auxiliary variables as output vectors of the neural network to form training samples; training the neural network with the training samples; when the rough model is optimized to the optimal state, selecting the design parameters X c 1 ; establishing a fine model, and making its design variables X f 1 = X c 1 to obtain a response R f ( X f 1 ); feeding the response R f ( X f 1 ) and the design variables X f 1 into the trained neural network to obtain auxiliary variables; feeding the auxiliary variables into the rough model, and when optimized to the optimal state, the design parameters are X c 2 ; making X f 2 = X c 2 to obtain a response R f ( X f 2 ); until the requirements are met. If not, return the response and design variables of the new accurate model to the auxiliary variable acquisition step to continue execution. The parameter extraction process in the implicit space mapping is simplified through the neural network, improving the design efficiency and accuracy.

Description

Technical Field

[0001] This application relates to the field of microstrip filter design, and particularly to a filter design method and device based on neural network assisted implicit space mapping. Background Art

[0002] With the rapid development of electronic technology, radio frequency (RF) and microwave circuits are widely used in various fields such as mobile communication, radar, test and measurement, etc. The requirements for RF and microwave devices are also constantly increasing. The research on microwave devices is continuously developing towards the trends of high bandwidth, high performance, low cost, miniaturization and integration, and its implementation structure is becoming increasingly complex.

[0003] As an important microwave passive device, how to achieve its high performance and improve its design speed has become an important challenge in filter design and research. In engineering design, full-wave electromagnetic simulation technology is often used for filter design. Full-wave electromagnetic simulation technology can accurately simulate the actual working response of the filter, but full-wave electromagnetic simulation often takes a lot of time to complete the filter design work; while the equivalent circuit model of the filter can achieve fast design and optimization work, but the simulation accuracy of this method is far less than that of full-wave electromagnetic simulation technology, and the results obtained are often very different from the actual results. Summary of the Invention

[0004] To solve the above deficiencies of the prior art, this application provides a filter design method and device based on neural network assisted implicit space mapping, which simplifies the parameter extraction process in implicit space mapping through a neural network, and can improve the design efficiency and design accuracy.

[0005] To achieve the above object, the present invention adopts the following technologies:

[0006] A filter design method based on neural network assisted implicit space mapping, comprising the steps of:

[0007] Establish an equivalent circuit model of the microstrip filter as a rough model;

[0008] Select a set of design parameters related to the filter structure X c , and select a set of auxiliary parameters X , to form a set of input vectors CX and send them into the rough model for simulation to obtain a set of response vectors CY ;

[0009] For each set of design parameters X c and its corresponding response vector CY as the input vector of the neural network NX , and the auxiliary variable XAs the output vector of the neural network NY , a training sample is formed;

[0010] Train the neural network using the training sample;

[0011] Optimize the rough model to make its response optimal, and select the design parameters at this time X c 1 ;

[0012] Establish a full-wave electromagnetic simulation model of the filter as the fine model, and set its design variables X f 1 = X c 1 Perform full-wave electromagnetic simulation to obtain the response R f ( X f 1 );

[0013] Auxiliary variable acquisition: Input the response R f ( X f 1 ) and the design variables X f 1 into the trained neural network to obtain the auxiliary variable X 1 ;

[0014] Send the auxiliary variable X 1 into the rough model, optimize the design parameters to make the response optimal, and the design parameters at this time X c 2 ;

[0015] Set the design variables X f 2 = X c 2 and send them into the fine model to obtain the response R f ( X f 2 );

[0016] If the response design requirements of the fine model have been met, end;

[0017] If not, return to the auxiliary variable acquisition step, use the response of the latest obtained accurate model and the design variables as the input of the trained neural network, and continue to execute.

[0018] A filter design setting for neural network-assisted implicit space mapping, including:

[0019] A rough model establishment unit for establishing an equivalent circuit model of a microstrip filter as a rough model;

[0020] A rough model simulation unit for selecting a set of design parameters related to the filter structure X c , and selecting a set of auxiliary parameters X , to form a set of input vectors CX and sending them into the rough model for simulation to obtain a set of response vectors CY ;

[0021] A sample unit for using each set of design parameters X c and their corresponding response vectors CY as the input vectors of the neural network NX , using the auxiliary variable X as the output vector of the neural network NY , and forming training samples;

[0022] A training unit for training the neural network using the training samples;

[0023] An initial optimization unit for optimizing the rough model to make its response optimal, and selecting the design parameters at this time X c 1 ;

[0024] A fine model establishment and simulation unit for establishing a full-wave electromagnetic simulation model of the filter as a fine model, and setting its design variables X f 1 = X c 1 to perform full-wave electromagnetic simulation to obtain the response R f ( X f 1 );

[0025] An auxiliary variable acquisition unit for using the response R f ( X f 1 ) and the design variables X f1 Input the trained neural network to obtain auxiliary variables X 1 ;

[0026] An iterative optimization unit for sending the auxiliary variables X 1 into a rough model and optimizing the design parameters to make the response optimal. At this time, the design parameters X c 2 ;

[0027] An iterative simulation unit for setting the design variables X f 2 = X c 2 and sending them into a fine model to obtain a response R f ( X f 2 )

[0028] A judgment and jump unit for judging whether the response obtained by the iterative simulation unit has met the response design requirements of the fine model:

[0029] If so, end;

[0030] If not, use the response and design variables of the latest obtained accurate model as the input of the trained neural network, and jump to the auxiliary variable acquisition unit to continue to execute the auxiliary variable acquisition unit, iterative optimization unit, iterative simulation unit, and judgment and jump unit in sequence.

[0031] The beneficial effects of the present invention are as follows:

[0032] 1. Simplify the parameter extraction process in implicit space mapping through a neural network, which can improve design efficiency and design accuracy;

[0033] 2. Use the neural network as the inverse model of the rough model, eliminating the need to repeatedly update the rough model circuit during parameter extraction and directly using software code to achieve high automation. Brief Description of the Drawings

[0034] Figure 1 is the flowchart of the method according to the embodiment of the present application.

[0035] Figure 2 is the block diagram of the device structure according to the embodiment of the present application. Detailed Embodiments

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will describe the embodiments of the present invention in detail with reference to the accompanying drawings. However, the embodiments described herein are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0037] In one aspect of the embodiments of the present application, a filter design method assisted by a neural network for implicit space mapping is provided. As Figure 1 shown, it includes the following steps:

[0038] S100. Establish a rough model

[0039] According to the design requirements, perform filter synthesis design to obtain a coupling matrix, select the type of filter, and establish an equivalent circuit simulation model of the microstrip filter based on the coupling matrix.

[0040] S200. Obtain a training set

[0041] Select a set of design parameters X c , such as the length of the resonator and the spacing between resonators in a microstrip filter. These design parameters are related to the specific filter structure. And select a relevant response vector R f ( X c ). Generally speaking, it is a vector composed of the real part and the imaginary part of the return loss of the filter and the real part and the imaginary part of the insertion loss. This model has a very fast operation speed and can optimize the design parameters X c to make the response of this model reach the optimal.

[0042] Additionally, select a set of auxiliary parameters X， such as the parameters of the substrate in the filter, such as the substrate thickness or the substrate dielectric constant, etc., which together with the design parameters X c constitute a set of input vectors CX for feeding into the rough model for simulation, and a set of response vectors CY can be obtained. Specifically, through different design parameters X c and auxiliary parameters X , N input vectors CX are formed, and the response R f ( X c ) can be obtained as CY .

[0043] For each set of design parameters X c and its corresponding response vector CYAs the input vector of the neural network NX , use the auxiliary variable X as the output vector of the neural network NY to form the training samples.

[0044] Before training, the training samples need to be normalized and randomly divided into a training set and a test set.

[0045] S300. Train the neural network

[0046] Use a BP neural network. The activation function of the hidden layer can be selected as the Sigmoid function, and the output layer uses a linear activation function. The number of neurons in the output layer is the dimension of the auxiliary parameter X . The number of hidden layers and the number of neurons depend on the specific training effect. The training algorithm selects the Levenberg-Marquardt algorithm to train the neural network using the training samples.

[0047] S400. Optimize the rough model

[0048] Optimize the rough model to make its response optimal, and select the design parameter X c 1 at this time. Specifically, certain auxiliary parameters X 1 can be selected and kept unchanged. Optimize the design parameter X c through the optimization algorithm to make the response of the rough model optimal, and let the design variable at this time be X c 1 .

[0049] S500. Establish a fine model and obtain the response of the fine model

[0050] Establish a full-wave electromagnetic simulation model of the filter as the fine model, and let its design variable X f 1 = X c 1 Conduct full-wave electromagnetic simulation to obtain the response R f ( X f 1 );

[0051] S600. Auxiliary variable acquisition

[0052] Use the response of the fine model R f ( X f1 ) and design variables X f 1 Input the trained neural network to obtain auxiliary variables X 1 , as the new auxiliary variables.

[0053] S700. Update the auxiliary variables and optimize the rough model

[0054] Send the new auxiliary variables X 1 into the rough model and keep it unchanged, and re-optimize the design parameters to make the response optimal. The design parameters at this time are X c 2 .

[0055] S800. Update the design variables of the accurate model and obtain the response of the accurate model

[0056] Let the design variables of the accurate model X f 2 = X c 2 , and send them into the fine model to obtain the response of the accurate model R f ( X f 2 ).

[0057] S900. Judge whether the requirements are met:

[0058] If the response design requirements of the fine model have been met, end;

[0059] If not, return to the step of obtaining auxiliary variables in S600, use the latest obtained response and design variables of the accurate model as the input of the trained neural network, and continue to execute in a loop.

[0060] On the other hand, an embodiment of the present application provides a filter design setting for neural network-assisted implicit space mapping, as Figure 2 shown, including a rough model establishment unit, a rough model simulation unit, a sample unit, a training unit, a preliminary optimization unit, a fine model establishment and simulation unit, an auxiliary variable acquisition unit, an iterative optimization unit, an iterative simulation unit, a judgment and jump unit.

[0061] Among them, the rough model establishment unit is used to establish an equivalent circuit model of the microstrip filter as the rough model; the rough model simulation unit is used to select a set of design parameters related to the filter structure X c , and select a set of auxiliary parametersX , constituting a set of input vectors CX are fed into the rough model for simulation to obtain a set of response vectors R f ( X c ) serves as CY . Among them, the design parameter X c can be the length of the resonator in the microstrip filter, the spacing between resonators, etc. This design parameter is related to the specific filter structure; the response vector R f ( X c ), generally speaking, is a vector composed of the real and imaginary parts of the return loss of the filter and the real and imaginary parts of the insertion loss;

[0062] The sample unit is used to use each set of design parameters X c and its corresponding response vector CY as the input vector of the neural network NX , and the auxiliary variable X as the output vector of the neural network NY , forming training samples; and performing normalization processing on the training samples and randomly dividing them into a training set and a test set.

[0063] The training unit is used to train the neural network using the training set and test set of the training samples. The neural network is a BP neural network. The activation function of its hidden layer can choose the Sigmoid function, the output layer uses a linear activation function, the number of neurons in the output layer is the dimension of the auxiliary parameter X , and the number of hidden layers and neurons depends on the specific training effect. The training algorithm chooses the Levenberg-Marquardt algorithm.

[0064] The initial optimization unit is used to optimize the rough model so that its response reaches the optimum, and select the design parameter X c 1 at this time; the fine model establishment and simulation unit is used to establish the full-wave electromagnetic simulation model of the filter as the fine model, and make its design variable X f 1 = X c 1 to perform full-wave electromagnetic simulation and obtain the response R f ( X f 1 ).

[0065] The auxiliary variable acquisition unit is used to obtain the responseR f ( X f 1 ) and design variables X f 1 Input the trained neural network to obtain auxiliary variables X 1 ; The iterative optimization unit is used to send the auxiliary variables X 1 to the rough model, optimize the design parameters to make the response optimal, and at this time the design parameters X c 2 ; The iterative simulation unit is used to make the design variables X f 2 = X c 2 and send them to the fine model to obtain the response R f ( X f 2 ).

[0066] The judgment and jump unit is used to judge whether the response obtained by the iterative simulation unit has reached the response design requirements of the fine model: if so, end; if not, use the response and design variables of the latest obtained accurate model as the input of the trained neural network, jump to the auxiliary variable acquisition unit, and continue to execute the auxiliary variable acquisition unit, iterative optimization unit, iterative simulation unit, and judgment and jump unit in sequence.

[0067] Another aspect of the embodiments of the present application provides an electronic device, including: at least one processor and a memory; wherein, the memory stores computer execution instructions; when the at least one processor executes the computer execution instructions stored in the memory, the at least one processor is caused to execute the filter design method of neural network-assisted implicit space mapping described in the foregoing embodiments.

[0068] Another aspect of the embodiments of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is run by a processor, it controls the device where the storage medium is located to execute the filter design method of neural network-assisted implicit space mapping described in the foregoing embodiments.

[0069] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these changes and modifications.

Claims

1. A filter design method assisted by neural network for implicit space mapping, characterized in that, Including the steps: Establish an equivalent circuit model of the microstrip filter as a rough model; Select a set of design parameters related to the filter structure X c , and select a set of auxiliary parameters X to form a set of input vectors CX and send them into the rough model for simulation to obtain a set of response vectors CY ; With each set of design parameters X c and its corresponding response vector CY as the input vector of the neural network NX , taking the auxiliary variable X as the output vector of the neural network NY , to form a training sample; Train a neural network using training samples; Optimize the rough model to make its response optimal, and select the design parameters at this time X c 1 ; Establish a full-wave electromagnetic simulation model of the filter as the fine model, and set its design variables X f 1 = X c 1 Perform full-wave electromagnetic simulation to obtain the response R f ( X f 1 ) Auxiliary variable acquisition: The response R f ( X f 1 ) and the design variables X f 1 are input into the trained neural network to obtain the auxiliary variable X 1 ; Send the auxiliary variable X 1 into the rough model, optimize the design parameters to make the response optimal, and at this time the design parameters X c 2 ; Let the design variables X f 2 = X c 2 and send them into the fine model to obtain the response R f ( X f 2 ); If the response design requirements of the fine model have been met, end; If not, return to the auxiliary variable acquisition step, use the response of the latest obtained accurate model and the design variables as the input of the trained neural network, and continue to execute.

2. The filter design method for neural network-assisted implicit space mapping according to claim 1, wherein The response vector is a vector composed of the real part and the imaginary part of the return loss of the filter and the real part and the imaginary part of the insertion loss.

3. The filter design method based on neural network-assisted implicit space mapping according to claim 1, wherein, Normalize the training samples, randomly divide them into a training set and a test set, and then perform the training of the neural network.

4. The filter design method of neural network-assisted implicit space mapping according to claim 1, wherein The activation function of the hidden layer of the neural network is the Sigmoid function, and the output layer uses a linear activation function. The number of neurons in the output layer is the dimension of the auxiliary variable X of.

5. The filter design method based on neural network-assisted implicit space mapping according to claim 1, characterized in that The training algorithm of the neural network adopts the Levenberg-Marquardt algorithm.

6. A filter design setting for neural network-assisted implicit space mapping, characterized in that, Including: A rough model establishment unit for establishing an equivalent circuit model of the microstrip filter as a rough model; A rough model simulation unit for selecting a set of design parameters related to the filter structure X c , and selecting a set of auxiliary parameters X to form a set of input vectors CX and sending them into the rough model simulation to obtain a set of response vectors CY ; Sample units for each set of design parameters X c and their corresponding response vectors CY as the input vectors of the neural network NX , taking the auxiliary variable X as the output vector of the neural network NY , to form training samples; A training unit for training a neural network using training samples; Initial optimization unit, used to optimize the rough model to make its response optimal, and select the design parameters at this time X c 1 ; The fine model establishment and simulation unit is used to establish the full-wave electromagnetic simulation model of the filter as the fine model and set its design variables X f 1 = X c 1 Perform full-wave electromagnetic simulation to obtain the response R f ( X f 1 ) Auxiliary variable acquisition unit for obtaining a response R f ( X f 1 ) and design variables X f 1 Input into the trained neural network to obtain auxiliary variables X 1 ; Iterative optimization unit for sending auxiliary variables X 1 to the rough model to optimize the design parameters to make the response optimal. At this time, the design parameters X c 2 ; Iterative simulation unit, used to make design variables X f 2 = X c 2 and send them into the refined model to obtain responses R f ( X f 2 ) A judgment and jump unit for judging whether the response obtained by the iterative simulation unit has met the response design requirements of the fine model: If so, end; If not, use the response of the latest obtained accurate model and the design variables as the input of the trained neural network, jump to the auxiliary variable acquisition unit, and continue to sequentially execute the auxiliary variable acquisition unit, the iterative optimization unit, the iterative simulation unit, and the judgment and jump unit.

7. The filter design setting for neural network-assisted implicit space mapping according to claim 6, wherein The sample unit is also used to normalize the training samples, randomly divide them into a training set and a test set, and then send them to the training unit for training the neural network.

8. The filter design setting for neural network-assisted implicit space mapping according to claim 6, characterized in that The activation function of the hidden layer of the neural network is the Sigmoid function, and the output layer uses a linear activation function. The number of neurons in the output layer is the dimension of the auxiliary variable X of.

9. An electronic device, comprising: At least one processor and a memory; wherein, the memory stores computer execution instructions; characterized in that when the at least one processor executes the computer execution instructions stored in the memory, the at least one processor executes the filter design method of neural network-assisted implicit space mapping according to any one of claims 1 to 5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it controls the device where the storage medium is located to execute the filter design method of neural network-assisted implicit space mapping according to any one of claims 1 to 5.

Images