CN113128119A - Filter reverse design and optimization method based on deep learning - Google Patents
Filter reverse design and optimization method based on deep learning Download PDFInfo
- Publication number
- CN113128119A CN113128119A CN202110431909.2A CN202110431909A CN113128119A CN 113128119 A CN113128119 A CN 113128119A CN 202110431909 A CN202110431909 A CN 202110431909A CN 113128119 A CN113128119 A CN 113128119A
- Authority
- CN
- China
- Prior art keywords
- filter
- network
- neural network
- reverse
- response curve
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/17—Mechanical parametric or variational design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/06—Multi-objective optimisation, e.g. Pareto optimisation using simulated annealing [SA], ant colony algorithms or genetic algorithms [GA]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Artificial Intelligence (AREA)
- Geometry (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Hardware Design (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Mathematics (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Analysis (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physiology (AREA)
- Genetics & Genomics (AREA)
- Control Of Motors That Do Not Use Commutators (AREA)
Abstract
The invention belongs to the technical field of microwave filter design, and particularly relates to a filter reverse design and optimization method based on a deep learning algorithm. The invention designs the structure parameters of the filter according to the electromagnetic response change of the filter; in the design, a reverse network is trained by using forward network generated data, deep learning is carried out, and genetic algorithm optimization is combined: training a reverse network by using data generated by the forward network and combining the data obtained by simulation; the electromagnetic response curve of the filter is obtained by synthesizing Chebyshev polynomials; taking the target electromagnetic response curve as the input of a reverse neural network to obtain the initial value of the structural parameter; inputting the initial value into a genetic algorithm and a forward neural network for iterative optimization; and the optimization target is that the difference between the electromagnetic response curve output by the forward neural network and the target electromagnetic response curve is minimum, and finally, the optimized filter response curve is output, and the structural parameters of the final filter are obtained.
Description
Technical Field
The invention belongs to the technical field of filters, and particularly relates to a design method of a filter.
Background
In a modern base station electromagnetic wave transceiving system, frequency bands in which different antennas operate are often different, and a filter is generally required at the back end of the modern base station electromagnetic wave transceiving system to control transceiving of electromagnetic waves in a required frequency band. With the rapid development of communication systems, the requirements for communication quality are continuously increasing, and the design indexes of back-end microwave filters become more severe. With the increasing number of terminal devices, the increasingly crowded frequency spectrum and the complex electromagnetic environment, the function of the filter as a passive device is more and more important. In the traditional filter design, a corresponding filter function is often selected according to performance indexes such as pass-band return loss, out-of-band rejection and the like, a proper filter prototype is determined, coupling coefficients are extracted by simulation software, the relation between the actual model coupling coefficients and the ideal filter coupling coefficients is established, then a circuit model is used for adjustment, and finally overall optimization is carried out. The conventional design method is very cumbersome and the potential coupling affects the extraction of parameters, so that a new method for automatically, accurately and rapidly designing the filter is needed.
In recent years, along with the continuous improvement of computer performance and the continuous improvement of artificial intelligence theory, the method has been effectively applied to many fields. The well-trained neural network has strong generalization capability and robustness, and is very suitable for realizing rapid and accurate design under different design indexes. For reverse design of the device, too few data sets may cause the neural network to be affected by non-unique solutions and fall into local optima. The existing solution is to obtain more data sets through simulation, but without doubt increases the time consumption. The existing design based on evolutionary algorithms such as a particle swarm optimization algorithm has strong dependence on initial values, simulation software needs to be continuously used in an iteration process, time is consumed, a trained reverse neural network can provide a value close to global optimum, the forward network can be used as a proxy model and combined with the optimization algorithm to complete the design of a filter within dozens of seconds, and due to the fact that the calculation resources are not greatly occupied, parallel design of multiple design targets can be conveniently conducted, and the method has obvious advantages compared with a traditional design method and an optimization algorithm.
Disclosure of Invention
The invention aims to provide an automatic, rapid and accurate filter reverse design and optimization method based on a deep learning algorithm.
The method for reversely designing and optimizing the filter based on the deep learning algorithm, provided by the invention, gives an explanation by taking the design of the cavity interdigital filter as an example, and is also suitable for the structural design of other types of filters, including the design of a microstrip filter, a dielectric filter and the like.
The invention provides a reverse design and optimization method of a filter based on a deep learning algorithm, which is mainly designed aiming at the structural parameters of the filter, wherein the structural parameters of the filter are reflected by an electromagnetic response curve of the filter, and particularly, the electromagnetic response curve of the filter is obtained by synthesizing Chebyshev polynomials; selecting a Chebyshev polynomial with a proper order according to the insertion loss, the center frequency and the working bandwidth of the required filter, and synthesizing an electromagnetic response curve in a frequency band range;
the reverse design and optimization method of the filter adopts a deep learning algorithm, comprises the steps of training a reverse neural network to predict an initial value of a structural parameter by using a forward network generation data set, optimizing the forward neural network and a genetic algorithm by using the predicted initial value, performing deep learning, and finally obtaining the structural parameter of the filter: the method comprises the following specific steps:
(1) firstly, importing a data set obtained by simulation software into a forward neural network, training, and obtaining a forward neural network model after a convergence condition is reached;
(2) enabling the forward neural network model to generate more data by self, and forming a mixed data set by the data and the simulation data to train a reverse neural network to obtain a reverse network model;
(3) then, a target electromagnetic response curve synthesized by the Chebyshev polynomial is used as the input of a reverse network model, and the initial value of the structural parameter is predicted by a reverse neural network;
(4) inputting the initial value into a genetic algorithm, and generating a structural parameter in a range near the initial value by the genetic algorithm; and then taking the structural parameters as the input of a forward network to carry out iterative optimization: optimizing according to an electromagnetic response curve which is synthesized by Chebyshev polynomials and accords with design indexes, wherein the optimization target is that the difference between the electromagnetic response curve output by the forward neural network and the response curve according to the target is minimum, when a loss function reaches a threshold value or an algorithm reaches a specified maximum iteration step number, the optimization is stopped, and finally, the optimized response curve is output, and the final structural parameters are obtained.
The reverse design and optimization method of the present invention is further specifically described below by taking the design of the cavity interdigital filter as an example.
The cavity interdigital filter can be processed by adopting a standard aluminum alloy integral processing and forming process. The basic structure is as follows: the outer cavity is a metal wall and comprises an input tap, an output tap and four sections of coupling resonance metal rods; key structural parameters of the filter include: the length of the four metal rods is sequentially a first section L1, a second section L2, a third section L2, a fourth section L1, a first second section spacing d12, a second third section spacing d23 and a third fourth section spacing d12 from top to bottom; the widths of the four sections of coupling metal rods are equal, the lengths of the first section of metal rod and the fourth section of metal rod are equal and are L1, and the lengths of the second section of metal rod and the third section of metal rod are equal and are L2; namely, the cavity interdigital filter needs to optimize the structural parameters of L1, L2, d12 and d 23.
In the design method, the electromagnetic response S11 curve of the filter is obtained by synthesizing the Chebyshev polynomials, and particularly, the Chebyshev polynomials with proper orders are selected according to the insertion loss, the center frequency and the working bandwidth of the required filter, so that the filter response curve in a frequency band range is synthesized. For example, the fourth order chebyshev polynomial used is:
T 4=8x 4-8x 2+1；
the target filter response curve can be obtained by a Chebyshev polynomial:
wherein k isδTo control the parameters, k determines the maximum amplitude of the filter S11 within the desired bandwidth,δis a constant small quantity.
In the design method, because the precision of the reverse neural network on a small number of data sets is not high, and extra time consumption is caused by using simulation software to obtain a large amount of data, firstly, simulation data is input into the forward network for training, then, a large amount of data is generated by using a forward model, and at the moment, the generation speed is thirty times faster than the simulation speed; then, the generated data and the simulated data are mixed to train an inverse neural network, and a target filtering S11 curve synthesized by Chebyshev polynomials is used as input to obtain more accurate structural parameter prediction; because a plurality of local optimal values exist in the solution space, optimization can be prevented from falling into the local optimal values by using more accurate reverse prediction, structural parameters and a filter response curve are directly output after the optimization meets design indexes, and otherwise, next-step optimization is carried out; and taking the output value of the reverse neural network as an initial value of the forward neural network and the genetic algorithm, performing iterative optimization, wherein the target curve is still a filter response curve synthesized by the Chebyshev polynomial, the optimization target is that the difference between the filter response curve output by the forward neural network and the target filter response curve is minimum, and when the loss function reaches a threshold value or the algorithm reaches a specified maximum iteration step number, the optimization is stopped, so that the final structural parameters are obtained.
In the design method of the present invention, the architecture of the inverse neural network is shown in fig. 3, and specifically includes: the system comprises an input layer, twelve fully-connected hidden layers, corresponding Dropout layers, a LeakyReLU activation function and an output layer; wherein each of the twelve hidden layers has 140 neurons, and the value of Dropout is 0.07; the loss function is an MSE mean square error function, the training period is 1000 epochs, and the training target is that the loss function of the predicted structure parameters and the real structure parameters or the structure parameters generated by the forward network is the minimum; judging whether the network is trained well or not by observing the convergence of the loss functions of the training set and the test set and the prediction performance of the network on the new data set; if the network performance after training is poor, the hyper-parameters such as the size of the data set generated by the forward network, the training period, the loss function and the like need to be adjusted.
In the design method of the present invention, the architecture of the forward neural network is shown in fig. 4, and specifically includes: the input layer, 8 layers of full connection layer containing LeakyReLU activation function, 16 layers of full connection layer containing Dropout layer and LeakyReLU activation function, and output layer; wherein, the full-connection layer of the input layer has 4 neurons, and each of the 8 full-connection layers has 80 neurons; among 16 full-junction layers, each layer has 80 neurons; in the output layer, the total connection layer is 1 neuron; the loss function is an MSE mean square error function; training period 1000 epochs; the training target is that the loss function of the predicted S11 curve and the real S11 curve is minimum, and whether the network is trained well is judged by observing the convergence of the loss functions of the training set and the test set and the prediction performance of the network on a new data set; if the network performance after training is poor, the super parameters such as the size of the data set, the training period, the loss function and the like need to be adjusted.
In the design method, a forward network is trained by simulation data, a more data set is generated after a trained forward model is obtained, and the generated data and the simulation data form mixed data to train a reverse neural network; the size of the data set is adjusted according to the complexity of the problem, and a better neural network can be obtained by using fewer data sets as far as possible. The obtained data set is divided into a training set and a test set to check the convergence of network training. The input of the reverse neural network is the electromagnetic response of the filter, here is a filter response curve S11 (for example, 251 frequency points are selected), and the output is four key structure parameters (L1, L2, d12 and d 23) of the filter; the input of the forward neural network is four structural parameters and a frequency value, and the output is an S11 value at the frequency point; the input of the reverse neural network is a complete S11 curve, and the output is four structural parameters; the function of electromagnetic simulation software can be realized after the forward neural network training is finished, and the performance of the filter can be quickly predicted; the reverse network trained by the mixed data set can obtain more accurate structural parameter values and can be used as initial values, and optimization is prevented from falling into local optimal values.
In the design method, the population number adopted by the genetic algorithm is 20, and the iteration number is 15. The objective function is the MAE mean absolute error function F:
where n and m are the number of key frequency points selected from the target curve of S11, m represents four points relatively above, n represents three points relatively below,
for the value of S11 at the nth critical frequency point located relatively below on the designed target curve,
predicting the value of S11 at the nth key frequency point for the forward neural network.
The invention provides a reverse design and optimization method, the simulation data and the generated data of the forward network are used for training the reverse network together, compared with the method only using the simulation data to train, the time is saved, and the accuracy of the reverse design can be greatly improved; the initial values of the structural parameters generated by good reverse prediction can avoid some local optimal values before optimization, so that the genetic algorithm can realize rapid optimization in small-scale population and few iteration times; the forward neural network is used for replacing the function of electromagnetic simulation software, so that the rapid iteration of the genetic algorithm can be realized, and the rapid and accurate design of the filter can be finally realized.
Drawings
Fig. 1 shows the basic structure of a cavity interdigital filter.
Fig. 2 shows the overall architecture of the filter design method.
Fig. 3 is a reverse network structure of the artificial neural network.
Fig. 4 is a forward network structure of an artificial neural network.
Fig. 5 shows the design result of filter a. Comprises an objective function curve and an output result curve of the inverse neural network.
Fig. 6 shows the design result of the filter B. The method comprises an objective function curve, an output result curve of the inverse neural network and an optimization result of the set calculation method.
Reference numbers in the figures: 1. 2 is an input or output tap, 3 is a metal resonance rod, and 4 is a metal cavity.
Detailed Description
As shown in fig. 1, the four-step metal cavity interdigital filter is selected, wherein 1 and 2 are input or output taps, 3 is a metal resonant rod, and 4 is a metal cavity. The four key structural parameters of the filter are L1, L2, d12 and d 23. A range of suitable structural parameters was set, within which 900S 11 curves were generated using electromagnetic simulation software, the frequency range of the S11 curve was 0.5 GHz to 4.5 GHz, and there were 251 data points per curve within this frequency range. The 900 samples are made into data sets, 800 are training sets, 100 are testing sets, firstly, the forward neural network shown in fig. 4 is trained, after 1000 epochs are trained, the forward network model automatically generates 3600 data sets, and the 3600 data sets and the former 900 data are combined to form mixed 4500 data so as to train the reverse neural network shown in fig. 3. The loss functions of the two networks are MSE mean square error functions, and good results are obtained after the reverse neural network is trained for 1000 epochs.
As shown in fig. 3, the schematic diagram of the structure of the inverse neural network includes an input layer, an output layer and twelve hidden layers, and after training, a LeakyReLU activation function is used, and a target filter curve is input, and a desired filter structure parameter is output.
As shown in fig. 4, it is a schematic diagram of a forward neural network structure, which includes an input layer, an output layer and twenty-four hidden layers, and adopts the leak relu function, the input is the structure parameters and the frequency, and the output is the S11 value at the corresponding frequency.
As shown in FIG. 5, for the design result of the filter A, the horizontal axis is frequency, the vertical axis is S11 parameter, and the goal is to make S11 within 1.5-2.1 GHz<-24 dB. The target curve is given by a fourth order chebyshev polynomial,δsetting the value to be 0.0002, wherein the value is shown by dotted lines in the figure and comprises 251 frequency points, inputting the frequency points into the trained reverse neural network, outputting predicted filter structure parameters (L1, L2, d12 and d 23), simulating an S11 curve in simulation software based on the structure parameters, and comparing the structure parameters with a target curve as shown by solid lines in the figure, so that the reverse design of the reverse neural network meets the design index and directly outputs the structure parameters and the corresponding S11 curve.
As shown in FIG. 6, for the design result of the filter B, the horizontal axis is frequency, the vertical axis is S11 parameter, and the goal is to make S11 in the 2.1-2.9 GHz band<-24 dB. The target curve is given by a fourth order chebyshev polynomial,δset to 0.0002. As shown by dotted lines in the figure, the simulated model comprises 251 frequency points, the 251 frequency points are input into the trained inverse neural network, predicted filter structure parameters (L1, L2, d12 and d 23) are output, an S11 curve is simulated in simulation software based on the structure parameters, and as shown by the dotted lines in the figure, the simulated model is compared with a target curve, the inverse neural network gives a result approximate to a target, but has a gap with the target, and the passband bandwidth is not enough. Inputting the predicted filter structure parameters into a trained forward neural network, predicting a corresponding S11 curve, and constructing an MAE mean absolute error target function F with the target curve;
the key points are selected as extreme points in the passband, as indicated by the circle point identifiers in the figure. And if the F value is lower than the threshold value 10, the design index is considered to be met, and the structural parameters and the corresponding S11 curve are output. And (3) optimizing the target function by a genetic algorithm due to the fact that the requirement of the threshold is not met, setting the threshold to be 10, and continuously calling the forward neural network to complete optimization through 20 populations and iteration for 15 times. The final optimization result is shown as a solid line in the figure, is basically consistent with the target curve, and reaches the design target.
The two design examples shown in fig. 5 and fig. 6 demonstrate the effectiveness of the design method described in fig. 2, which shows effectiveness for filter designs of different center frequencies and different bandwidths.
Claims (5)
1. A reverse design and optimization method of a filter based on a deep learning algorithm is carried out aiming at the structural parameter design of the filter, the structural parameter of the filter is reflected by an electromagnetic response curve of the filter, and particularly the electromagnetic response curve of the filter is obtained by synthesizing Chebyshev polynomials; the method is characterized in that a deep learning algorithm is adopted, the method comprises the steps of training a reverse neural network to predict an initial value of a structural parameter by using a forward network generation data set, optimizing the forward neural network and a genetic algorithm by using the predicted initial value, performing deep learning, and finally obtaining the structural parameter of a filter: the method comprises the following specific steps:
(1) firstly, importing a data set obtained by simulation software into a forward neural network, training, and obtaining a forward neural network model after a convergence condition is reached;
(2) enabling the forward neural network model to generate more data by self, and forming a mixed data set by the data and the simulation data to train a reverse neural network to obtain a reverse network model;
(3) then, a target electromagnetic response curve synthesized by the Chebyshev polynomial is used as the input of a reverse network model, and the initial value of the structural parameter is predicted by a reverse neural network;
(4) inputting the initial value into a genetic algorithm, and generating a structural parameter in a range near the initial value by the genetic algorithm; and then taking the structural parameters as the input of a forward network to carry out iterative optimization: optimizing according to an electromagnetic response curve which is synthesized by Chebyshev polynomials and accords with design indexes, wherein the optimization target is that the difference between the electromagnetic response curve output by the forward neural network and the response curve according to the target is minimum, when a loss function reaches a threshold value or an algorithm reaches a specified maximum iteration step number, the optimization is stopped, and finally, the optimized response curve is output, and the final structural parameters are obtained.
2. The method of claim 1, wherein the cavity interdigital filter has the following basic structure: the outer cavity is a metal wall and comprises an input tap, an output tap and four sections of coupling resonance metal rods; key structural parameters of the filter include: the length of the four metal rods is sequentially a first section L1, a second section L2, a third section L2, a fourth section L1, a first second section spacing d12, a second third section spacing d23 and a third fourth section spacing d12 from top to bottom; the widths of the four sections of coupling metal rods are equal, the lengths of the first section of metal rod and the fourth section of metal rod are equal and are L1, and the lengths of the second section of metal rod and the third section of metal rod are equal and are L2; namely, the cavity interdigital filter needs to optimize the structural parameters of L1, L2, d12 and d 23;
the electromagnetic response S11 curve of the filter is obtained by synthesizing Chebyshev polynomials, specifically, the Chebyshev polynomials with proper orders are selected according to the insertion loss, the center frequency and the working bandwidth of the required filter, and the filtering response curve in a frequency band range is synthesized; the fourth order chebyshev polynomial used is:
T 4=8x 4-8x 2+1；
obtaining a target filter response curve by a Chebyshev polynomial:
wherein k isδTo control the parameters, k determines the maximum amplitude of the filter S11 within the desired bandwidth,δis a constant small quantity.
3. The method for inverse design and optimization of a filter according to claim 1 or 2, wherein the inverse neural network specifically comprises: the system comprises an input layer, twelve fully-connected hidden layers, corresponding Dropout layers, a LeakyReLU activation function and an output layer; wherein each of the twelve hidden layers has 140 neurons, and the value of Dropout is 0.07; the loss function is an MSE mean square error function, the training period is 1000 epochs, and the training target is that the loss function of the predicted structure parameters and the real structure parameters or the structure parameters generated by the forward network is the minimum; judging whether the network is trained well or not by observing the convergence of the loss functions of the training set and the test set and the prediction performance of the network on the new data set; if the network performance after training is poor, the hyper-parameters of the size of the data set generated by the forward network, the training period and the loss function need to be adjusted.
4. The method for inverse design and optimization of a filter according to claim 1 or 2, wherein the forward neural network specifically comprises: the input layer, 8 layers of full connection layer containing LeakyReLU activation function, 16 layers of full connection layer containing Dropout layer and LeakyReLU activation function, and output layer; wherein, the full-connection layer of the input layer has 4 neurons, and each of the 8 full-connection layers has 80 neurons; among 16 full-junction layers, each layer has 80 neurons; in the output layer, the total connection layer is 1 neuron; the loss function is an MSE mean square error function; training period 1000 epochs; the training target is that the loss function of the predicted S11 curve and the real S11 curve is minimum, and whether the network is trained well is judged by observing the convergence of the loss functions of the training set and the test set and the prediction performance of the network on a new data set; if the network performance after training is poor, the hyper-parameters of the data set size, the training period and the loss function need to be adjusted.
5. The method for reverse design and optimization of filters according to claim 1 or 2, characterized in that the genetic algorithm uses a population number of 20 and the number of iterations is 15; the objective function is the MAE mean absolute error function F:
where n and m are the number of key frequency points selected from the target curve of S11, m represents four points relatively above, n represents three points relatively below,
for the value at the nth critical frequency point whose position on the designed S11 target curve is relatively lower,
the value at the nth critical frequency point on the target curve of S11 predicted for the forward neural network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110431909.2A CN113128119B (en) | 2021-04-21 | 2021-04-21 | Filter reverse design and optimization method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110431909.2A CN113128119B (en) | 2021-04-21 | 2021-04-21 | Filter reverse design and optimization method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113128119A true CN113128119A (en) | 2021-07-16 |
| CN113128119B CN113128119B (en) | 2022-04-12 |
Family
ID=76778961
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110431909.2A Active CN113128119B (en) | 2021-04-21 | 2021-04-21 | Filter reverse design and optimization method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113128119B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116111984A (en) * | 2022-12-06 | 2023-05-12 | 中国电子科技集团公司信息科学研究院 | Filter design optimization method and device, filter, equipment and medium |
| WO2023155558A1 (en) * | 2022-02-21 | 2023-08-24 | 浙江大学 | On-chip filter inverse design method based on spatial mapping of equivalent circuit |
| WO2023240772A1 (en) * | 2022-06-16 | 2023-12-21 | 苏州大学 | Silicon-based optical micro-ring filter reverse design method based on sparsity calculation |
| CN117668954A (en) * | 2024-01-30 | 2024-03-08 | 中国电子科技集团公司信息科学研究院 | Design method and system for resonance structure of super-surface band-pass filter |
| CN117668954B (en) * | 2024-01-30 | 2024-05-31 | 中国电子科技集团公司信息科学研究院 | Design method and system for resonance structure of super-surface band-pass filter |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101997516A (en) * | 2009-08-20 | 2011-03-30 | 中国计量学院 | Method for designing combined filter based on improved back propagation (BP) neutral network |
| CN108170922A (en) * | 2017-12-21 | 2018-06-15 | 中国地质大学（武汉） | A kind of aided debugging method of microwave filter, equipment and storage device |
| CN110443364A (en) * | 2019-06-21 | 2019-11-12 | 深圳大学 | A kind of deep neural network multitask hyperparameter optimization method and device |
-
2021
- 2021-04-21 CN CN202110431909.2A patent/CN113128119B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101997516A (en) * | 2009-08-20 | 2011-03-30 | 中国计量学院 | Method for designing combined filter based on improved back propagation (BP) neutral network |
| CN108170922A (en) * | 2017-12-21 | 2018-06-15 | 中国地质大学（武汉） | A kind of aided debugging method of microwave filter, equipment and storage device |
| CN110443364A (en) * | 2019-06-21 | 2019-11-12 | 深圳大学 | A kind of deep neural network multitask hyperparameter optimization method and device |
Non-Patent Citations (1)
| Title |
|---|
| LI YAPING ETAL: "an artificial neural network assisted optimization system for analog design space exploration", 《IEEE》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2023155558A1 (en) * | 2022-02-21 | 2023-08-24 | 浙江大学 | On-chip filter inverse design method based on spatial mapping of equivalent circuit |
| WO2023240772A1 (en) * | 2022-06-16 | 2023-12-21 | 苏州大学 | Silicon-based optical micro-ring filter reverse design method based on sparsity calculation |
| CN116111984A (en) * | 2022-12-06 | 2023-05-12 | 中国电子科技集团公司信息科学研究院 | Filter design optimization method and device, filter, equipment and medium |
| CN117668954A (en) * | 2024-01-30 | 2024-03-08 | 中国电子科技集团公司信息科学研究院 | Design method and system for resonance structure of super-surface band-pass filter |
| CN117668954B (en) * | 2024-01-30 | 2024-05-31 | 中国电子科技集团公司信息科学研究院 | Design method and system for resonance structure of super-surface band-pass filter |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113128119B (en) | 2022-04-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113128119B (en) | Filter reverse design and optimization method based on deep learning | |
| CN109783905B (en) | Intelligent microwave cavity filter debugging method based on particle swarm optimization algorithm | |
| CN112989508B (en) | Filter optimization design method based on deep learning algorithm | |
| CN108182316B (en) | Electromagnetic simulation method based on artificial intelligence and electromagnetic brain thereof | |
| CN115809624B (en) | Automatic analysis design method for integrated circuit microstrip line transmission line | |
| Yu et al. | State-of-the-art: AI-assisted surrogate modeling and optimization for microwave filters | |
| CN114444390A (en) | Design method of vacuum electronic device slow wave structure based on deep learning and machine learning | |
| CN112149360A (en) | ALO-LMBP neural network-based double-notch characteristic ultra-wideband antenna reverse modeling research method | |
| Hammoodi et al. | Bayesian regularization based ANN for the design of flexible antenna for UWB wireless applications | |
| CN112669216B (en) | Super-resolution reconstruction network of parallel cavity new structure based on federal learning | |
| Gu et al. | Intelligent design of arbitrary bandstop FSS through deep learning and genetic algorithm | |
| Shibata et al. | A Novel Convolutional-Autoencoder Based Surrogate Model for Fast S-parameter Calculation of Planar BPFs | |
| Liu et al. | Adapruner: Adaptive channel pruning and effective weights inheritance | |
| Raya et al. | Design of inset fed rectangular shaped microstrip patch antenna using deep neural networks | |
| Roy et al. | ANN model development for tunable bandpass filter | |
| Wei et al. | Machine Learning-Assisted Automatic Filter Synthesis with Prior Knowledge and Its Application to Single-Mode Bandpass Filter Design | |
| CN117688901B (en) | Circuit design method, system, device and storage medium for high-frequency discrete filter | |
| CN117494565A (en) | Non-traditional low-pass filter based on artificial intelligence and design method | |
| CN114492253B (en) | Microstrip circuit half-space mapping rapid optimization method combined with Gaussian process | |
| Li et al. | Bandpass filter design by artificial neural network modeling | |
| Angiulli et al. | Metamaterial unit cell characterization by using a multi-fidelity surrogate modelling approach | |
| CN117077604B (en) | CSRR filter design method and device based on reinforcement learning | |
| CN116663435B (en) | Damping bus structure parameter optimization method, damping bus structure parameter optimization device and storage medium | |
| Liu et al. | Geometry scaling of microwave filters using an adaptive homotopy continuation method | |
| Dautov et al. | ML-enabled Resonance Prediction for SGP-based WPT |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |