CN109284541A - A kind of more Method of Physical Modeling of neural network for microwave passive component - Google Patents
A kind of more Method of Physical Modeling of neural network for microwave passive component Download PDFInfo
- Publication number
- CN109284541A CN109284541A CN201811016571.9A CN201811016571A CN109284541A CN 109284541 A CN109284541 A CN 109284541A CN 201811016571 A CN201811016571 A CN 201811016571A CN 109284541 A CN109284541 A CN 109284541A
- Authority
- CN
- China
- Prior art keywords
- physical
- model
- neuron
- output
- neural network
- 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.)
- Pending
Links
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
- G06F30/23—Design optimisation, verification or simulation using finite element methods [FEM] or finite difference methods [FDM]
Abstract
The invention belongs to microwave circuits and devices to model field, provide a kind of more Method of Physical Modeling of the neural network for microwave passive component.The more physical models of microwave passive component are established using artificial neural network.Compared with finite element simulation, the method proposed realizes similar precision with the time using less calculating cost.Sample result shows that device architecture is more complicated, and advantage of the mentioned method in terms of saving calculating time and cost is more obvious.Housebroken model can accurately indicate electromagnetic response of the microwave passive component relative to more physical input parameters.The model proposed can be used for providing the prediction of electromagnetic response accurately and quickly for more physical design processes, this can further shorten the design cycle.
Description
Technical field
The present invention relates to the more physical modeling fields of microwave passive component more particularly to neural net model establishing technology microwave without
The application in source organs weight field.
Background technique
In recent years, the electromagnetic behavior parametric modeling of microwave passive components becomes more and more important electromagnetic design.It is right
It is designed in high-performance radio-frequency/microwave component and system, we are frequently necessary to consider device in the multiple physical field environment of real world
Part behavior further includes other physical domains wherein including not only electromagnetic arts.Therefore the device model in microwave radio field is not only wanted
The behavioral trait of microwave passive component can be described, also to accurately reflect device with the characteristic that physics is different with geometric parameter and causes
Change.More physical Designs centered on electromagnetic property are related to more physical fields such as emi analysis and ermal physics and structural mechanics
Influence, more physical parameters usually require duplicate Electromagnetic Simulation as design variable, therefore more physical simulations are very time-consuming, need
Want a large amount of computer resource.Although already present modeling technique comparative maturity, it is suitable for most existing products, needs pair
When circuit or system modelling containing new device, existing modeling method tends not to reach model accuracy height and emulation speed simultaneously
Spend fast requirement.Therefore the more physical characteristic modeling methods of passive device that research precision is high, modeling period is short are urgently to be resolved
Hot issue.
The more Method of Physical Modeling researchs of passive device have become the important developing direction in microwave designing field.Currently, micro-
More Method of Physical Modeling of wave passive device mainly include two kinds: equivalent-circuit model and limit element artificial module.Correlative study
Existing modeling method is proved with certain application limitation.
Limitation 1: equivalent-circuit model is the model built according to the electromagnetic property of passive device with electronic component, the model
It is more traditional, universal, but need constantly to adjust Model Parameter value in modeling process, parameter is directly to the mutual shadow of model performance
It rings, therefore needs to take a substantial amount of time and obtain suitable parameter value with energy.It is existing with device application environment more sophisticated
Often there is distortion phenomenon in complicated more physical environments in equivalent-circuit model.
Limitation 2: limit element artificial module is to carry out Electromagnetic Simulation based on the more physical messages of device, being capable of accurate outlines device
The model of performance, but this method must rely on numerical simulation software, and the model emulation time is long, and needs are built and prototype
The identical structure of part, this process is difficult, and electromagnetism calculates and needs powerful computer resource, the object during modelling
When reason and geometric parameter are adjusted repeatedly, this is particularly problematic.
Therefore, the purpose of the present invention is to by proposing a kind of more physical modeling sides of the neural network for microwave passive component
Method reduces Modeling Calculation cost, shortens the design cycle, provides and quickly prediction accurate to device electromagnetic behavior.
Summary of the invention
The purpose of the present invention is overcoming the above-mentioned deficiency of the prior art, a kind of nerve net for microwave passive component is proposed
The more Method of Physical Modeling of network.This method training artificial neural network is to learn microwave passive component electromagnetic property and more physical Designs
Relationship between parameter.The trained more physical models of neural network can be provided to be rung about the electromagnetism of more physical design parameters
The efficiently and quickly prediction answered.
A kind of more Method of Physical Modeling of neural network for microwave passive component, including the following steps:
Step 1: building three-layer artificial neural network is classified as three parts: bottom is that input neuron is outer for receiving
Portion's input, last layer are output neuron for exporting more physical responses, are implicit nerve among input layer and output layer
Member handles input signal according to activation primitive and result is transferred to last layer output neuron；
Step 2: the geometric parameter x of passive deviceg, more physical parameter xmWith frequency xfAs the more physical models of passive device
The input signal for inputting neuron, is denoted as x=[xg, xm, xf]T；
Step 3: the set of all output responses of passive device, i.e., the electromagnetic response under more physical environments is as passive device
The output signal of more physical model output neurons, is denoted as y；
Step 4: emulation or actual measurement device are carried out to the passive device in more physical environments with finite element emulation software,
Obtained data or measured data are emulated for training pattern.It obtains establishing accurate model with simulation software or actual measurement in the step
Data, carry out preliminary preparation for subsequent modeling work；
Step 5: the initial ginseng of model is rule of thumb arranged to model training in the device data for being measured or being emulated with step 4
Numerical value, the electromagnetic property of passive device under more physical environments can be expressed by making model substantially；
Step 6: the parameter value of the more physical models of coarse adjustment reduces model error, so that model matches with device data；
Step 7: the parameter of the more physical models of neural network being finely tuned, model accuracy is further increased, makes the electricity of model
Magnetic characteristic is identical as measuring or emulating passive device performance, i.e., when input data is identical, the output signal y and simulation software of model
Gained output (or output obtained by measured data) y*It is equal.
It is realized in the more physical models of passive device proposed by the present invention using three layer perceptron neural network structure, expression
Formula is
Wherein p and q respectively represents the quantity for outputting and inputting neuron.L and k respectively indicate xgAnd xmQuantity, i.e. l+k
=p-1.N indicates the quantity of the hidden neuron determined during neural metwork training.WithIt is corresponding h, r
Or the weight parameter between p input neuron and i-th of hidden neuron,It is i-th of neuron in hidden layer and defeated
Weight parameter between j-th of neuron in layer out.WithRespectively indicate i-th of hidden neuron and j-th of output
The deviation of neuron.These weight parameters determine the non-linear relation output and input between variable.Excitation function uses
Sigmoid function, formula are σ (γ)=1/ (1+e-γ)。
Neural network space reflection modeling method proposed by the present invention does not need microwave passive component internal structure letter not only
Breath, and neural network structure is simple, and electromagnetic property exports under the more physical environments of Optimal Parameters Controlling model, and model respectively inputs ginseng
It is independent between number, reduce calculating cost, shortens modeling period.
Detailed description of the invention
Fig. 1 is structure of the invention block diagram；
Fig. 2 is according to the embodiment of the present invention to the more physical modeling flow charts of microwave passive component.
Fig. 3 is the flow chart that the present invention establishes model.
Fig. 4,5 are that the sample data of the embodiment of the present invention and model export electromagnetic property curve.
Specific embodiment
To make the object, technical solutions and advantages of the present invention clearer, below in conjunction with attached drawing to implementation of the invention
Example is described in detail.
As shown in Fig. 2, first having to obtain in a kind of more Method of Physical Modeling of the neural network of microwave passive component of the present invention
Sample data for model training.Sample data can be obtained by actual measurement device or simulation software.All samples
It is training data that data, which are divided into two groups: one groups, for training neural network model；Another group is test number for verifying model
According to.In order to improve modeling accuracy, data scaling method is used for training data.It is complete according to flow chart shown in Fig. 3 using sample data
It is built at neural network model as shown in figure 1, trained more physical parameter models can be used for fast and accurately completing with electricity
Multiple physical field analysis centered on magnetic.
Neural network structure as shown in Figure 1 is built, model uses 3 layers of perceptron structure: the 1st layer is input layer, is connect
Receiving external input includes geometric parameter, more physical parameters and frequency parameter；2nd layer be neural network hidden layer, it according to swash
Function processing extrinsic neural member living；3rd layer is output layer, it indicates the output response of proposed model.Input sample data are
The geometric parameter x of passive deviceg, more physical parameter xmWith frequency xf, it is denoted as x=[xg, xm, xf]T；Exporting sample data is physics
Electromagnetic response under environment, is denoted as y.
The flow chart shown in Fig. 3 is to the model training in Fig. 1, based on the initial number of experience selection hidden neuron, and
And an appropriate number of hidden neuron can be obtained by repetition test process.Change weight using Fast Newton's training method to join
Number is so that the error between neural network prediction and training sample is as small as possible.For training goal, we use training error
Come measure the more physical nerval network models of microwave passive component learning performance and test error with the predictive ability of measurement model.
The weighted value and hidden layer number in model are adjusted, reduces the error between the output response of model and sample data constantly.
If test error is unsatisfactory for required precision, continue with training data training or adjustment neural network structure, change hidden layer
The number of neuron, re -training.Until training data calculates training error and is expired with the test error that test data calculates
Training process is executed before sufficient required precision always.If test error meets required precision, deconditioning makes model investment
With.
Fig. 4,5 can be seen to establish model output characteristic curve figure compared with sample data using the modeling method of the invention
The curve of output of model is consistent with sample data out.
Claims (2)
1. a kind of more Method of Physical Modeling of neural network for microwave passive component, including the following steps:
Step 1: building three-layer artificial neural network is classified as three parts: bottom is that input neuron is external defeated for receiving
Entering, it is hidden neuron among input layer and output layer that last layer, which is output neuron for exporting more physical responses,
Input signal is handled according to activation primitive and result is transferred to last layer output neuron；
Step 2: the geometric parameter x of passive deviceg, more physical parameter xmWith frequency xfAs the more physical model inputs of passive device
The input signal of neuron is denoted as x=[xg, xm, xf]T；
Step 3: the set of all output responses of passive device, i.e., the electromagnetic response under more physical environments is as the more objects of passive device
The output signal for managing model output neuron, is denoted as y；
Step 4: emulation or actual measurement device, emulation being carried out to the passive device in more physical environments with finite element emulation software
Obtained data or measured data are used for training pattern.Obtain establishing the number of accurate model in the step with simulation software or actual measurement
According to carrying out preliminary preparation for subsequent modeling work；
Step 5: the initial parameter value of model is rule of thumb arranged to model training in the device data for being measured or being emulated with step 4,
The electromagnetic property of passive device under more physical environments can be expressed by making model substantially；
Step 6: the parameter value of the more physical models of coarse adjustment reduces model error, so that model matches with device data；
Step 7: the parameter of the more physical models of neural network being finely tuned, model accuracy is further increased, keeps the electromagnetism of model special
Property with measure or emulation passive device performance is identical, i.e., when input data is identical, obtained by the output signal y and simulation software of model
Output (or output obtained by measured data) y*It is equal.
2. the more Method of Physical Modeling of a kind of neural network for microwave passive component according to claim 1, feature
It is, realized in the more physical models of passive device using three layer perceptron neural network structure, expression formula is
Wherein p and q respectively represents the quantity for outputting and inputting neuron.L and k respectively indicate xgAnd xmQuantity, i.e. l+k=p-
1.N indicates the quantity of the hidden neuron determined during neural metwork training.WithIt is corresponding h, r or p
The weight parameter between neuron and i-th of hidden neuron is inputted,It is i-th of neuron and output in hidden layer
The weight parameter between j-th of neuron in layer.WithRespectively indicate i-th of hidden neuron and j-th of output nerve
The deviation of member.These weight parameters determine the non-linear relation output and input between variable.Excitation function uses
Sigmoid function, formula are σ (γ)=1/ (1+e-γ)。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811016571.9A CN109284541A (en) | 2018-08-31 | 2018-08-31 | A kind of more Method of Physical Modeling of neural network for microwave passive component |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811016571.9A CN109284541A (en) | 2018-08-31 | 2018-08-31 | A kind of more Method of Physical Modeling of neural network for microwave passive component |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109284541A true CN109284541A (en) | 2019-01-29 |
Family
ID=65184058
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811016571.9A Pending CN109284541A (en) | 2018-08-31 | 2018-08-31 | A kind of more Method of Physical Modeling of neural network for microwave passive component |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109284541A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110533319A (en) * | 2019-08-27 | 2019-12-03 | 西安电子科技大学 | A kind of microwave components gold ribbon interconnection transmission performance prediction technique based on interconnection form |
| CN110765704A (en) * | 2019-11-28 | 2020-02-07 | 北京工业大学 | Novel automatic deep neural network modeling method applied to microwave device |
| CN111695230A (en) * | 2019-12-31 | 2020-09-22 | 天津工业大学 | Neural network space mapping multi-physics modeling method for microwave passive device |
| CN113221503A (en) * | 2020-12-31 | 2021-08-06 | 芯和半导体科技(上海)有限公司 | Passive device modeling simulation engine based on machine learning |
| CN115458143A (en) * | 2022-09-16 | 2022-12-09 | 兰州大学 | Radio frequency heating evaluation method of passive implanted medical device based on neural network |
| CN117272778A (en) * | 2023-07-12 | 2023-12-22 | 上海交通大学 | Design method and device of microwave passive device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101971177A (en) * | 2007-11-30 | 2011-02-09 | 科文托尔公司 | System and method for three-dimensional schematic capture and result visualization of multi-physics system models |
| CN104134001A (en) * | 2014-07-30 | 2014-11-05 | 大连海事大学 | Metal back-coated asymmetric coplanar waveguide physical size detection method based on neural network |
| CN106202794A (en) * | 2016-07-21 | 2016-12-07 | 哈尔滨工业大学 | Relay class unit output characteristics acquisition methods based on many finite element simulations |
| CN107308895A (en) * | 2017-06-09 | 2017-11-03 | 四川大学 | A kind of thermostatic type microwave continuous flow reactor |
-
2018
- 2018-08-31 CN CN201811016571.9A patent/CN109284541A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101971177A (en) * | 2007-11-30 | 2011-02-09 | 科文托尔公司 | System and method for three-dimensional schematic capture and result visualization of multi-physics system models |
| CN104134001A (en) * | 2014-07-30 | 2014-11-05 | 大连海事大学 | Metal back-coated asymmetric coplanar waveguide physical size detection method based on neural network |
| CN106202794A (en) * | 2016-07-21 | 2016-12-07 | 哈尔滨工业大学 | Relay class unit output characteristics acquisition methods based on many finite element simulations |
| CN107308895A (en) * | 2017-06-09 | 2017-11-03 | 四川大学 | A kind of thermostatic type microwave continuous flow reactor |
Non-Patent Citations (1)
| Title |
|---|
| SHUXIA YAN等: "Multi-Physics Parametric Modeling of Microwave Passive Components Using Artificial Neural Networks", 《PROGRESS IN ELECTROMAGNETICS RESEARCH M》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110533319A (en) * | 2019-08-27 | 2019-12-03 | 西安电子科技大学 | A kind of microwave components gold ribbon interconnection transmission performance prediction technique based on interconnection form |
| CN110765704A (en) * | 2019-11-28 | 2020-02-07 | 北京工业大学 | Novel automatic deep neural network modeling method applied to microwave device |
| CN110765704B (en) * | 2019-11-28 | 2024-02-06 | 北京工业大学 | Novel deep neural network automatic modeling method applied to microwave device |
| CN111695230A (en) * | 2019-12-31 | 2020-09-22 | 天津工业大学 | Neural network space mapping multi-physics modeling method for microwave passive device |
| CN111695230B (en) * | 2019-12-31 | 2023-05-02 | 天津工业大学 | Neural network space mapping multi-physical modeling method for microwave passive device |
| CN113221503A (en) * | 2020-12-31 | 2021-08-06 | 芯和半导体科技(上海)有限公司 | Passive device modeling simulation engine based on machine learning |
| CN115458143A (en) * | 2022-09-16 | 2022-12-09 | 兰州大学 | Radio frequency heating evaluation method of passive implanted medical device based on neural network |
| CN115458143B (en) * | 2022-09-16 | 2023-05-23 | 兰州大学 | Passive implantable medical device radio-frequency heating evaluation method based on neural network |
| CN117272778A (en) * | 2023-07-12 | 2023-12-22 | 上海交通大学 | Design method and device of microwave passive device |
| CN117272778B (en) * | 2023-07-12 | 2024-03-12 | 上海交通大学 | Design method and device of microwave passive device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109284541A (en) | A kind of more Method of Physical Modeling of neural network for microwave passive component | |
| Kou et al. | Multi-kernel neural networks for nonlinear unsteady aerodynamic reduced-order modeling | |
| Soleimani et al. | Biologically inspired spiking neurons: Piecewise linear models and digital implementation | |
| Zhang et al. | A novel training approach for parametric modeling of microwave passive components using Padé via Lanczos and EM sensitivities | |
| Guo et al. | Physics embedded deep neural network for solving volume integral equation: 2-D case | |
| CN107784180B (en) | Design method of time-varying convex quadratic programming solver | |
| CN111898734B (en) | NMR relaxation time inversion method based on MLP | |
| CN106777620A (en) | A kind of neutral net space reflection modeling method for power transistor | |
| Zhang et al. | Parallel decomposition approach to wide-range parametric modeling with applications to microwave filters | |
| Magerl et al. | Echo state networks for black-box modeling of integrated circuits | |
| Gomar et al. | Digital realization of PSTDP and TSTDP learning | |
| Yan et al. | Multi-physics parametric modeling of microwave passive components using artificial neural networks | |
| Liu et al. | Electromagnetic parametric modeling using combined neural networks and RLGC-based eigenfunctions for two-port microstrip structures | |
| Zhong et al. | An intelligent calculation method of volterra time-domain kernel based on time-delay artificial neural network | |
| CN104571086B (en) | Temperature controller emulation test method based on transmission function | |
| CN107957685B (en) | Neurodynamics method for solving noise-containing time-varying problem | |
| CN114692529B (en) | CFD high-dimensional response uncertainty quantification method and device, and computer equipment | |
| Cao et al. | Transient behavioral modeling of nonlinear I/O drivers combining neural networks and equivalent circuits | |
| Zhang | Parametric modeling and optimization of electromagnetic and multiphysics behaviors of microwave components | |
| CN111695230B (en) | Neural network space mapping multi-physical modeling method for microwave passive device | |
| Ersoy et al. | Realization of artificial neural networks on FPGA | |
| Khan et al. | Generalizable deep neural network based multi-material hysteresis modeling | |
| Wang et al. | Research on the Role of Hybrid Mesh Warm-up in Flow Prediction Based on Deep Learning | |
| CN110298073A (en) | The shift of integrated neural network and physical system model loads Dynamic Simulation Method | |
| Gubsky et al. | Computer modeling of measurement devices and tools |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190129 |
|
| WD01 | Invention patent application deemed withdrawn after publication |