CN108733852A - A kind of power amplifier behavior modeling method based on extreme learning machine - Google Patents
A kind of power amplifier behavior modeling method based on extreme learning machine Download PDFInfo
- Publication number
- CN108733852A CN108733852A CN201710247970.5A CN201710247970A CN108733852A CN 108733852 A CN108733852 A CN 108733852A CN 201710247970 A CN201710247970 A CN 201710247970A CN 108733852 A CN108733852 A CN 108733852A
- Authority
- CN
- China
- Prior art keywords
- model
- error
- power amplifier
- input
- learning machine
- 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]
- G06F30/30—Circuit design
- G06F30/36—Circuit design at the analogue level
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/30—Circuit design
- G06F30/36—Circuit design at the analogue level
- G06F30/367—Design verification, e.g. using simulation, simulation program with integrated circuit emphasis [SPICE], direct methods or relaxation methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/10—Numerical modelling
Landscapes
- Engineering & Computer Science (AREA)
- Computer Hardware Design (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Amplifiers (AREA)
Abstract
The theory of extreme learning machine is introduced into the practical application of amplifier behavior modeling by a kind of power amplifier behavior modeling method based on extreme learning machine, by determining mode input x j (t,t‑τ,…,t‑mτ) and target output t j (t) and model error ε, sample time intervalτ（Also referred to as memory interval）, model orderm；Determine the structure of model（Input layer, hidden layer, output layer）；Training parameter；Calculate error and etc. realize high-precision model（Least mean-square error reaches 10‑4Rank）；The modeling method has the advantages that low complex degree, low manual intervention, quick training pattern simultaneously.
Description
Technical field
The present invention relates to the fields such as power amplifier, behavior modeling, extreme learning machine, CAD, especially relate to
A kind of and power amplifier behavior modeling method based on extreme learning machine.
Background technology
Power amplifier（Power Amplifier, i.e. PA）It is one of key modules in wireless communication system, system performance
Quality be strongly dependent upon the performance of power amplifier.Power amplifier unintentional nonlinearity characteristic can lead to output signal matter
Amount declines, or even signal is made to fall flat, and influences normal communication [1].Therefore, in order to design high performance power amplification
The ancillary technique for improving power amplifier linearity, such as pre-distortion technology, negative-feedback technology, Power back can be used in device,
These ancillary techniques are required for accurately modeling the nonlinear characteristic of power amplifier, to describe the non-of power amplifier
Linear characteristic.
The modeling of power amplifier is divided into physical model and behavior model.Physical model needs to know circuit inner structure,
Understanding for the non-linear behavior characteristic of power amplifier requires no knowledge about circuit inner structure, therefore usage behavior models.Often
Behavior model is divided into following several：（1）Memoryless nonlinear model：What the output signal of system was only inputted with current time
Signal is related, such as Saleh models, power series model.（2）Memory nonlinear model：The output signal of system not only with currently
The input signal at moment is related, also related to the signal of the input of moment before, as Volterra models, Wiener models,
Hammerstein model, memory polynomial model, artificial neural network（Artificial Neural Networks ,ANN）
Model etc..
In PA modeling techniques, using relatively broad for memory polynomial model, Volterra series models and ANN
Model etc..Wherein memory polynomial model can be with the increase of circuit non-linearity characteristic with the complexity of Volterra series models
And increase, it is suitable for the modeling [2] [3] of small nonlinearity circuit.ANN model is a kind of model based on study, but is based on it
The problem of this body structure, can make the parameter for needing to adjust excessive, in addition, its calculating speed is also not best [4] [5].For
The above problem, invention introduces it is a kind of be all the model modelling approach based on learning algorithm --- extreme learning machine（Extreme
Learning Machine, i.e. ELM）[6].This method is applicable not only to small nonlinearity circuit modeling and is suitable for strong nonlinearity
Circuit modeling, meanwhile, under the premise of ensureing modeling accuracy, modeling speed is faster.
【Bibliography】
[1] S. C. Cripps, RF power amplifiers for wireless communications2nd ed.,
Artech House, Norwood, MA, 2006.
[2] E. Ngoya, and S. Mons, “Progress for behavioral challenges: a summary
of time-domain behavioral modeling of RF and microwave subsystems,” IEEE Microwave Mag., vol. 15, no. 6, pp. 91-105, Sep. 2014.
[3] J. C. Pedro, T. R. Cunha, “Predictable behavior: behavioral modeling
from measured data,” IEEE Microwave Mag., vol. 15, no. 6, pp. 75-90., Sep.
2014.
[4] S. X. Yan, C. Zhang, Q. J. Zhang, “Recurrent neural network technique
for behavioral modeling of power amplifier with memory effects”, Int. J. RF Microw. Comput.-Aided Eng., vol. 25, no. 4, pp. 289-298, May 2015.
[5] J. Xu, M. Yagoub, R. Ding, and Q. J. Zhang, “Neural-based dynamic
modeling of nonlinear microwave circuits,” IEEE Trans. Microw. Theory Tech.,
vol. 50, no.12, pp. 2769-2780, Dec. 2002.
[6] G. B. Huang, Q. Y. Zhu, and C. K. Siew, “Extreme learning machine: A
new learning scheme of feedforward neural networks,” in Proc. Int. Joint Conf. Neural Networks (IJCNN2004), Budapest, Hungary, Jul. 2004, vol. 2, pp.
985-990。
Invention content
For existing modeling method there are the problem of, a kind of power amplifier behavior based on extreme learning machine of the present invention is built
The theory of extreme learning machine is introduced into the practical application of amplifier behavior modeling by mould method, realizes that high-precision models（It is minimum
Mean square error reaches 10-4Rank）；Have the advantages that low complex degree, low manual intervention, quick training pattern simultaneously.
A kind of power amplifier behavior modeling method based on extreme learning machine, modeling procedure is as shown in Figure 1, including such as
Lower step：
Step (1)：Determine mode input x j (t, t - τ, … , t - mτ) and target output t j (t) and model error
ε,τFor sample time interval（Also referred to as memory interval）,mFor the exponent number of model；
Step (2)：Determine that the structure of model, model are divided into input layer, hidden layer and output layer；
Step (3)：Training parameter.By adjusting hidden layer neuron numberLCarry out training parameter with the type of excitation function g ()；
Step (4)：Calculate error.
In the step (1), power amplifier is seen as a flight data recorder by the present invention, determines the input x of behavior model j
(t, t - τ, … , t - mτ) be amplifier input signal vector, such as input power, frequency.Behavior model it is defeated
Incoming vector can be the vector of single type（If input signal vector is only input power or only frequency）Can also be multiple types
Type mixed vector（As existing input power also has frequency in input signal vector）.Target exports t j (t) it is power amplifier
Output signal vector, such as output power, gain, efficiency.It is determined that the input vector of behavior model determines the step
(2) input vector of input layer in, it is determined that the object vector of behavior model is determined in the step (4) and missed for calculating
The object vector of difference.
In the step (2), the present invention is in the structure of behavior model, input layer, hidden layer, the number of plies point of output layer
It Gui Ding not only one layer.ForNA random sample point (x j (t, t-τ, … ,t-mτ), t j (t))∈R d ×R m ,(j =
1, … ,N), haveLThe model structure of a hidden layer neuron as shown in Fig. 2, and model can be expressed as：
（1）
Wherein, x j (t, t-τ, … ,t-mτ) it is the mode input vector for being located at input layer in the step (1).o j (t)
For the output vector of model.b i (t) be each hidden layer neuron bias function, according to continuous distribution function（Just such as standard
State is distributed, is uniformly distributed, exponential distribution etc.）It is random to generate.w i (t) withβ i (t) it is respectively the weight for outputting and inputting vector.g
() is the nonlinear activation function of hidden layer, is used for the object function of fit non-linear（It is used for describing power amplifier non-thread
The function of property characteristic）, can be SIN function, S type functions, threshold function table etc., the type of excitation function is according to power amplifier
Nonlinear degree determines.For convenience of description, in description latert, t-τ, … ,t–mτIt will be ignored, such as x j (t, t-τ, … ,t-mτ)、o j (t)、t j (t)、w i (t)、β i (t) withb i (t), x will be respectively expressed as j 、o j 、t j 、w i 、β i Withb i .Model
The parameter training in step (3) that is determined as of structure is prepared.
In the step (3), the present invention greatly shortens the parameter training operating limit learning machine algorithm of behavior model
Model training time, but not loss model precision.The purpose of behavior modeling is to find one to can be very good description x j With t j It
Between relationship（I.e. to the description of non-linearity of power amplifier characteristic）Nonlinear function.Operating limit learning machine algorithm description x j With
t j Between relational approach it is as follows：o j For model output vector, t j For desired value（That is actual value）, to find suitable (w i , b i )
Withβ i So that
（2）
I.e.
（3）
Above formula can be written as T=Hβ.Training parameter can be equivalent to find least square solution：
（4）
Wherein, the number of hidden layer neuronLIt must not exceed the number of sample pointN.Parameter (the w of hidden layer neuron i , b i )
(i = 1, …, L) it is to randomly generate.Adjust hidden layer neuron numberLSize and excitation function g () type, and count
Calculate formula (1) so that o j Value close to t j , calculation formula (4) obtainsβ i Value.The training of parameter is for confirming row
For the parameter (w of model i , b i ) withβ i 。
In the step (4), according to the parameter that the step (3) trains, o is calculated j With t j Between error, obtain mould
Type error.The calculation of error can be relative error, absolute error etc., for different error calculation modes, determine
'sεIt will be different.Error is calculated for confirming whether established model meets target, i.e., is set up in the described step (1)
Error criterion.If being unsatisfactory for error criterion, return to step (3) re -training parameter is needed；If meeting error criterion, illustrate model
Meet target, modeling is completed.
A kind of power amplifier behavior modeling method based on extreme learning machine, compared with prior art, present invention modeling
The advantages of method is：
(1) according to the variability of excitation function g (), method of the invention is not only applicable to small nonlinearity circuit, is also applied for
Strong nonlinearity circuit；
(2) modeling process of the invention only needs to adjust hidden layer neuron numberLWith the type of excitation function g (), adjust
Parameter is few, and modeling speed is fast, and modeling method is easy to use.
Description of the drawings
Fig. 1 is the flow chart of the modeling method specific implementation mode of the present invention；
Fig. 2 is modeling method principle assumption diagram of the present invention；
Fig. 3 is the modeling result comparison diagram of the present invention and a kind of nonlinear characteristic of an E classes PA of ANN methods pair；
Fig. 4 is the present invention and comparison of the ANN methods for input power and gain modeling result.
Specific implementation mode
Illustrate the preferred forms of the present invention below according to attached drawing：
Step (1)：In present embodiment, for the E class PA of a strong nonlinearity, the Advanced of Agilent companies is used
Design System (ADS) software is emulated.Scan power variation range is from 10 dBm to 30 dBm and step-length is 1
DBm, frequency range is from 2.80 GHz to 3.00 GHz and step-length is 0.05 GHz, and the sample point of emulation is obtained with this
(x j (t, t - τ, … , t - mτ), t j (t)) ∈R d × R m ,(j = 1, … , N).Wherein, x j (t, t -τ, … , t - mτ) be power amplifier input power and frequency vector；t j (t) be power amplifier gain to
Amount；τFor sample time interval, it is chosen as 0.001；m1 is selected as this preferred forms for the exponent number of model.Mould
Type error may be configured asε=10-3。
Step (2)：Input layer, hidden layer, the output layer of behavior model structure are disposed as one layer.For in the present embodiment
Non-linearity of power amplifier characteristic, in formula (1),b i (t) select submit between (0,1) standard uniformly continuous distribution
Random number, w i (t) random number for submitting to the distribution of the standard uniformly continuous between (0,2) is selected,Select SIN function.
Step (3)：According to (the w generated at random i , b i ) calculateβ i So that
（5）
Wherein, the number of hidden layer neuronLIt must not exceed the number of sample pointN, and the precision of model can be withLIncrease and
It improves.It adjustsLSize and g () type, calculation formula (5) obtainsValue.
Step (4)：According to (the w obtained in step (2) and step (3) i , b i ) withβ i , using formula (6), calculate o j With t j
Between least mean-square error（Mean Square Error, MSE）：
（6）
Work as MSE<εWhen, show that established function model can describe x well j With t j Between relationship, i.e., modeling reach fine
Effect；Work as MSE>εWhen, then need return to step (3) re -training parameter.The present embodiment existsLBe selected as 10 andIt is selected as just
Error reaches 10 when string function-4Rank meets MSE<ε.The method for calculating error is not limited solely to use least mean-square error,
Can be other error formulas.
To show the effect of the present invention, an E class PA with strong nonlinearity characteristic is selected, uses ELM methods and ANN respectively
Nonlinear characteristic of the method between the output power of PA, frequency and gain models, and modeling result is as shown in Figure 3 and Figure 4.Its
In, Fig. 3 is the design sketch that ELM and ANN emulates ADS data modeling；Fig. 4 is that ELM and ANN emulates data, modeling for ADS
The comparison of required training time and modeling result precision.By the above results figure it is found that the model that the present embodiment is built has very
High precision, and it is apparent less the time required to model training.And then it is a kind of to power amplification to prove that the method for the present invention can provide
The method that device carries out behavior modeling, this method modeling speed is fast and institute's established model precision is high.
The above, best specific implementation mode only of the invention, but scope of protection of the present invention is not limited thereto,
Any one skilled in the art in the technical scope disclosed by the present invention, the change or replacement that can be readily occurred in,
It should be covered by the protection scope of the present invention.
Claims (5)
1. a kind of power amplifier behavior modeling method based on extreme learning machine, it is characterised in that：Include the following steps：
Step (1)：Determine mode input x j (t, t - τ, … , t - mτ) and target output t j (t) and model error
ε,τFor sample time interval（Also referred to as memory interval）,mFor the exponent number of model；
Step (2)：Determine that the structure of model, model are divided into input layer, hidden layer and output layer；
Step (3)：Training parameter；
By adjusting hidden layer neuron numberLCarry out training parameter with the type of excitation function g ()；
Step (4)：Calculate error.
2. a kind of power amplifier behavior modeling method based on extreme learning machine according to claim 1, it is characterised in that：
In the step (1), power amplifier is seen as a flight data recorder, determines the input x of behavior model j (t, t - τ, … ,t - mτ) be amplifier input signal vector, such as input power, frequency, and pass through sample time intervalτ（Also referred to as remember
Recall interval）With model ordermTo embody the Memorability of signal；The input vector of behavior model can be that the vector of single type also may be used
To be the vector of multiple types mixing；Target exports t j (t) vectorial for the output signal of power amplifier, such as output power increases
Benefit, efficiency etc.；It is determined that the input vector of behavior model determines the input vector of input layer in the step (2), determines
The object vector of behavior model is the object vector determined in the step (4) for calculating error.
3. a kind of power amplifier behavior modeling method based on extreme learning machine according to claim 1, it is characterised in that：
In the step (2), the present invention is in the structure of behavior model, input layer, hidden layer, and the number of plies of output layer respectively provides only
There is one layer.
4. a kind of power amplifier behavior modeling method based on extreme learning machine according to claim 1, it is characterised in that：
In the step (3), the present invention substantially reduces model instruction for the parameter training operating limit learning machine algorithm of behavior model
Practice the time, but not loss model precision.
5. a kind of power amplifier behavior modeling method based on extreme learning machine according to claim 1, it is characterised in that：
In the step (4), according to the parameter that the step (3) trains, o is calculated j With t j Between error, obtain model error；Accidentally
The calculation of difference can be relative error, absolute error etc., identified for different error calculation modesεIt will be different；
Error is calculated for confirming whether established model meets target, i.e., the error criterion set up in the described step (1)；If
It is unsatisfactory for error criterion, needs return to step (3) re -training parameter；If meeting error criterion, illustrate that model meets expected mesh
Mark, modeling are completed.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710247970.5A CN108733852A (en) | 2017-04-17 | 2017-04-17 | A kind of power amplifier behavior modeling method based on extreme learning machine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710247970.5A CN108733852A (en) | 2017-04-17 | 2017-04-17 | A kind of power amplifier behavior modeling method based on extreme learning machine |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108733852A true CN108733852A (en) | 2018-11-02 |
Family
ID=63923883
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710247970.5A Pending CN108733852A (en) | 2017-04-17 | 2017-04-17 | A kind of power amplifier behavior modeling method based on extreme learning machine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108733852A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109858616A (en) * | 2019-02-15 | 2019-06-07 | 东南大学 | Power amplifier behavior level modeling system neural network based and method |
| CN113221308A (en) * | 2021-06-11 | 2021-08-06 | 北京邮电大学 | Transfer learning rapid low-complexity modeling method facing power amplifier |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6047168A (en) * | 1996-06-28 | 2000-04-04 | Telefonaktiebolaget Lm Ericsson | Device and method for radio transmitters |
| CN103728431A (en) * | 2014-01-09 | 2014-04-16 | 重庆科技学院 | Industrial sewage COD (chemical oxygen demand) online soft measurement method based on ELM (extreme learning machine) |
| CN105224985A (en) * | 2015-09-28 | 2016-01-06 | 南京航空航天大学 | A kind of power amplifier behavior modeling method based on degree of depth reconstruction model |
-
2017
- 2017-04-17 CN CN201710247970.5A patent/CN108733852A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6047168A (en) * | 1996-06-28 | 2000-04-04 | Telefonaktiebolaget Lm Ericsson | Device and method for radio transmitters |
| CN103728431A (en) * | 2014-01-09 | 2014-04-16 | 重庆科技学院 | Industrial sewage COD (chemical oxygen demand) online soft measurement method based on ELM (extreme learning machine) |
| CN105224985A (en) * | 2015-09-28 | 2016-01-06 | 南京航空航天大学 | A kind of power amplifier behavior modeling method based on degree of depth reconstruction model |
Non-Patent Citations (2)
| Title |
|---|
| 张呈宇等: "Extreme Learning Machine for the Behavioral Modeling of RF", 《 2017 IEEE MTT-S INTERNATIONAL MICROWAVE SYMPOSIUM (IMS)》 * |
| 柯海森: "一种改进极限学习机方法的研究", 《第三十二届中国控制会议》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109858616A (en) * | 2019-02-15 | 2019-06-07 | 东南大学 | Power amplifier behavior level modeling system neural network based and method |
| CN113221308A (en) * | 2021-06-11 | 2021-08-06 | 北京邮电大学 | Transfer learning rapid low-complexity modeling method facing power amplifier |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Mkadem et al. | Physically inspired neural network model for RF power amplifier behavioral modeling and digital predistortion | |
| He et al. | Adaptive neural network control of unknown nonlinear affine systems with input deadzone and output constraint | |
| Abdelhafiz et al. | A PSO based memory polynomial predistorter with embedded dimension estimation | |
| Yan et al. | Recurrent neural network technique for behavioral modeling of power amplifier with memory effects | |
| CN108733852A (en) | A kind of power amplifier behavior modeling method based on extreme learning machine | |
| You et al. | An optimized-load-impedance calculation and mining method based on I–V curves: Using broadband class-E power amplifier as example | |
| CN102983819A (en) | Imitating method of power amplifier and imitating device of power amplifier | |
| CN103296976B (en) | The method, apparatus and magnetic resonance equipment of the quiescent point of stable radio frequency amplifier | |
| Yu et al. | A nonlinear behavioral modeling approach for voltage-controlled oscillators using augmented neural networks | |
| Tran et al. | Ringing test for third-order ladder low-pass filters | |
| Feng et al. | Digital predistortion method combining memory polynomial and feed‐forward neural network | |
| CN110705085B (en) | Control method and system for high-frequency digital low level in single sine mode of accelerator | |
| CN102081751A (en) | Method for modeling synchronous double-frequency power amplifier based on real number time delay neural network | |
| Aguilar‐Lobo et al. | Application of the NARX neural network as a digital predistortion technique for linearizing microwave power amplifiers | |
| Zhou et al. | A nonlinear memory power amplifier behavior modeling and identification based on memory polynomial model in soft-defined shortwave transmitter | |
| Semyonov | Synthesis of behavioral models for circuits with nonlinearity less than model error | |
| Gong et al. | A Memorized Recurrent Neural Network Design for Wide Bandwidth PA Linearization | |
| Devi et al. | Broadband RF power amplifier modeling using an enhanced Wiener model | |
| Song et al. | A Novel Digital Predistortion Identification Algorithm Based on Variable Forgetting Factor Recursive Least Square Method | |
| CN108268700A (en) | A kind of RF power amplification temperature characterisitic modeling method based on BPNN | |
| CN114598578A (en) | Digital predistortion training method and device and storage medium | |
| Ming et al. | Modeling Broad-band Power Amplifier Using Hybrid Tinte-Delay Neural Network | |
| Nuñez-Perez et al. | FPGA realization of RF-PA models with memory effects based on ANFIS | |
| Wang et al. | A Novel Transfer Learning Approach for Efficient RF Device Behavior Model Parameter Extraction | |
| CN110598261A (en) | Power amplifier frequency domain modeling method based on complex reverse neural network |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20181102 |