CN116245064A - Power amplifier predistortion processing method based on simplified GMP variant model - Google Patents

Power amplifier predistortion processing method based on simplified GMP variant model Download PDF

Info

Publication number
CN116245064A
CN116245064A CN202310110118.9A CN202310110118A CN116245064A CN 116245064 A CN116245064 A CN 116245064A CN 202310110118 A CN202310110118 A CN 202310110118A CN 116245064 A CN116245064 A CN 116245064A
Authority
CN
China
Prior art keywords
model
power amplifier
gmp
variant
coefficient
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
Application number
CN202310110118.9A
Other languages
Chinese (zh)
Inventor
严家树
郭春炳
杜志侠
林鑫
肖亦成
陈思宇
邱鼎
孔祥键
简明朝
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Guangdong University of Technology
Original Assignee
Guangdong University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Guangdong University of Technology filed Critical Guangdong University of Technology
Priority to CN202310110118.9A priority Critical patent/CN116245064A/en
Publication of CN116245064A publication Critical patent/CN116245064A/en
Pending legal-status Critical Current

Links

Images

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
    • 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
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H03—ELECTRONIC CIRCUITRY
    • H03F—AMPLIFIERS
    • H03F1/00—Details of amplifiers with only discharge tubes, only semiconductor devices or only unspecified devices as amplifying elements
    • H03F1/32—Modifications of amplifiers to reduce non-linear distortion
    • H03F1/3241—Modifications of amplifiers to reduce non-linear distortion using predistortion circuits
    • H—ELECTRICITY
    • H03—ELECTRONIC CIRCUITRY
    • H03F—AMPLIFIERS
    • H03F3/00—Amplifiers with only discharge tubes or only semiconductor devices as amplifying elements
    • H03F3/20—Power amplifiers, e.g. Class B amplifiers, Class C amplifiers
    • H03F3/21—Power amplifiers, e.g. Class B amplifiers, Class C amplifiers with semiconductor devices only
    • H03F3/213—Power amplifiers, e.g. Class B amplifiers, Class C amplifiers with semiconductor devices only in integrated circuits
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Hardware Design (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Microelectronics & Electronic Packaging (AREA)
  • Artificial Intelligence (AREA)
  • Geometry (AREA)
  • Power Engineering (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Medical Informatics (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Nonlinear Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Amplifiers (AREA)

Abstract

The invention discloses a power amplifier predistortion processing method based on a simplified GMP variant model, which comprises the following steps: s1: building a neural network model for selecting the basis functions, and selecting the basis functions; s2: constructing a simplified GMP variant model based on the selected basis function as a power amplifier distortion model; s3: solving the power amplifier distortion model coefficient to obtain a global optimal solution of the power amplifier distortion model coefficient; s4: according to the obtained global optimal solution of the power amplifier distortion model coefficient, the baseband signal is processedThe signal u is pre-distorted, the input signal of the pre-distortion is the baseband signal u, the baseband signal u is pre-distorted to obtain the pre-distorted signal x as the input of the power amplifier, the pre-distorted signal x is processed by the power amplifier to obtain the output signal after linear amplification
Figure DDA0004076515920000011
The invention reduces the number of coefficients used by the model, reduces the calculation complexity and improves the self-adaptive capacity of the model.

Description

Power amplifier predistortion processing method based on simplified GMP variant model
Technical Field
The invention relates to the technical field of wireless communication, in particular to a power amplifier predistortion processing method based on a simplified GMP variant model.
Background
Wireless communication is increasingly developed, and for a power amplifier module, many input signals have the characteristics of high peak-to-average ratio, broadband and high-order modulation at present, so that the power amplifier can enter a nonlinear region more easily, and a serious memory effect is generated by the power amplifier, and the gain effect of the power amplifier is further influenced. Aiming at the technology of power amplification, the digital predistortion method has the advantages of simplicity, high model precision, good stability, self-adaptability and the like, and has been widely used.
The core of the digital predistortion method is to fit various distortion characteristics of the power amplifier. Because of the memory effect and nonlinear change of the power amplifier, under the permission of the implementation condition, the memory term and Gao Jiexiang should be added to perform high-performance fitting on the distortion model, wherein the generalized memory polynomial GMP model with the memory lead and lag cross term and the higher-order term is added, and is widely accepted. In recent years, neural networks have also become of increasing interest as an emerging technique for predistortion. The NN model of the neural network can be theoretically fit with any nonlinear model, can perform parameter calculation and updating through undetermined coefficients and fixed input and output and iterative operation, and has strong learning and self-adaption capability.
In practice, models with high fitting performance, such as a Vortla series VS model and an NN model, such as a GMP model, often have the defects of numerous coefficients and complex structure. In order to increase the application of the method in various scenes and improve the defects, a plurality of self-adaptive methods are proposed, and the main aspects are coefficient self-adaptation and structure self-adaptation at present. The former means that the coefficients can be iteratively updated until the model is within allowable fitting errors; the method also comprises self-adaptive modification of parameters in the convergence process, such as parameters of convergence step length and the like, so that fitting precision is improved. The latter means that the memory depth and the polynomial order can be adaptively adjusted to change the model structure; or according to the input and output signals, the model becomes sparse, so that a plurality of coefficients do not need to be calculated; or the weight parameter calculation with smaller importance is reduced, and finally the complexity of the model is reduced.
In the future, coefficients can be effectively reduced under the condition of keeping fitting accuracy, so that the hardware structure is more simplified, the realization cost is lower, the model has self-adaptive characteristics, and the method is a main predistortion improvement direction.
The method has the defects that the coefficient simplifying scheme is low in flexibility, the coefficient quantity is large due to the fact that a GMP model is easy to fit to nonlinearity, the neural network model is limited by data quantity, training duration and complexity, when the power amplifier performance changes, the whole structure is easy to lose efficacy or the change cost is large, and the self-adaption capability is insufficient.
Disclosure of Invention
The invention provides a power amplifier predistortion processing method based on a simplified GMP variant model, which aims to solve the problems of large number of model coefficients and high calculation complexity in the prior art, reduces the use of redundancy coefficients and reduces the calculation complexity.
In order to solve the technical problems, the invention adopts the following technical scheme:
a power amplifier predistortion processing method based on a simplified GMP variant model comprises the following steps:
s1: building a neural network model for selecting the basis functions, and selecting the basis functions;
s2: constructing a simplified GMP variant model based on the selected basis function as a power amplifier distortion model;
s3: solving the power amplifier distortion model coefficient to obtain a global optimal solution of the power amplifier distortion model coefficient;
s4: according to the obtained global optimal solution of the power amplifier distortion model coefficient, carrying out predistortion processing on a baseband signal u, wherein an input signal of the predistortion processing is the baseband signal u, the baseband signal u is subjected to predistortion processing to obtain a predistortion signal x which is used as an input of a power amplifier, and the predistortion signal x is subjected to power amplifier processing to obtain a linear amplified output signal
Figure BDA0004076515900000021
The working principle of the invention is as follows:
and (3) finishing the selection of the basis functions through the constructed neural network model, constructing a simplified GMP variant model based on the selected basis functions as a power amplifier distortion model, solving the global optimal solution of the power amplifier distortion model coefficient, obtaining the input of an actual power amplifier by using the power amplifier distortion model, obtaining the output signal after linear amplification, and adapting the power amplifier distortion model to different input or power amplifier change conditions by changing the basis functions so as to realize the adaptive iteration of the power amplifier distortion model.
Preferably, the neural network model uses a 3-CNN neural network model; the 3-CNN neural network model comprises a 1-layer convolution layer, a 1-layer pooling layer and a 1-layer full-connection layer.
Further, the process of constructing the neural network model is as follows:
training the neural network model by using the training set, and performing model evaluation on the neural network model by using the testing set; the training set and the testing set come from input and output signals with different time lengths generated by different power amplifiers at different moments.
Acquiring characteristics through a power amplifier input and output sample set; the feature vector is obtained through calculation, and the label corresponding to the feature vector is the kind of the basis function.
And placing the features of the training set and the corresponding labels into a neural network model for model training, placing the features of the test set into the neural network model for model evaluation, comparing the result of the selection of the basis function of the neural network model with the correct labels to calculate a model fitting error, and adjusting the structural parameters of the neural network model to ensure that the model fitting error is smaller than a fitting error threshold.
Furthermore, the characteristic is time-frequency characteristic of the sequence, and let the data points of the baseband signal and the original power amplifier output signal be N, the baseband signal be u (N), n=1, N, the original power amplifier output signal be y (N), n=1, N.
If p time-frequency characteristics calculated according to a section of sequence are provided, calculating a 2 p-long characteristic vector according to a group of u and y; g groups of signals are arranged, and a feature matrix with the size of 2p x q is calculated; for any sequence s, the time-frequency characteristic and the calculation formula representing the arbitrary sequence s are as follows:
the time domain features are s_mean, s_std, s_max, s_min.
s_mean is the average of the vector s modulus, expressed as:
Figure BDA0004076515900000031
s_std is the variance of the vector s modulus, expressed as:
Figure BDA0004076515900000032
s_max is the maximum value of the vector s modulus, and the expression is s_max=max (|s|).
s_min is the minimum of the vector s modulus, and the expression is s_min=min (|s|).
The frequency domain features are p_max and f_pmax; p_max is the maximum value of the power amplitude by
Figure BDA0004076515900000033
-pi < omega < pi to obtain a power spectrum, wherein j is an imaginary number and omega is a frequency; the power amplitude maximum is calculated from the power spectrum.
f_pmax is the frequency value corresponding to the maximum value of the power amplitude in the power spectrum.
Further, the types of the basis functions comprise a memory polynomial fitting basis function, a Fourier polynomial fitting basis function and an Euler polynomial fitting basis function; the basis function representation and the corresponding fitted polynomial are as follows:
the expression of the memory polynomial fitting basis function is:
u(n-l)|u(n-l)| k
where n is the data point number of u, l is the memory depth, and k is the order.
The polynomial fit of the memory polynomial fit basis function is:
Figure BDA0004076515900000034
wherein a is k，l u is the coefficient of the basis function corresponding to different memory depths and orders.
The fourier polynomial fit basis function is expressed as:
coS[u(n-l)]|cos[u(n-l)]| k
the fourier polynomial fit of the fourier polynomial fit basis function is:
Figure BDA0004076515900000041
the expression of the Euler polynomial fitting basis function is:
Figure BDA0004076515900000042
Figure BDA0004076515900000043
Figure BDA0004076515900000044
wherein A is n-l For the amplitude coefficient of the basic function, theta, under different memory depths n-l Is the frequency coefficient under different memory depth.
The euler polynomial fit of the euler polynomial fit basis function is:
Figure BDA0004076515900000045
where K is the order and L is the memory depth.
The baseband signal u corresponding to the characteristic vector and the original power amplifier output signal y are fitted by using polynomials of different basis functions, the memory depth and the order are set, the corresponding coefficient is calculated by using a least square method, and the fitting result is that
Figure BDA0004076515900000046
Error of fit error passing
Figure BDA0004076515900000047
And (5) calculating.
Setting a fitting error threshold, and if the fitting error is not smaller than the fitting error threshold, increasing the memory depth or the order until the fitting error is smaller than the threshold; the coefficients used by the polynomials of the different basis functions are compared, the basis function with the least coefficients used being the label that best fits the non-linear characteristics of the power amplifier model, the kind of basis function being marked as the eigenvector.
Further, the steps of constructing the simplified GMP variant model are as follows:
the traditional GMP model is as follows:
Figure BDA0004076515900000048
where u is the baseband signal, y GMP To output a signal a k，l 、b k，l，m C k，l，m For coefficients of different terms, L a 、L b 、L c To memorize depth, K a 、K b 、K c For the order, M b 、M c For lead or lag length; calculating the number of coefficients of the traditional GMP model to be K a L a +K b L b M b +K c L c M c A plurality of; bringing the original power amplifier output signal y into y GMP Then, coefficient calculation is carried out;
when the memory polynomial fitting basis function and the fourier polynomial fitting basis function are selected separately, the conventional GMP model performs the following variants:
y GMP1 (n)＝y GMP (n)
Figure BDA0004076515900000051
Figure BDA0004076515900000052
let f (n, l, k) =f 1 (n，l)·|f 2 (n，l)| k As a basis function, the variant GMP model and corresponding values are as follows:
when f 1 (n, l) is u (n-l), f 2 When (n, l) is u (n-l), the corresponding GMP variant model is expressed as y GMP1 ；
When f 1 (n, l) is cos [ u (n-l)]，f 2 (n, l) is cos [ u (n-l)]When the corresponding GMP variant model is denoted as y GMP2 ；
When f 1 (n, l) is A n-l ，f 2 (n, l) is
Figure BDA0004076515900000053
When the corresponding GMP variant model is denoted as y GMP3 ；
And (3) carrying out coefficient simplification on the GMP variant model, and carrying out simplification on summation items of the GMP variant model after coefficient simplification, wherein the matrix form of the GMP variant model after simplification is expressed as follows:
Figure BDA0004076515900000054
where u' is a memory depth L based on the setting from the baseband signal data point u (n) a Advance length M b Length of hysteresis M c The calculated vector has a length of 2 (L a +L b M b +L c M c ) The method comprises the steps of carrying out a first treatment on the surface of the w' is the corresponding y GMP′ A coefficient vector of (n).
Further, the steps for coefficient reduction of the GMP variant model are as follows:
when present
Figure BDA0004076515900000061
In the form, the definition domain is used as arbitrary value
Figure BDA0004076515900000062
Figure BDA0004076515900000063
/>
The two taylor formulas perform coefficient conversion, and the expression is:
Figure BDA0004076515900000064
when z l ＝|f 2 (n, l) is represented by the formula
Figure BDA0004076515900000065
If it is
Figure BDA0004076515900000066
q is an even number from 0 to K-1, and
Figure BDA0004076515900000067
q is an odd number from 0 to K-1;
Figure BDA0004076515900000068
further, the step of simplifying each summation term of the GMP variant model after coefficient simplification is as follows:
Figure BDA0004076515900000071
Figure BDA0004076515900000072
/>
Figure BDA0004076515900000073
the above expression is written in the following matrix form in turn:
Figure BDA0004076515900000074
Figure BDA0004076515900000075
Figure BDA0004076515900000076
Figure BDA0004076515900000081
Figure BDA0004076515900000082
Figure BDA0004076515900000083
finally, a matrix form of the simplified GMP variant model is obtained.
Further, the step of solving the global optimal solution of the power amplifier distortion model coefficient is as follows:
according to the simplified GMP variant model, calculating a corresponding predistortion coefficient by utilizing the input and output of the power amplifier; that is, u' is the memory depth L according to the setting by the distortion output y (n) of the actual power amplifier a Advance length M b Length of hysteresis M c W ' is a coefficient vector corresponding to u (n), i.e., u (n) =u ' ·w ';
all u (N) use the same set of coefficients w', if the data points of u (N) and y (N) are N
u＝Y·ω′，u∈R N×1 ，
Figure BDA0004076515900000084
Calculating a global optimal solution using a least squares method: w' = (Y) H Y) -1 Y H u。
The parameter is set to L a ＝L b ＝L c ＝3，M b ＝M c =2 for adapting to the power amplifier distortion situation.
The lookup table is used for assisting the operation of the sine base function or the cosine base function in the solving process, and the method comprises the following steps:
all results of the logic circuit are calculated, a table with x as an index value is manufactured, the table is written into the RAM in advance, and the corresponding value in the table is selected as a result to be output during calculation.
Further, the input baseband signal u is calculated by:
x＝U·w′，x∈R N×1 ，
Figure BDA0004076515900000085
obtaining a predistortion signal x as an input of an actual power amplifier, and finallyObtaining a linearly amplified output signal
Figure BDA0004076515900000086
And (4) carrying out self-adaptive iteration on the power amplifier distortion model, and repeating the steps S1 to S4 to finish self-adaptive adjustment.
Compared with the prior art, the invention has the beneficial effects that:
1. by constructing a neural network model, the optimal basis function selection is performed, and a simplified GMP variant model is constructed, so that the use of redundancy coefficients is reduced, the calculation complexity is reduced, and the calculation speed of the model is increased.
2. And the operation is assisted by a lookup table method, so that the flexibility of a coefficient simplification scheme is improved.
3. And the power amplifier distortion model is adaptively adjusted according to the power amplifier change condition, so that the adaptive capacity of the power amplifier distortion model is improved.
4. The training time cost before use is reduced through the 3-CNN neural network model, the actual power amplifier distortion relation is quickly approximated, and the cost is reduced.
Drawings
Fig. 1 is a flow chart of a power amplifier predistortion processing method based on a simplified GMP variant model.
Fig. 2 is a flow diagram of an embodiment of the present invention.
Detailed Description
The invention is described in detail below with reference to the drawings and the detailed description.
Example 1
In this embodiment, as shown in fig. 1 and 2, a power amplifier predistortion processing method based on a simplified GMP variant model includes the following steps:
s1: and building a neural network model for selecting the basis functions, and selecting the basis functions.
S2: and constructing a simplified GMP variant model based on the selected basis function as a power amplifier distortion model.
S3: and solving the power amplifier distortion model coefficient to obtain a global optimal solution of the power amplifier distortion model coefficient.
S4: according to the obtained power amplifierThe global optimal solution of the distortion model coefficient carries out predistortion processing on a baseband signal u, an input signal of the predistortion processing is the baseband signal u, the baseband signal u is subjected to predistortion processing to obtain a predistortion signal x which is used as the input of a power amplifier, and the predistortion signal x is subjected to power amplifier processing to obtain an output signal of which the linear amplification is carried out
Figure BDA0004076515900000091
The working principle of the invention is as follows:
and (3) finishing the selection of the basis functions through the constructed neural network model, constructing a simplified GMP variant model based on the selected basis functions as a power amplifier distortion model, solving the global optimal solution of the power amplifier distortion model coefficient, obtaining the input of an actual power amplifier by using the power amplifier distortion model, obtaining the output signal after linear amplification, and adapting the power amplifier distortion model to different input or power amplifier change conditions by changing the basis functions so as to realize the adaptive iteration of the power amplifier distortion model.
In this embodiment, the 3-CNN neural network model includes a 1-layer convolutional layer, a 1-layer pooling layer, and a 1-layer full-connection layer.
More specifically, the process of constructing the neural network model is as follows:
training the neural network model by using the training set, and performing model evaluation on the neural network model by using the testing set; the training set and the testing set are from input and output signals with different time lengths generated by different power amplifiers at different moments; the ratio of the duty ratio of the training set to the test set is 4:1, a step of;
acquiring characteristics through a power amplifier input and output sample set; the feature obtains a feature vector through calculation, and a label corresponding to the feature vector is a kind of a basic function;
and placing the features of the training set and the corresponding labels into a neural network model for model training, placing the features of the test set into the neural network model for model evaluation, comparing the result of the selection of the basis function of the neural network model with the correct labels to calculate a model fitting error, and adjusting the structural parameters of the neural network model to ensure that the model fitting error is smaller than a fitting error threshold.
More specifically, the characteristic is time-frequency characteristic of the sequence, and let the baseband signal and the original power amplifier output signal data point be N, the baseband signal be u (N), =1,, the original power amplifier output signal be y (N), =1,,.
If p time-frequency characteristics calculated according to a section of sequence are provided, calculating a 2 p-long characteristic vector according to a group of u and y; q groups of signals are arranged, and a feature matrix with the size of 2p x q is calculated; for an arbitrary sequence s, the time-frequency characteristics representing the arbitrary sequence s are calculated as shown in table 1;
Figure BDA0004076515900000101
TABLE 1
The label corresponding to the feature vector is the kind of the basis function; the types of the basis functions comprise a memory polynomial fitting basis function, a Fourier polynomial fitting basis function and an Euler polynomial fitting basis function; the basis function representation and the corresponding fitted polynomial are shown in table 2:
Figure BDA0004076515900000102
/>
Figure BDA0004076515900000111
TABLE 2
A in Table A k,l The basic function coefficients corresponding to different memory depths and orders; a is that n-l For the amplitude coefficient of the basic function, theta, under different memory depths n-l Is the frequency coefficient under different memory depth.
The baseband signal u corresponding to the characteristic vector and the original power amplifier output signal y are fitted by using polynomials of different basis functions, the memory depth and the order are set, the corresponding coefficient is calculated by using a least square method, and the fitting result is that
Figure BDA0004076515900000113
Error of fit error passing
Figure BDA0004076515900000112
And (5) calculating.
Setting a fitting error threshold, and if the fitting error is not smaller than the fitting error threshold, increasing the memory depth or the order until the fitting error is smaller than the threshold; the coefficients used by the polynomials of the different basis functions are compared, the basis function with the least coefficients used being the label that best fits the non-linear characteristics of the power amplifier model, the kind of basis function being marked as the eigenvector. For the basis function selection, the polynomial fitting model does not need to select a traditional GMP model, namely only a high-order term and a simple memory term are needed, after the basis function matched with the current power amplification characteristic is selected, the nonlinear fitting effect of the model in the GMP form can be further enhanced, and the nonlinear problem under the broadband condition can be solved.
More specifically, a general three-layer convolutional neural network model 3-CNN is selected, a feature vector is used as the input of the neural network model, and an adaptive basis function is selected for modeling of a power amplifier distortion model.
More specifically, the steps for constructing the simplified GMP variant model are as follows:
the traditional GMP model is as follows:
Figure BDA0004076515900000121
where u is the baseband signal, y GMP To output a signal a k，l 、b k，l，m 、c k，l，m For coefficients of different terms, L a 、L b 、L c To memorize depth, K a 、K b 、K c For the order, M b 、M c For lead or lag length; thus the number of model coefficients can be calculated as K a L a +K b L b M b +K c L c M c A plurality of; in actual calculation, the original power amplifier output y is directly substituted into y GMP Then, coefficient calculation is carried out;
when the memory polynomial fitting basis function and the fourier polynomial fitting basis function are selected separately, the conventional GMP model performs the following variants:
y GMP1 (n)＝y GMP (n)
Figure BDA0004076515900000122
Figure BDA0004076515900000123
let f (n, l, k) =f 1 (n，l)·f 2 (n, l, k) is the basis function, the variant GMP model and corresponding values are shown in Table 3:
Figure BDA0004076515900000131
TABLE 3 Table 3
And (3) carrying out coefficient simplification on the GMP variant model, and carrying out simplification on summation items of the GMP variant model after coefficient simplification, wherein the matrix form of the GMP variant model after simplification is expressed as follows:
Figure BDA0004076515900000132
where u' is a memory depth L based on the setting from the baseband signal data point u (n) a Advance length M b Length of hysteresis M c The calculated vector has a length of 2 (L a +L b M b +L c M c ) The method comprises the steps of carrying out a first treatment on the surface of the w' is the corresponding y GMP′ A coefficient vector of (n).
More specifically, because the GMP model is essentially a Taylor series generalized VS model, the main calculation module of the GMP variant formula can use the Taylor formula to carry out auxiliary calculation, effectively expand the higher-order terms to the infinite-order terms, and simultaneously carry out coefficient combination of the same memory depth, so that hardware implementation difficulty caused by a large amount of redundancy is avoided; the steps for coefficient reduction of the GMP variant model are as follows:
when present
Figure BDA0004076515900000133
In the form, the definition domain is used as arbitrary value
Figure BDA0004076515900000134
Figure BDA0004076515900000141
The two taylor formulas perform coefficient conversion, and the expression is:
Figure BDA0004076515900000142
when z l ＝|f 2 (n, l) is represented by the formula
Figure BDA0004076515900000143
If it is
Figure BDA0004076515900000144
q is an even number from 0 to K-1, and
Figure BDA0004076515900000145
q is an odd number from 0 to K-1;
Figure BDA0004076515900000146
the above formula is an expression of coefficient simplification of the GMP variant model, and the above coefficient merging process, called "coefficient simplification" process, is based on the condition that the basis function is properly matched and fitted to infinity, and the fitting error caused by the coefficient simplification is acceptable, that is, the simplification result of the module can be widely applied to the GMP variant model.
More specifically, based on the simplified expression of the coefficient of the GMP variant model, the steps of simplifying each summation term of the GMP variant model after the coefficient is simplified are as follows:
Figure BDA0004076515900000151
Figure BDA0004076515900000152
/>
Figure BDA0004076515900000153
the above expression is written in the following matrix form in turn:
Figure BDA0004076515900000154
Figure BDA0004076515900000155
Figure BDA0004076515900000156
Figure BDA0004076515900000157
Figure BDA0004076515900000161
Figure BDA0004076515900000162
Figure BDA0004076515900000163
finally, a matrix form of the simplified GMP variant model is obtained.
More specifically, the step of solving the global optimal solution of the power amplifier distortion model coefficient is as follows:
according to the simplified GMP variant model, calculating a corresponding predistortion coefficient by utilizing the input and output of the power amplifier; that is, u' is the memory depth L according to the setting by the distortion output y (n) of the actual power amplifier a Advance length M b Length of hysteresis M c W ' is a coefficient vector corresponding to u (n), i.e., u (n) =u ' ·w '.
All u (N) use the same set of coefficients w', if the data points of u (N) and y (N) are N
u＝Y·ω′，u∈R N×1 ，
Figure BDA0004076515900000164
Calculating a global optimal solution using a least squares method: w' = (Y) H Y) -1 Y H u。
The parameter is set to L a ＝L b ＝L c ＝3，M b ＝M c =2 for adapting to the power amplifier distortion situation.
The lookup table is used for assisting the operation of the sine base function or the cosine base function in the solving process, and the method comprises the following steps:
all results of the logic circuit are calculated through PLD/FPGA development software, a table with x as an index value is manufactured, the table is written into the RAM in advance, and the corresponding value in the table is selected as the result to be output during calculation.
More specifically, the input baseband signal u is calculated by:
x＝U·ω′，x∈R N×1 ，
Figure BDA0004076515900000165
obtaining a predistortion signal x as an input of an actual power amplifier, and finally obtaining a linearly amplified output signal
Figure BDA0004076515900000166
And (4) carrying out self-adaptive iteration on the power amplifier distortion model, and repeating the steps S1 to S4 to finish self-adaptive adjustment.
And (3) self-adaptive iteration of a power amplifier distortion model: in the continuous operation process of the power amplifier, the distortion characteristics of the power amplifier are possibly changed due to the temperature change of components or the change of baseband signals, so that the basis functions and coefficients used for predistortion are updated in real time according to the latest power amplifier output; since the basis functions are updated at the same time, the number of coefficients does not increase greatly.
Example 2
In this embodiment, a power amplifier predistortion processing method based on a simplified GMP variant model includes the following steps:
s1: building a neural network model for selecting the basis functions, and selecting the basis functions;
s2: constructing a simplified GMP variant model based on the selected basis function as a power amplifier distortion model;
s3: solving the power amplifier distortion model coefficient to obtain a global optimal solution of the power amplifier distortion model coefficient;
s4: according to the obtained global optimal solution of the power amplifier distortion model coefficient, carrying out predistortion processing on a baseband signal u, wherein an input signal of the predistortion processing is the baseband signal u, the baseband signal u is subjected to predistortion processing to obtain a predistortion signal x which is used as an input of a power amplifier, and the predistortion signal x is subjected to power amplifier processing to obtain a linear amplified output signal
Figure BDA0004076515900000171
Example 3
In this embodiment, a power amplifier predistortion processing method based on a simplified GMP variant model includes the following steps:
building a neural network model for basis function selection; the neural network only bears the function of basic function selection, and the building of the reactive power amplification distortion model coefficient or parameter calculation is realized, so that the neural network does not depend on each data point of the input and output of the power amplifier seriously, and can be widely used; the formed neural network model can be directly used to effectively reduce the time cost of the hardware circuit before use, such as training time; a classical 3-layer convolutional neural network 3-CNN model may be used herein.
The neural network model performs basis function selection; the actual power amplifier distortion relation is approximated quickly by using the adapted basis functions.
And constructing a simplified GMP variant model as a power amplifier distortion model. Designing a corresponding GMP variant model based on the selected basis functions, and performing variants based on the conventional GMP model; the Taylor series is used for simplifying the model, and the number of coefficients is reduced. The coefficient calculation amount is reduced under the condition that necessary polynomials such as Gao Jiexiang and memory cross terms which can maintain the nonlinear fitting performance are reserved.
And calculating a global optimal solution of the power amplifier distortion model coefficient, wherein a lookup table is used for assisting in special basis function operation.
The predistortion processing of the original power amplifier input, namely the baseband signal u, namely the GMP variant model is used and the coefficient calculated in the last step is directly used, meanwhile, the power amplification multiple is set, the predistortion signal x is obtained as the input of the actual power amplifier, and finally the output signal after linear amplification is obtained
Figure BDA0004076515900000172
Look-up table assistance will be used here as well for special basis function operations.
Performing self-adaptive iteration on the power amplifier model; the basic function can be changed at any time to adapt to different input or power amplifier change conditions without adding a large number of coefficients to carry out self-adaptive adjustment.
It is to be understood that the above examples of the present invention are provided by way of illustration only and not by way of limitation of the embodiments of the present invention. Any modification, equivalent replacement, improvement, etc. which come within the spirit and principles of the invention are desired to be protected by the following claims.

Claims (10)

1. The power amplifier predistortion processing method based on the simplified GMP variant model is characterized by comprising the following steps of:
s1: building a neural network model for selecting the basis functions, and selecting the basis functions;
s2: constructing a simplified GMP variant model based on the selected basis function as a power amplifier distortion model;
s3: solving the power amplifier distortion model coefficient to obtain a global optimal solution of the power amplifier distortion model coefficient;
s4: according to the obtained global optimal solution of the power amplifier distortion model coefficient, carrying out predistortion processing on a baseband signal u, wherein an input signal of the predistortion processing is the baseband signal u, the baseband signal u is subjected to predistortion processing to obtain a predistortion signal x which is used as an input of a power amplifier, and the predistortion signal x is subjected to power amplifier processing to obtain a linear amplified output signal
Figure FDA0004076515880000011
2. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 1, wherein the neural network model uses a 3-CNN neural network model; the 3-CNN neural network model comprises a 1-layer convolution layer, a 1-layer pooling layer and a 1-layer full-connection layer.
3. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 2, wherein the process of constructing the neural network model is as follows:
training the neural network model by using the training set, and performing model evaluation on the neural network model by using the testing set; the training set and the testing set are from input and output signals with different time lengths generated by different power amplifiers at different moments;
acquiring characteristics through a power amplifier input and output sample set; the feature obtains a feature vector through calculation, and a label corresponding to the feature vector is a kind of a basic function;
and placing the features of the training set and the corresponding labels into a neural network model for model training, placing the features of the test set into the neural network model for model evaluation, comparing the result of the selection of the basis function of the neural network model with the correct labels to calculate a model fitting error, and adjusting the structural parameters of the neural network model to ensure that the model fitting error is smaller than a fitting error threshold.
4. A method for predistortion processing of a power amplifier based on a simplified GMP variant model according to claim 3, wherein the characteristic is time-frequency characteristic of a sequence, and the baseband signal and the original power amplifier output signal data point are set to N, the baseband signal is u (N), =1,, the original power amplifier output signal is y (N), =1,;
if p time-frequency characteristics calculated according to a section of sequence are provided, calculating a 2 p-long characteristic vector according to a group of u and y; if q groups of signals are arranged, calculating a feature matrix with the size of 2p x g; for any sequence s, the time-frequency characteristic and the calculation formula representing the arbitrary sequence s are as follows:
the time domain features are s_mean, s_std, s_max and s_min;
s_mean is the average of the vector s modulus, expressed as:
Figure FDA0004076515880000021
s_std is the variance of the vector s modulus, expressed as:
Figure FDA0004076515880000022
s_max is the maximum value of the vector s modulus, and the expression is s_max=max (|s|);
s_min is the minimum value of the vector s modulus value, and the expression is s_min=min (|s|);
the frequency domain features are p-max and f_pmax; p_max is the maximum value of the power amplitude by
Figure FDA0004076515880000023
-pi < omega < pi to obtain a power spectrum, wherein j is an imaginary number and omega is a frequency; calculating a power amplitude maximum value through a power spectrum;
f_pmax is the frequency value corresponding to the maximum value of the power amplitude in the power spectrum.
5. A power amplifier predistortion processing method based on a simplified GMP variant model according to claim 3, wherein the kinds of basis functions include a memory polynomial fitting basis function, a fourier polynomial fitting basis function, an euler polynomial fitting basis function; the basis function representation and the corresponding fitted polynomial are as follows:
the expression of the memory polynomial fitting basis function is:
u(n-l)|u(n-l)| k
wherein n is the data point sequence number of u, l is the memory depth, and k is the order;
the polynomial fit of the memory polynomial fit basis function is:
Figure FDA0004076515880000024
wherein a is k，l u is a coefficient of a base function corresponding to different memory depths and orders;
the fourier polynomial fit basis function is expressed as:
cos[u(n-l)]|cos[u(n-l)]| k
the fourier polynomial fit of the fourier polynomial fit basis function is:
Figure FDA0004076515880000025
the expression of the Euler polynomial fitting basis function is:
Figure FDA0004076515880000026
Figure FDA0004076515880000027
Figure FDA0004076515880000028
wherein A is n-l For the amplitude coefficient of the basic function, theta, under different memory depths n-l Frequency coefficients at different memory depths;
the euler polynomial fit of the euler polynomial fit basis function is:
Figure FDA0004076515880000031
wherein K is the order, L is the memory depth;
the baseband signal u corresponding to the characteristic vector and the original power amplifier output signal y are fitted by using polynomials of different basis functions, the memory depth and the order are set, the corresponding coefficient is calculated by using a least square method, and the fitting result is that
Figure FDA0004076515880000032
Error of fit error passing
Figure FDA0004076515880000033
Calculating;
setting a fitting error threshold, and if the fitting error is not smaller than the fitting error threshold, increasing the memory depth or the order until the fitting error is smaller than the threshold; the coefficients used by the polynomials of the different basis functions are compared, the basis function with the least coefficients used being the label that best fits the non-linear characteristics of the power amplifier model, the kind of basis function being marked as the eigenvector.
6. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 4, wherein the step of constructing the simplified GMP variant model is as follows:
the traditional GMP model is as follows:
Figure FDA0004076515880000034
where u is the baseband signal, y GMP To output a signal a k，l 、b k，l，m c k，l，m For coefficients of different terms, L a 、L b 、L c To memorize depth, K a 、K b 、K c For the order, M b 、M c For lead or lag length; calculating the number of coefficients of the traditional GMP model to be K a L a +K b L b M b +K c L c M c A plurality of; bringing the original power amplifier output signal y into y GMP Then, coefficient calculation is carried out;
when the memory polynomial fitting basis function and the fourier polynomial fitting basis function are selected separately, the conventional GMP model performs the following variants:
y GMP1 (n)＝y GMP (n)
Figure FDA0004076515880000041
Figure FDA0004076515880000042
let f (n, l, k) =f 1 (n，l)·|f 2 (n,l)| k As a basis function, the variant GMP model and corresponding values are as follows:
when f 1 (n, l) is u (n-l), f 2 When (n, l) is u (n-l), the corresponding GMP variant model is expressed as y GMP1 ；
When f 1 (n, l) is cos [ u (n-l)]，f 2 (n, l) is cos [ u (n-l)]When the corresponding GMP variant model is denoted as y GMP2 ；
When f 1 (n, l) is A n-l ，f 2 (n, l) is
Figure FDA0004076515880000043
When the corresponding GMP variant model is denoted as y GMP3 ；
And (3) carrying out coefficient simplification on the GMP variant model, and carrying out simplification on summation items of the GMP variant model after coefficient simplification, wherein the matrix form of the GMP variant model after simplification is expressed as follows:
Figure FDA0004076515880000044
where u' is a memory depth L based on the setting from the baseband signal data point u (n) a Advance length M b Length of hysteresis M c The calculated vector has a length of 2 (L a +L b M b +LcM c ) The method comprises the steps of carrying out a first treatment on the surface of the w' is the corresponding y GMP′ A coefficient vector of (n).
7. The method for predistortion processing of a power amplifier based on a simplified GMP variant model according to claim 6, wherein the step of coefficient simplifying the GMP variant model comprises the steps of:
when present
Figure FDA0004076515880000045
In the form, the definition domain is used as arbitrary value
Figure FDA0004076515880000046
Figure FDA0004076515880000051
/>
The two taylor formulas perform coefficient conversion, and the expression is:
Figure FDA0004076515880000052
when z l ＝|f 2 (n, l) is represented by the formula
Figure FDA0004076515880000053
K is an odd number;
or (b)
Figure FDA0004076515880000054
K is an even number;
if it is
Figure FDA0004076515880000055
q is an even number from 0 to K-1, and
Figure FDA0004076515880000056
q is an odd number from 0 to K-1;
Figure FDA0004076515880000057
8. the method for predistortion processing of a power amplifier based on a simplified GMP variant model according to claim 7, wherein the step of simplifying each summation term of the GMP variant model with simplified coefficients is as follows:
Figure FDA0004076515880000058
Figure FDA0004076515880000061
/>
Figure FDA0004076515880000062
Figure FDA0004076515880000063
the above expression is written in the following matrix form in turn:
Figure FDA0004076515880000064
finally, a matrix form of the simplified GMP variant model is obtained.
9. The method for predistortion processing of a power amplifier based on a simplified GMP variant model according to claim 8, wherein the step of solving a global optimal solution of the power amplifier distortion model coefficients is as follows:
Figure FDA0004076515880000065
Figure FDA0004076515880000071
according to the simplified GMP variant model, calculating a corresponding predistortion coefficient by utilizing the input and output of the power amplifier; that is, u' is the memory depth L according to the setting by the distortion output y (n) of the actual power amplifier a Advance length M b Length of hysteresis M c W ' is a coefficient vector corresponding to u (n), i.e., u (n) =u ' ·w ';
all u (N) use the same set of coefficients w', if the data points of u (N) and y (N) are N
u＝Y·ω′，u∈R N×1 ，
Figure FDA0004076515880000072
Calculating a global optimal solution using a least squares method: w' = (Y) H Y) -1 Y H u；
The parameter is set to L a ＝L b ＝L c ＝3，M b ＝M c =2, for adapting to the power amplifier distortion condition;
the lookup table is used for assisting the operation of the sine base function or the cosine base function in the solving process, and the method comprises the following steps:
all results of the logic circuit are calculated, a table with x as an index value is manufactured, the table is written into the RAM in advance, and the corresponding value in the table is selected as a result to be output during calculation.
10. The method for predistortion processing of a power amplifier based on a simplified GMP variant model according to claim 9, wherein the input baseband signal u is calculated by:
x＝U·w′，x∈R N×1 ，
Figure FDA0004076515880000073
obtaining a predistortion signal x as an input of an actual power amplifier, and finally obtaining a linearly amplified output signal
Figure FDA0004076515880000074
And (4) carrying out self-adaptive iteration on the power amplifier distortion model, and repeating the steps S1 to S4 to finish self-adaptive adjustment. />
CN202310110118.9A 2023-02-10 2023-02-10 Power amplifier predistortion processing method based on simplified GMP variant model Pending CN116245064A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310110118.9A CN116245064A (en) 2023-02-10 2023-02-10 Power amplifier predistortion processing method based on simplified GMP variant model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310110118.9A CN116245064A (en) 2023-02-10 2023-02-10 Power amplifier predistortion processing method based on simplified GMP variant model

Publications (1)

Publication Number Publication Date
CN116245064A true CN116245064A (en) 2023-06-09

Family

ID=86629112

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310110118.9A Pending CN116245064A (en) 2023-02-10 2023-02-10 Power amplifier predistortion processing method based on simplified GMP variant model

Country Status (1)

Country Link
CN (1) CN116245064A (en)

Similar Documents

Publication Publication Date Title
US9225501B2 (en) Non-linear modeling of a physical system using look-up table with polynomial interpolation
CN110414565B (en) Group Lasso-based neural network cutting method for power amplifier
CN104796111B (en) It is a kind of to be used for Dynamic Hysteresis system modelling and the nonlinear adaptable filter of compensation
CN111859795B (en) Polynomial aided neural network behavior modeling system and method for power amplifier
CN111245375B (en) Power amplifier digital predistortion method of complex-valued fully-connected recurrent neural network model
CN110765720B (en) Power amplifier predistortion method of complex-valued pipeline recurrent neural network model
CN109314390B (en) Equivalent conductance compensation type global linear symmetry method for obtaining DC power network power flow
WO2015107392A1 (en) Systems and methods for basis function orthogonalization for digital predistortion
CN113489014B (en) Quick and flexible full-pure embedded power system optimal power flow evaluation method
CN113037226B (en) Digital predistortion design method and device based on adaptive step length clipping method
US20230006611A1 (en) Ai-assisted power amplifier optimization
CN104796364A (en) Pre-distortion parameter solving method and pre-distortion processing system
Cao et al. Complexity optimized digital predistortion model of RF power amplifiers
CN102611661B (en) Predistortion device and method based on precise inverse solution memory polynomial model equation
CN116245064A (en) Power amplifier predistortion processing method based on simplified GMP variant model
CN107786174B (en) Circuit and method for predistortion
CN113055323B (en) Method and system for digital predistortion processing of communication system
CN108111448A (en) Generation method, device and the pre-distortion calibration equipment of predistortion lookup table
CN114598274B (en) Low-complexity lookup table construction method oriented to broadband predistortion
CN111884602B (en) Power amplifier predistortion method based on single-output-node neural network
CN109257948B (en) Power compensation type global linear symmetry method for obtaining DC power network tide
CN102522957B (en) A kind of method improving predistortion performance of radio-frequency power amplifier
Gong et al. A Memorized Recurrent Neural Network Design for Wide Bandwidth PA Linearization
CN108733852A (en) A kind of power amplifier behavior modeling method based on extreme learning machine
CN109257949B (en) Equivalent conductance compensation type global linear eccentricity method for obtaining DC power network power flow

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