CN111884602B - Power amplifier predistortion method based on single-output-node neural network - Google Patents
Power amplifier predistortion method based on single-output-node neural network Download PDFInfo
- Publication number
- CN111884602B CN111884602B CN202010608963.5A CN202010608963A CN111884602B CN 111884602 B CN111884602 B CN 111884602B CN 202010608963 A CN202010608963 A CN 202010608963A CN 111884602 B CN111884602 B CN 111884602B
- Authority
- CN
- China
- Prior art keywords
- signal
- neural network
- path
- power amplifier
- output
- 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.)
- Active
Links
Images
Classifications
-
- 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
Abstract
The invention discloses a power amplifier predistortion method based on a single output node neural network, and belongs to the technical field of communication. Different from the traditional predistortion method based on a neural network, the method only needs to fit any input signal in the I or Q path, and the other path of signal can be directly obtained by changing the input vector of the neural network without learning again. The reduction of the number of the output nodes of the neural network can effectively improve the fitting precision, reduce the calculated amount, save hardware resources and enhance the corresponding predistortion linearization performance.
Description
Technical Field
The invention belongs to the technical field of communication, and particularly relates to a power amplifier predistortion method based on a single-output node neural network.
Background
With the continuous development of communication technology, new modulation techniques with high spectral efficiency are widely used in order to more efficiently utilize limited spectrum resources. However, these modulation methods make the envelope fluctuation of the signal large and the peak-to-average ratio is high. Signals with the characteristics can generate serious nonlinear distortion after passing through a radio frequency power amplifier; the method not only causes the EVM of the signals in the frequency band to be seriously degraded, but also can generate serious out-of-band distortion, causes interference to adjacent channels and seriously influences the communication quality.
In order to guarantee communication quality and correct nonlinear distortion of signals generated by a power amplifier, the power amplifier needs to be linearized. In the current linearization technology, the predistortion technology is widely applied by the characteristics of low complexity, high stability, wide bandwidth, good linearization performance and the like.
The predistortion technology is that a predistorter with the characteristic opposite to that of a power amplifier is added at the front end of the power amplifier, a signal is subjected to predistortion treatment through the predistorter, the preprocessed signal passes through the power amplifier, and the characteristics of the predistorter and the power amplifier are mutually compensated, so that the final output signal of the power amplifier and the original input signal are in a linear relation.
As shown in the figure1 is shown as ViIs the input original signal, the function F (-) is the transfer function of the predistorter, the function G (-) is the transfer function of the power amplifier, the original input signal ViAn output signal V is obtained after passing through the predistorter and the power amplifierO＝G(F(Vi) Predistortion technique, i.e. by designing the transfer function F (-) of the predistorter such that the output signal is the same as the original input signal ViIn a linear relationship.
The predistortion technology needs to construct a behavior model of a power amplifier and an inverse model thereof, a digital predistortion system structure is shown in fig. 2, and the circuit mainly comprises: a predistorter, a digital-to-analog converter (D/A), a power amplifier, an attenuator, and an analog-to-digital converter (A/D). The pre-distortion processing of the signal is done in the digital domain: the input signal x (n) passes through a predistorter to obtain a predistortion signal Z (n), and the predistortion signal Z (n) becomes the input signal Z (n) of the radio frequency power amplifier after D/A conversion, modulation and up-conversionRF(n), and obtaining an output signal y after amplification by the power amplifierRF(t)，yRF(t) forming feedback through an attenuator with gain of 1/k (k is the expected gain of the power amplifier), wherein the feedback signal becomes another input signal y (n)/k of the predistorter after down-conversion, demodulation and A/D conversion, and the predistorter calculates the coefficient for updating the predistortion model according to x (n) and y (n)/k; usually, the hardware of the predistorter is implemented by an FPGA.
The power amplifier and the predistorter need to be behavioral modeled, the input z (n) of the power amplifier can be used as the input of a mathematical model, the output y (n) of the power amplifier can be used as the output of the mathematical model to obtain a behavioral model G (-) of the power amplifier, the predistorter and the power amplifier have inverse characteristics, the predistorter model can be solved by two methods, one is to solve the inverse model according to the behavioral model of the power amplifier, the other is to use y (n)/k as the input of the system, and x (n) as the output of the system to obtain a mathematical model, and the characteristics of the mathematical model are opposite to the characteristics of the power amplifier and can be used as a predistorter model function F (-).
One general class of existing power amplifier predistortion behavior models is based on neural networks. As shown in fig. 3, the input vector of such a neural network can be represented as:
Y＝[yr(n),yr(n-1),...,yr(n-D),
yi(n),yi(n-1)...yi(n-D),
|y(n)|,|y(n-1)|...|y(n-D)|,
|y(n)|2,|y(n-1)|2...|y(n-D)|2...
|y(n)|Q,|y(n-1)|Q...|y(n-D)|Q]
where y (n) represents the baseband signal output by the power amplifier, subscripts r and I represent the I/Q components, respectively, D represents the maximum memory depth, | · | represents the amplitude of the signal, and Q represents the order of maximum nonlinearity of the amplitude.
According to FIG. 3, the conventional neural network needs to fit two I/Q inputs, and the fitted target vector is
X＝[xr(n),xi(n)]
Where x (n) represents the baseband signal at the input of the power amplifier.
According to fig. 3, the output of each layer can be obtained
Wherein
Representing the weight connecting the ith neuron to the jth neuron output of the kth layer,
the bias of the jth neuron of the kth layer,
represents the jth output of the kth layer, fk(. DEG) represents an activation function of the k-th layer, the activation function of the hidden layer usually adopts a tansig function, and the activation function of the output layer adopts a linear activation function, wherein the tansig functionIs counted as
Model fitting the target to minimize the loss function as follows
Wherein xr(n) and xi(n) represents the actual I/Q component of the power amplifier input,
and
representing the I/Q component fitted by the neural network, N representing the number of samples, and when the LOSS function is minimum, the output of the neural network is closest to the true value, namely representing that the training of the neural network is completed.
The predistortion of the neural network adopts an indirect learning structure, and only the input vector correspondence needs to be changed into the input vector correspondence after the training is finished according to the principle of the indirect learning structure
X＝[xr(n),xr(n-1),...,xr(n-D),
xi(n),xi(n-1)...xi(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n|Q,|x(n-1)|Q...|x(n-D)|Q]
And then the I/Q component of the signal after predistortion can be obtained through the neural network trained before. The combined I/Q component is a baseband signal, and the predistortion signal passes through a power amplifier and then is in a linear relation with an original signal x (n).
However, the forward modeling and predistortion method based on the neural network has the following problems:
1. the neural network has large calculation amount and is not easy to realize engineering;
2. the neural network is based on a non-convex optimization algorithm, and when the network structure is large, the performance is easily influenced by an initial value;
3. the traditional neural network has less than ideal performance and needs to be improved.
Disclosure of Invention
The invention provides a power amplifier predistortion method based on a single-output node neural network, aiming at overcoming the defects of large calculation amount and the like of the original neural network.
The method comprises the following specific steps:
step one, constructing a single-output-node neural network on the basis of a traditional neural network, and only outputting an I path signal or a Q path signal;
the single output node neural network inputs are:
Y＝[yr(n),yr(n-1),...,yr(n-D),
yi(n),yi(n-1)...yi(n-D),
|y(n)|,|y(n-1)|...|y(n-D)|,
|y(n)|2,|y(n-1)|2...|y(n-D)|2...
|y(n)|Q,|y(n-1)|Q...|y(n-D)|Q]
wherein y isr(n) represents the I-component baseband signal of the power amplifier output, yi(n) represents a Q component baseband signal output by the power amplifier, D represents the maximum memory depth, | · | represents the signal amplitude, and Q represents the nonlinear order with the maximum amplitude.
And secondly, aiming at the single-output node neural network, only fitting the I path signal or the Q path signal of single-path input by the LOSS function, and realizing training by minimizing LOSS through a back propagation algorithm.
The loss function that fits the I-path signal for a single input only is:
fitting only single-input Q-path signalsThe loss function of (d) is:
xr(n) represents the I component of the true power amplifier output;
representing the component I fitted by the neural network, and N representing the number of samples; x is the number ofi(n) represents the Q component of the true power amplifier output,
representing the Q component fitted by the neural network;
step three, judging whether the rate of the predistortion signal processed by the user exceeds 1/2 of the hardware processing speed, if not, constructing two groups of different input vectors, respectively sequentially passing through the trained neural network to obtain the predistortion signal, and entering step five; otherwise, entering the step four;
specifically, for the case where the rate of the predistorted signal does not exceed 1/2, which is the speed of the hardware processing, if the I-path signal is fitted in step one, two sets of input signals X are constructed1And X2：
X1＝[xr(n),xr(n-1),...,xr(n-D),
xi(n),xi(n-1)...xi(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X2＝[xi(n),xi(n-1)...xi(n-D),
-xr(n),-xr(n-1),...,-xr(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
respectively obtaining I path components and Q path components of the predistortion signal by the two groups of input signals through a trained neural network, and combining the two paths of components to obtain the predistortion signal;
if step one fits the Q-way signals, then two sets of input signals X 'are constructed'1And X'2：
X'1＝[xr(n),xr(n-1),...,xr(n-D),
xi(n),xi(n-1)...xi(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X'2＝[-xi(n),-xi(n-1)...-xi(n-D),
xr(n),xr(n-1),...,xr(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
similarly, two groups of input signals respectively pass through the trained neural network to obtain Q path components and I path components, and the two path components are combined to obtain a predistortion signal.
Fourthly, copying the trained neural network, constructing two groups of different input vectors, and obtaining signals after pre-distortion processing through the trained neural network and the copied neural network respectively;
fitting the I-path signals to construct two groups of input signals X1And X2(ii) a Two paths of signals are respectively obtained through a neural networkAnd (3) pre-distorting the I path signal component and the Q path signal component after the pre-distortion treatment, and then combining the two paths of signal components to obtain a complete signal after the pre-distortion treatment.
Fitting the Q-path signals to construct two groups of input signals X'1And X'2(ii) a And respectively passing the two paths of signals through a neural network to respectively obtain the I path signal components and the Q path signal components after the pre-distortion treatment, and then combining the two paths of signal components to obtain the complete signal after the pre-distortion treatment.
And step five, enabling the predistortion signal to pass through a power amplifier, wherein the obtained output signal and the input signal x (n) of the power amplifier are in a linear relation.
The invention has the advantages that:
1) a power amplifier predistortion method based on neural network of single output node, different from traditional method, the fitting process of the invention only needs to fit any way in I/Q component, therefore train the goal to be less, only need use less neuron can reach better effect, have saved the hardware resource;
2) a power amplifier predistortion method based on single output node neural network, because there is only one training target, the neural network structure output layer has only one neuron, the computational rate is greatly improved.
3) A power amplifier predistortion method based on single output node neural network, because there is only one training target, has excluded the interference to fitting process of another way, the performance is superior to the traditional neural network.
4) Compared with the traditional neural network, the predistortion method of the power amplifier based on the single-output-node neural network has the advantages that the complexity of the network is low, hardware storage resources are saved, and the predistortion performance is improved.
Drawings
Fig. 1 is a circuit block diagram of a predistortion implementation principle in the prior art;
fig. 2 is a mechanism diagram of the implementation of predistortion in the prior art.
FIG. 3 is a conventional neural network model requiring fitting of two inputs to the I/Q.
Fig. 4 is a diagram of a single output node neural network architecture for power amplifier predistortion in accordance with the present invention.
Fig. 5 is a flowchart of a predistortion method of a power amplifier based on a neural network with a single output node according to the present invention.
Detailed Description
For the purpose of making the objects, aspects and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings.
The invention relates to a single-output node neural network for predistortion of a power amplifier, which can realize the construction of a predistorter only by fitting any one path of signal in I/Q; as is known, the calculated amount of the neural network is related to the coefficient amount contained in the neural network, namely weight and bias, the neural network adopted by the invention only has a single output node, compared with the traditional neural network with double output nodes, the output layer of the neural network is reduced by half of the weight, and due to the property of the neural network, two paths are fitted and one path is compared with the other path, and the interference of the other path is avoided by the fitting one path, so that the complexity of the single-node neural network is low compared with the traditional neural network, the hardware consumption can be saved, the performance can be improved, and the calculated amount is effectively reduced.
As shown in fig. 5, the specific steps are as follows:
step one, constructing a single-output-node neural network on the basis of a traditional neural network, and only outputting an I path signal or a Q path signal;
the input of the neural network is composed of a base function composed of orthogonal and homodromous components of the original current input of the power amplifier and the time delay input of the power amplifier, and the output of the neural network only contains one path of baseband signal in an I path or a Q path to be fitted. As shown in fig. 4, the constructed single-output node neural network inputs are:
Y＝[yr(n),yr(n-1),...,yr(n-D),
yi(n),yi(n-1)...yi(n-D),
|y(n)|,|y(n-1)|...|y(n-D)|,
|y(n)|2,|y(n-1)|2...|y(n-D)|2...
|y(n)|Q,|y(n-1)|Q...|y(n-D)|Q]
where y (n) represents the baseband signal output by the power amplifier, subscripts r and I represent the I/Q components, respectively, D represents the maximum memory depth, | · | represents the amplitude of the signal, and Q represents the nonlinear order of maximum amplitude. The input to be fitted is the I-way component xr(n) or Q-path component xi(n), with x in FIG. 4P(n) denotes, P ∈ { i, r }.
The single output node neural network is different from the traditional neural network in that: differences in network structure and differences in training patterns. Wherein the network result is that the output layer only has I path or Q path; the difference of the training mode is that the neural network with a single output node can obtain one path only by adjusting the input of the neural network after training and fitting the other path, while the traditional neural network directly fits two paths to directly obtain two paths of signals, so that the calculation amount is large and the performance is poor.
And secondly, aiming at the single-output node neural network, only fitting the I path signal or the Q path signal of single-path input by the LOSS function, and realizing training by minimizing LOSS through a back propagation algorithm.
If the I-path signal is fitted, minimizing the loss function through a back propagation algorithm as follows:
otherwise, the minimization of the loss function by the back propagation algorithm is:
xr(n) represents the I component of the true power amplifier output;
representing the component I fitted by the neural network, and N representing the number of samples; x is the number ofi(n) representsThe Q component of the true power amplifier output,
representing the Q component fitted by the neural network; when the LOSS function is minimum, the output of the neural network is closest to the true value, namely representing that the training of the neural network is completed. The trained neural network can be used as a predistorter.
Back-propagation algorithms, i.e. error back-propagation algorithms, such as Gradient device and Gradient device with Momentum, etc. Back propagation consists of two processes, forward propagation of the signal and back propagation of the error.
During forward propagation, an input sample enters a network from an input layer, is transmitted to an output layer by layer through a hidden layer, whether an error between actual output and expected output of the output layer is within an acceptance range is judged, if yes, modeling or predistortion of a power amplifier is directly carried out, and a learning algorithm is finished. Otherwise, turning to error reverse propagation;
the error satisfies within the acceptable range: the loss function < u; and u is a threshold value set artificially, the smaller u is, the higher the fitting precision is, and when the loss function is smaller than u, the algorithm is ended.
When the data is reversely transmitted, the output error (the difference between the expected output and the actual output) is reversely transmitted and calculated according to the original path, the data is reversely transmitted to the input layer through the hidden layer, the error is distributed to each unit of each layer by using a BP algorithm in the process of reverse transmission, the error signal of each unit of each layer is obtained, and the error signal is used as the basis for correcting the weight of each unit. After the weights and thresholds of the neurons in each layer are continuously adjusted, the error signals are minimized.
The process of continuously adjusting the weight and the threshold is the learning and training process of the network, and the adjustment of the weight and the threshold is repeatedly carried out through signal forward propagation and error backward propagation until the preset learning and training times or the output error is reduced to an allowable degree.
Step three, judging whether the speed of the predistortion signal processed by the user exceeds 1/2 of the hardware processing speed, if not, constructing two groups of different input vectors, respectively and sequentially passing through the trained neural network to obtain the predistortion signal, and entering the step five; otherwise, entering the step four;
both implementations are applicable to the case of low-speed signals and high-speed signals, respectively. For low-speed signals, two groups of different input signals are constructed and sequentially pass through the trained neural network to obtain an I/Q component of a predistortion signal, and the signal of the predistortion signal amplified by a power amplifier is in a linear relation with an original signal.
For high-speed signals, the trained neural network needs to be copied, two groups of different input signals are constructed and simultaneously pass through the two networks to obtain the I/Q component of a predistortion signal, and the signal of the predistortion signal amplified by a power amplifier is in a linear relation with the original signal.
Implementation mode 1: constructing input vector to obtain predistortion signal through neural network just trained
If the I path signals are fitted in the second step, two groups of input signals X are constructed1And X2：
X1＝[xr(n),xr(n-1),...,xr(n-D),
xi(n),xi(n-1)...xi(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X2＝[xi(n),xi(n-1)...xi(n-D),
-xr(n),-xr(n-1),...,-xr(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
and respectively obtaining the components of an I path and a Q path of the predistortion signal from the two groups of input signals through a trained neural network, combining the components of the two paths to obtain the predistortion signal, wherein the predistortion signal is in a linear relation with the input signal x (n) through the output of the power amplifier.
If the Q-path signals are fitted in the second step, two groups of input signals X 'are constructed'1And X'2：
X'1＝[xr(n),xr(n-1),...,xr(n-D),
xi(n),xi(n-1)...xi(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X'2＝[-xi(n),-xi(n-1)...-xi(n-D),
xr(n),xr(n-1),...,xr(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
the Q path component and the I path component are obtained by sequentially passing the two groups of input signals through the trained neural network respectively and are expressed as
And
combining two paths of components into
And obtaining a complete signal after predistortion processing, wherein the output of the predistortion signal passing through the power amplifier and the input signal x (n) are in a linear relation. .
Fourthly, copying the trained neural network, constructing two groups of different input vectors, and simultaneously obtaining a signal after pre-distortion treatment through the trained neural network and the copied neural network;
implementation mode 2: copying the neural network just trained to generate the same neural network as the trained neural network, and if the I-path signal is fitted in the step two, constructing two groups of input signals X1And X2：
X1＝[xr(n),xr(n-1),...,xr(n-D),
xi(n),xi(n-1)...xi(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X2＝[xi(n),xi(n-1)...xi(n-D),
-xr(n),-xr(n-1),...,-xr(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
and respectively passing the two paths of signals through a neural network to respectively obtain I path signal components and Q path signal components after pre-distortion treatment, then combining the two paths of signal components to obtain a complete signal after pre-distortion treatment, wherein the output of the pre-distortion signal after passing through a power amplifier and an input signal x (n) are in a linear relation.
If the Q-path signals are fitted in the second step, two groups of input signals X 'are constructed'1And X'2
X'1＝[xr(n),xr(n-1),...,xr(n-D),
xi(n),xi(n-1)...xi(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X'2＝[-xi(n),-xi(n-1)...-xi(n-D),
xr(n),xr(n-1),...,xr(n-D),
|x(n)|,|x(n-1)|...|x(n-D)|,
|x(n)|2,|x(n-1)|2...|x(n-D)|2...
|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
the two paths of signals respectively pass through a neural network in the same way, and the components of the Q path signal and the I path signal after the pre-distortion treatment are respectively obtained
And
then two paths of signal components are combined into
And obtaining a complete signal after predistortion processing, wherein the output of the predistortion signal after passing through the power amplifier and the input signal x (n) are in a linear relation.
And step five, enabling the predistortion signals obtained in the step three and the step four to pass through a power amplifier, wherein the obtained output signals and the input signals x (n) of the power amplifier are in a linear relation.
The two implementation modes adopted in the invention are different in that: the implementation mode 1 respectively requires faster hardware processing rate to meet the signal rate requirement through two groups of signals, and is suitable for low-speed signals; implementation 2 avoids the situation where the hardware processing rate does not meet the signal rate, and is suitable for high-speed signals, but uses additional hardware resources to process another set of the same network structures.
The invention has been carried out many times of experiments, and the experiments on most of the current power amplifiers are successful, and the method can obtain better predistortion performance than the traditional neural network.
Claims (2)
1. A power amplifier predistortion method based on a single output node neural network is characterized by comprising the following specific steps:
step one, constructing a single-output node neural network;
the constructed single-output node neural network has the following inputs:
Y＝[yr(n),yr(n-1),...,yr(n-D),yi(n),yi(n-1)...yi(n-D),|y(n)|,|y(n-1)|...|y(n-D)|,|y(n)|2,|y(n-1)|2...|y(n-D)|2...|y(n)|Q,|y(n-1)|Q...|y(n-D)|Q]
wherein y (n) represents the baseband signal output by the power amplifier, subscripts r and I represent the I/Q components of the sublist, D represents the maximum memory depth, | · | represents the amplitude of the signal, and Q represents the maximum order of the amplitude; the input to be fitted is the I-way component xr(n) or Q-path component xi(n) with xP(n) denotes, P ∈ { i, r };
step two, training the single-output node neural network, and if the I-path signal is fitted in the step 1, minimizing the I-path signal through a back propagation algorithm
If the Q path signal is fitted in the step 1, minimizing the Q path signal through a back propagation algorithm
Wherein x isr(n) and xi(n) represents the I/Q component of the true power amplifier output,
and
representing the I/Q component fitted by the neural network, wherein N represents the number of samples, and when the LOSS function is minimum, the output of the neural network is closest to a true value, namely representing that the training of the neural network is finished;
step three, step three has two different implementation modes
Implementation mode 1: construction of input vector predistortion signal obtained by neural network just trained if step 1 is performed on the I-path signal, two sets of input signals are constructed
X1＝[xr(n),xr(n-1),...,xr(n-D),xi(n),xi(n-1)...xi(n-D),|x(n)|,|x(n-1)|...|x(n-D)|,|x(n)|2,|x(n-1)|2...|x(n-D)|2...|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X2＝[xi(n),xi(n-1)...xi(n-D),-xr(n),-xr(n-1),...,-xr(n-D),|x(n)|,|x(n-1)|...|x(n-D)|,|x(n)|2,|x(n-1)|2...|x(n-D)|2...|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
respectively obtaining the components of the path I and the path Q of the predistortion signal by the two groups of input signals through the neural network trained in the step 2, combining the components of the two paths to obtain the predistortion signal, wherein the predistortion signal is in a linear relation with the input signal x (n) through the output of the power amplifier, and D represents the maximum memory depth; if the Q-path signals are fitted in the step 1, two groups of input signals are constructed
X1＝[xr(n),xr(n-1),...,xr(n-D),xi(n),xi(n-1)...xi(n-D),|x(n)|,|x(n-1)|...|x(n-D)|,|x(n)|2,|x(n-1)|2...|x(n-D)|2...|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X2＝[-xi(n),-xi(n-1)...-xi(n-D),xr(n),xr(n-1),...,xr(n-D),|x(n)|,|x(n-1)|...|x(n-D)|,|x(n)|2,|x(n-1)|2...|x(n-D)|2...|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
similarly, two groups of input signals are sequentially and respectively passed through the neural network trained in the step 2 to obtain the components of the path I and the path Q of the predistortion signal, the two paths of components are combined to obtain the predistortion signal, and the predistortion signal is in a linear relation with the input signal x (n) through the output of the power amplifier; the implementation mode 1 respectively requires faster hardware processing rate to meet the signal rate requirement through two groups of signals, and is suitable for low-speed signals;
implementation mode 2: duplicating the neural network just trained to generate two neural networks identical to those trained before, and if the I-path signal is fitted in step 1, constructing two groups of input signals
X1＝[xr(n),xr(n-1),...,xr(n-D),xi(n),xi(n-1)...xi(n-D),|x(n)|,|x(n-1)|...|x(n-D)|,|x(n)|2,|x(n-1)|2...|x(n-D)|2...|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X2＝[xi(n),xi(n-1)...xi(n-D),-xr(n),-xr(n-1),...,-xr(n-D),|x(n)|,|x(n-1)|...|x(n-D)|,|x(n)|2,|x(n-1)|2...|x(n-D)|2...|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
two paths of signals simultaneously pass through the two neural networks to obtain I path signal components and Q path signal components after pre-distortion treatment, the two paths of signal components are combined to obtain a complete signal after the pre-distortion treatment, and the output of the pre-distortion signal and an input signal x (n) form a linear relation through a power amplifier; if the Q-path signals are fitted in the step 1, two groups of input signals are constructed
X1＝[xr(n),xr(n-1),...,xr(n-D),xi(n),xi(n-1)...xi(n-D),|x(n)|,|x(n-1)|...|x(n-D)|,|x(n)|2,|x(n-1)|2...|x(n-D)|2...|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
And
X2＝[-xi(n),-xi(n-1)...-xi(n-D),xr(n),xr(n-1),...,xr(n-D),|x(n)|,|x(n-1)|...|x(n-D)|,|x(n)|2,|x(n-1)|2...|x(n-D)|2...|x(n)|Q,|x(n-1)|Q...|x(n-D)|Q]
the two paths of signals simultaneously pass through the two neural networks to obtain I path signal components and Q path signal components after pre-distortion treatment, the two paths of components are combined to obtain a complete signal after the pre-distortion treatment, and the output of the pre-distortion signal through a power amplifier and an input signal x (n) are in a linear relation; implementation 2 avoids the situation where the hardware processing rate does not meet the signal rate, and is suitable for high-speed signals, but uses additional hardware resources to process another set of the same network structures.
2. The method according to claim 1, wherein the neural network in step one is configured as a proposed single-output-node neural network, that is, the input of the neural network is formed by basis functions consisting of orthogonal and homodromous components of the original current input of the power amplifier and the time delay input of the power amplifier, and the output of the neural network only contains one baseband signal of the I path or the Q path to be fitted.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010608963.5A CN111884602B (en) | 2020-06-29 | 2020-06-29 | Power amplifier predistortion method based on single-output-node neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010608963.5A CN111884602B (en) | 2020-06-29 | 2020-06-29 | Power amplifier predistortion method based on single-output-node neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111884602A CN111884602A (en) | 2020-11-03 |
| CN111884602B true CN111884602B (en) | 2022-05-20 |
Family
ID=73157609
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010608963.5A Active CN111884602B (en) | 2020-06-29 | 2020-06-29 | Power amplifier predistortion method based on single-output-node neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111884602B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113325375B (en) * | 2021-05-25 | 2022-12-13 | 哈尔滨工程大学 | Self-adaptive cancellation method based on deep neural network |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101001327A (en) * | 2006-10-12 | 2007-07-18 | 成都新光微波工程有限责任公司 | Excitation apparatus of digital TV transmitter based on nerval net predistortion technology |
| WO2019190036A1 (en) * | 2018-03-27 | 2019-10-03 | 경상대학교산학협력단 | Method and apparatus for transmission power control based on neural network in wireless communication system |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6128609A (en) * | 1997-10-14 | 2000-10-03 | Ralph E. Rose | Training a neural network using differential input |
| US6771710B1 (en) * | 2000-10-17 | 2004-08-03 | Northrop Grumman Corporation | Partial frequency offset digital upconversion |
| CN100594669C (en) * | 2008-07-18 | 2010-03-17 | 东南大学 | Power amplifier predistortion method of Hammerstein model based on fuzzy neural network |
| CN102075469B (en) * | 2010-12-06 | 2013-01-02 | 北京邮电大学 | Estimation method for signal delay time of digital pre-distortion system |
| US8976893B2 (en) * | 2012-06-25 | 2015-03-10 | Telefonaktiebolaget L M Ericsson (Publ) | Predistortion according to an artificial neural network (ANN)-based model |
| WO2014184774A1 (en) * | 2013-05-16 | 2014-11-20 | Telefonaktiebolaget L M Ericsson (Publ) | Baseband equivalent volterra series for digital predistortion in multi-band power amplifiers |
| CN103856429A (en) * | 2014-03-26 | 2014-06-11 | 西安电子科技大学 | Adaptive predistortion system and method based on hybrid indirect learning algorithm |
-
2020
- 2020-06-29 CN CN202010608963.5A patent/CN111884602B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101001327A (en) * | 2006-10-12 | 2007-07-18 | 成都新光微波工程有限责任公司 | Excitation apparatus of digital TV transmitter based on nerval net predistortion technology |
| WO2019190036A1 (en) * | 2018-03-27 | 2019-10-03 | 경상대학교산학협력단 | Method and apparatus for transmission power control based on neural network in wireless communication system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111884602A (en) | 2020-11-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110414565B (en) | Group Lasso-based neural network cutting method for power amplifier | |
| CN111490737B (en) | Nonlinear compensation method and equipment for power amplifier | |
| CN111245375B (en) | Power amplifier digital predistortion method of complex-valued fully-connected recurrent neural network model | |
| US20220200540A1 (en) | Model trainer for digital pre-distorter of power amplifiers | |
| WO2003092154A1 (en) | The method of improving the radio frequency power amplifier efficiency based on the baseband digital pre-distortion technique | |
| CN110765720A (en) | Power amplifier predistortion method of complex value assembly line recurrent neural network model | |
| Lee et al. | Adaptive neuro-fuzzy inference system (ANFIS) digital predistorter for RF power amplifier linearization | |
| Tripathi et al. | Swish activation based deep neural network predistorter for RF-PA | |
| CN115278746A (en) | Self-adaptive fast error correction digital predistortion method for 5G broadband power amplifier | |
| CN111884602B (en) | Power amplifier predistortion method based on single-output-node neural network | |
| CN111064439A (en) | System and method for improving short-wave digital predistortion performance | |
| CN111200470A (en) | High-order modulation signal transmission control method suitable for being interfered by nonlinearity | |
| CN102522957B (en) | A kind of method improving predistortion performance of radio-frequency power amplifier | |
| CN115913844A (en) | MIMO system digital predistortion compensation method, device, equipment and storage medium based on neural network | |
| Li et al. | Data predistortion with adaptive fuzzy systems | |
| Al-Ja'afari et al. | A simplified adaptive digital predistorter for a high power amplifier with memory effects based on direct learning architecture | |
| Elaskary et al. | Closed-form analysis for indirect learning digital pre-distorter for wideband wireless systems | |
| Long et al. | Indirect learning hybrid memory predistorter based on polynomial and look-up-table | |
| TWI810056B (en) | Method performed by rf communication assembly and rf communication assembly | |
| CN115913140B (en) | Piecewise polynomial digital predistortion device and method for controlling operation precision | |
| Uncini et al. | Fast complex adaptive spline neural networks for digital signal processing | |
| RU2731128C1 (en) | Method for combined digital linearization of power amplifier and quadrature modulator | |
| TWI830276B (en) | Method of compensating for power amplifier distortions and system | |
| Guo et al. | Research for adaptive digital predistortion based on BP-LMS | |
| JP2005079935A (en) | Adaptive predistortion type distortion compensation power amplifier |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |