CN108768550B - Wide-band transmitter nonlinear modeling method based on dynamic multi-core bandwidth generalized regression neural network algorithm - Google Patents
Wide-band transmitter nonlinear modeling method based on dynamic multi-core bandwidth generalized regression neural network algorithm Download PDFInfo
- Publication number
- CN108768550B CN108768550B CN201810648237.9A CN201810648237A CN108768550B CN 108768550 B CN108768550 B CN 108768550B CN 201810648237 A CN201810648237 A CN 201810648237A CN 108768550 B CN108768550 B CN 108768550B
- Authority
- CN
- China
- Prior art keywords
- fingerprint
- amplitude
- frequency
- neural network
- dynamic
- 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
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/10—Monitoring; Testing of transmitters
- H04B17/11—Monitoring; Testing of transmitters for calibration
- H04B17/13—Monitoring; Testing of transmitters for calibration of power amplifiers, e.g. gain or non-linearity
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/0082—Monitoring; Testing using service channels; using auxiliary channels
- H04B17/0087—Monitoring; Testing using service channels; using auxiliary channels using auxiliary channels or channel simulators
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/10—Monitoring; Testing of transmitters
- H04B17/11—Monitoring; Testing of transmitters for calibration
- H04B17/12—Monitoring; Testing of transmitters for calibration of transmit antennas, e.g. of the amplitude or phase
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B1/00—Details of transmission systems, not covered by a single one of groups H04B3/00 - H04B13/00; Details of transmission systems not characterised by the medium used for transmission
- H04B1/02—Transmitters
- H04B1/04—Circuits
- H04B2001/0408—Circuits with power amplifiers
- H04B2001/0425—Circuits with power amplifiers with linearisation using predistortion
Abstract
The invention discloses a wide-band transmitter nonlinear modeling method based on a dynamic multi-core bandwidth generalized regression neural network algorithm, which comprises the following steps: building a test platform, measuring and recording the nonlinear characteristics of the full working frequency band of the wide-frequency-band transmitter, and collecting the input and output amplitudes and phase values of a test signal and the corresponding carrier frequency; carrying out segmentation processing on the signal sample according to carrier frequency, quantizing the amplitude of an input signal by using an amplitude non-uniform quantization technology, and realizing the construction of a dynamic non-linear fingerprint sample data set by using a dynamic memory fingerprint technology; and training the generalized regression neural network by using the dynamic nonlinear fingerprint sample data set, and combining an optimization algorithm to obtain a dynamic multi-core bandwidth generalized regression neural network model so as to realize modeling of full-band nonlinear characteristics of the wideband transmitter.
Description
Technical Field
The invention relates to the technical field of design of transmitters in wireless communication, in particular to a nonlinear distortion modeling and correcting technology of a transmitter. The signal of the wide-band transmitter is measured when the transmitter works in the full band, the characteristic modes in the signal are extracted to form training sample data, the nonlinear characteristic of the transmitter is fitted by utilizing a learning algorithm and a neural network model, and a model basis is provided for linearization technologies such as transmitter predistortion correction and the like.
Background
The nonlinear distortion of the wideband transmitter is mainly caused by the nonlinear characteristic of the rf power amplifier. Mainly represented by amplitude distortion and phase distortion of the signal: after the signal passes through the nonlinear power amplifier, the gain of the amplitude of the output signal changes along with the amplitude of the input signal, and meanwhile, the change amount of the phase also changes along with the amplitude of the input signal. The nonlinear distortion not only causes compression of an amplified signal, but also generates new frequency components, causes spectrum expansion, seriously affects the signal-to-noise ratio of a signal in an operating frequency band, also generates interference on adjacent channels, and particularly has great harm to a networking type communication system.
Transmitter non-linearity can be generally represented by AM-AM distortion and AM-PM distortion. Wherein AM-AM represents amplitude distortion generated after signal amplification, and AM-PM represents phase distortion generated after signal amplification.
Besides AM-AM distortion and AM-PM distortion, the radio frequency power amplifier also has memory effect. From the time domain perspective, the memory effect is that the current output signal of the rf power amplifier not only depends on the current input signal, but also depends on the past input signal, i.e. the distortion of the device depends on the previous rf output power of 10 to 20 nanoseconds. It is essential that the instantaneous temperature in the device channel affects the distortion; from the perspective of the frequency domain, the memory effect can be defined as the phenomenon that the amplitude and phase characteristics of the radio frequency power amplifier change along with the change of the envelope frequency of the input signal. The largest source of memory effect is the fluctuation of output amplitude and phase caused by the change of impedance of nodes of a matching network and a bias network due to the change of signal envelope frequency. Therefore, with the increase of communication bandwidth, the memory effect of the radio frequency power amplifier becomes more and more obvious, the description of the memory effect is added on the basis of the non-linear model of the memory-free radio frequency power amplifier to establish the memory radio frequency power amplifier model, and the complexity of the memory radio frequency power amplifier model is in direct proportion to the depth of the memory effect.
Due to the wide operating band of wideband transmitters, multiple octaves may be spanned. It is known from practical measurement that in this mode, the nonlinear characteristic of the transmitter changes significantly with the frequency change, the magnitude of the change does not change monotonically with the change in frequency, and the wide-band operation also brings about a serious memory effect, so that it is difficult to describe the nonlinear characteristic of the wide-band transmitter by using a simple mathematical expression.
The Generalized Regression Neural Network (GRNN) model is a nuclear Regression mathematical model based on density estimation. The classical generalized regression neural network model consists of an input layer, a mode layer, a summation layer and an output layer. The kernel function of the method follows multivariate Gaussian distribution, and the method has good nonlinear fitting performance and anti-noise capability. The nuclear estimate is given by
Where σ is the bandwidth of the kernel function, Xi is the observed sample value, yi is the output value, and X is the input value. The estimate may be considered as a weighted average of all observed sample values.
Disclosure of Invention
The purpose of the invention is as follows: in order to overcome the defects in the prior art, the invention provides a modeling method for nonlinear characteristics of a wide-band transmitter, which is based on the thought of machine learning, extracts and screens nonlinear characteristic patterns of the wide-band transmitter through a heuristic learning and incremental learning method, trains a model by using a neural network algorithm, has flexible model structure, less parameters to be trained, good fitting performance and strong anti-noise capability, and can accurately represent the nonlinear characteristics of the transmitter in a wide band.
In order to achieve the above and other objects, the present invention provides a wide band transmitter nonlinear modeling method based on dynamic multi-core bandwidth generalized regression neural network algorithm, comprising the steps of: building a test platform, measuring and recording the nonlinear characteristics of the full working frequency band of the wide-frequency-band transmitter, and collecting the input and output amplitudes and phase values of a test signal and the corresponding carrier frequency; carrying out segmentation processing on the signal sample according to carrier frequency, quantizing the amplitude of an input signal by using an amplitude non-uniform quantization technology, and realizing the construction of a dynamic non-linear fingerprint sample data set by using a dynamic memory fingerprint technology; and training the generalized regression neural network by using the dynamic nonlinear fingerprint sample data set, and combining an optimization algorithm to obtain a dynamic multi-core bandwidth generalized regression neural network model so as to realize modeling of full-band nonlinear characteristics of the wideband transmitter.
The test platform consists of a vector signal generator, a vector signal analyzer, an attenuator, a PC and other equipment. The testing process is completed through an instrument automatic measurement software program, the PC machine controls the testing process through a network cable or a GPIB (general purpose interface bus), and the testing data is stored in the PC machine.
The signals tested were: amplitude A of input digital baseband signal in vector signal generatorinPhase thetainAnd a carrier frequency F of the up-converted RF signalc(ii) a Amplitude A of demodulated digital baseband signal in vector signal analyzerfbPhase thetafb。
The test procedure flow is as follows:
step 1: setting the bandwidth, sampling rate, modulation mode, roll-off coefficient, etc. of digital baseband signal in vector network generator and vector signal analyzer, wherein the sampling rate f of digital signalsampleMust be greater than 100 Mps;
step 2: setting the frequency of a radio frequency carrier signal of a vector network generator to be FiThe signal source is a section of code comprising a full dynamic range of baseband signal amplitude A and phase theta;
and step 3: extracting data and test results of the vector network generator and the vector signal analyzer, and sending the data and the test results to a database of the PC;
and 4, step 4: and (3) increasing the frequency value of the radio frequency carrier signal in the vector network generator according to preset frequency stepping delta F, and repeating the steps 2-3 until the test of all working frequency points of the transmitter is completed (the working bandwidth of the transmitter is B, and n is mod (B/delta F) +1 frequency point in total).
After the test process is completed, processing the signal data obtained after measurement to obtain a dynamic nonlinear fingerprint sample data set:
step 1: the frequency segmentation technology can classify and process the working frequency band according to the frequency-nonlinearity of the transmitter, determine the numerical value and the stepping of the frequency fingerprint point, and control the number and the bandwidth of the neurons in the mode layer of the dynamic multi-core bandwidth generalized regression neural network.
Step 1: determining full-band transmission gain value G of transmitter0Generally determined by the lowest transmission gain value of the transmitter in the full operating band;
step 2: according to the measured data, calculating the gain vector of the transmitter under each carrier frequency, and normalizing G0 to obtain { Gi,k,i∈(0,max(A)],k∈[1,n]};
And step 3: calculating Euclidean distance d between frequency gain vectorsi,jObtaining a distance matrix D:
and 4, step 4: classifying the frequency vectors by a clustering algorithm, such as a K-means clustering algorithm, a DBSCAN density clustering algorithm or an AGNES hierarchical clustering algorithm, dividing the working frequency band into p continuous sub-frequency bands, and obtaining a frequency fingerprint set { f) consisting of the initial frequency values of each frequency bandi,(i∈[1,p])}。
The amplitude non-uniform quantization technology can perform segmented processing on the amplitude of an input signal according to the signal amplitude gain compression ratio, namely AM-AM distortion, determine the numerical value and the stepping of an amplitude fingerprint point, and control the number and the bandwidth of neurons in a dynamic multi-core bandwidth generalized regression neural network mode layer:
step 1: calculating the mean value of the gain vectors in each frequency band { MGi,k,i∈(0,max(A)],k∈[1,p]};
Step 2: each mean vector obtained in step 1 is subjected to the following processing using a linear rectification function (RELU):
w(MGi,k)=RELU(G0-MGi,k)=max(0,G0-MGi,k) (6)
here, the point w at which the gain is larger than G0 is 0, and the nonlinearity of the points corresponding to these amplitudes is considered to be small, and therefore the amplitude fingerprint value is classified as 0. According to the AM-AM characteristic of the power amplifier, the gain compression degree is in direct proportion to the amplitude of the input signal, and can be approximately regarded as monotone change. To eliminate noise interference, the data set w is subjected toi,k,i∈(0,max(A)],k∈[1,p]And (4) performing polynomial fitting treatment to obtain a continuous smooth gain compression curve g (w).
And step 3: setting a threshold value delta G, wherein the larger the threshold value is, the longer the length of a segmentation interval is, the fewer the number of segments is, and the lower the model precision is; the smaller the threshold value is, the shorter the length of the segmentation interval is, the more the number of segments is, and the higher the model precision is;
and 4, step 4: and (3) carrying out segmentation processing on the gain compression curve G (w) obtained in the step (2) according to a threshold value delta G to obtain an amplitude fingerprint value:
setting the starting point of the current amplitude fingerprint to correspond to the z (initial value is 1) th amplitude value AkLet j be z +1,
if g (w)j)-g(wz) If Δ G, z is j +1, and A is substitutedkSetting the next amplitude fingerprint value, and then repeating the step 4;
if g (w)j)-g(wz) If j is not larger than deltaG, j +1, and repeating the step 4;
and 5: after the step 4 is completed on the whole data set, an amplitude fingerprint set { alpha ] is obtainedi,i∈(0,max(A)]};
Step 6: repeating the steps 4 and 5 on the data set { MG } obtained by the formula (6) of all the p frequency bands to obtain a full-frequency-band amplitude fingerprint data set { alphai,k,i∈(0,max(A)],k∈[1,p]}。
The phase of the signal is fingerprint at 0, 2 pi]And (3) internal uniform quantization, setting a phase interval delta theta, and setting the length of the phase fingerprint set as follows: r-INT (2 pi/Δ θ). The smaller r, the lower the model accuracy; conversely, the higher the model accuracy. Quantized set of phase fingerprints [ theta ]j,j∈[1,r]}
The dynamic memory fingerprint technology can control the number of neurons of an input layer of the dynamic multi-core bandwidth generalized regression neural network and the dimensionality of a mode layer kernel function by changing the memory depth m of the amplitude and the carrier frequency of an input signal in the nonlinear fingerprint sample data.
The dynamic nonlinear fingerprint sample structure:
{[α(t),α(t-1),...,α(t-m)],[f(t),f(t-1),...,f(t-m)],θ}
the dynamic multi-core bandwidth generalized regression neural network topology is composed of an input layer, a mode layer, a summation layer and an output layer, and refer to fig. 3.
The number of neurons in the input layer is v ═ 2 × m +1, and the number of neurons corresponds to an amplitude vector having a length of m and m frequency vectors in the fingerprint data, and 1 phase data, respectively;
the number of the neurons in the mode layer is jointly determined by the length p of the frequency fingerprint set, the length r of the amplitude fingerprint set, the length r of the phase fingerprint set and the memory depth m;
the summation layer comprises 4 neurons;
the output layer comprises 2 neurons which respectively correspond to the amplitude value and the phase value estimated by the model.
Training the generalized regression neural network by using a dynamic nonlinear fingerprint sample data set, wherein the trained target parameters are a Gaussian kernel bandwidth set { sigma (sigma) corresponding to pattern layer neuronsi,i∈[1,v]}。
The Gaussian kernel bandwidth parameter can be obtained by an optimization algorithm, such as a mathematical method of a gradient descent method, a Newton method, a particle swarm algorithm, a genetic algorithm, an annealing algorithm and the like. The modeling of the full-band nonlinear characteristics of the wide-band transmitter is realized.
The dynamic multi-core bandwidth generalized regression neural network model adopts an incremental learning mode, can further improve the model precision by inputting new training samples, can track the nonlinear change of the power amplifier in time, corrects parameters and improves the robustness of the model.
The invention can realize accurate modeling of the nonlinear characteristic of the wide-band radio transmitter, extracts and screens the nonlinear characteristic mode of the wide-band transmitter by a heuristic learning and incremental learning method based on the thought of machine learning, trains the model by utilizing a neural network algorithm, has flexible model structure, less parameters to be trained, good fitting performance and strong anti-noise capability, can accurately represent the nonlinear characteristic of the transmitter in the wide band, can track the nonlinear change of a power amplifier in time, corrects the parameters and improves the robustness of the model.
Drawings
Fig. 1 is a flow chart of wideband transmitter nonlinear characteristic modeling according to the present invention.
FIG. 2 is a schematic diagram of a wideband transceiver nonlinear characteristic testing platform according to the present invention.
FIG. 3 is a schematic diagram of dynamic nonlinear fingerprint sample data set construction in the present invention.
FIG. 4 is a schematic structural diagram of a dynamic multi-core bandwidth generalized regression neural network model in the present invention.
Detailed Description
Other advantages and effects of the present invention will become apparent to those skilled in the art from the following description, which, taken in conjunction with the annexed drawings, discloses embodiments of the present invention. The invention is capable of other and different embodiments and its several details are capable of modification in various other respects, all without departing from the spirit and scope of the present invention.
Referring to fig. 1, the method includes the steps of: the method comprises the steps of setting up a test platform, measuring and recording nonlinear characteristics of the full-working frequency band of the wide-band transmitter, collecting input and output amplitude and phase values of test signals and corresponding carrier frequencies, carrying out segmentation processing on signal samples according to the carrier frequencies, then carrying out non-uniform quantization processing on the amplitude of the input signals, realizing construction of a dynamic nonlinear fingerprint sample data set by adopting a dynamic memory fingerprint technology, training a generalized regression neural network by using the fingerprint sample data, obtaining a dynamic multi-core bandwidth generalized regression neural network model by combining an optimization algorithm, and realizing modeling of the full-band nonlinear characteristics of the wide-band transmitter.
Referring to fig. 2, the test platform is composed of a vector signal generator, a vector signal analyzer, an attenuator, and a PC. The testing process is completed through an instrument automatic measurement software program, the PC machine controls the testing process through a network cable or a GPIB (general purpose interface bus), and the testing data is stored in the PC machine.
The signals tested were: amplitude A of input digital baseband signal in vector signal generatorinPhase thetainAnd a carrier frequency F of the up-converted RF signalc(ii) a Amplitude A of demodulated digital baseband signal in vector signal analyzerfbPhase thetafb。
The test procedure flow is as follows:
step 1: setting the bandwidth, sampling rate, modulation mode, roll-off coefficient, etc. of digital baseband signal in vector network generator and vector signal analyzer, wherein the sampling rate f of digital signalsampleMust be greater than 100 Mps;
step 2: setting the frequency of a radio frequency carrier signal of a vector network generator to be FiThe signal source is a section of code comprising a full dynamic range of baseband signal amplitude A and phase theta;
and step 3: extracting data and test results of the vector network generator and the vector signal analyzer, and sending the data and the test results to a database of the PC;
and 4, step 4: and (3) increasing the frequency value of the radio frequency carrier signal in the vector network generator according to preset frequency stepping delta F, and repeating the steps 2-3 until the test of all working frequency points of the transmitter is completed (the working bandwidth of the transmitter is B, and n is mod (B/delta F) +1 frequency point in total).
Referring to fig. 3, after the test flow is completed, the signal data obtained after measurement is processed to obtain a dynamic nonlinear fingerprint sample data set:
step 1: the frequency segmentation technology can classify and process the working frequency band according to the frequency-nonlinearity of the transmitter, determine the numerical value and the stepping of the frequency fingerprint point, and control the number and the bandwidth of the neurons in the mode layer of the dynamic multi-core bandwidth generalized regression neural network.
Step 1: determining full-band transmission gain value G of transmitter0Generally determined by the lowest transmission gain value of the transmitter in the full operating band;
And step 3: calculating Euclidean distance d between frequency gain vectorsi,jObtaining a distance matrix D:
and 4, step 4: classifying the frequency vectors by a clustering algorithm, such as a K-means clustering algorithm, a DBSCAN density clustering algorithm or an AGNES hierarchical clustering algorithm, dividing the working frequency band into p continuous sub-frequency bands, and obtaining a frequency fingerprint set { f) consisting of the initial frequency values of each frequency bandi,(i∈[1,p])}。
The amplitude non-uniform quantization technology can perform segmented processing on the amplitude of an input signal according to the signal amplitude gain compression ratio, namely AM-AM distortion, determine the numerical value and the stepping of an amplitude fingerprint point, and control the number and the bandwidth of neurons in a dynamic multi-core bandwidth generalized regression neural network mode layer:
step 1: calculating the mean value of the gain vectors in each frequency band { MGi,k,i∈(0,max(A)],k∈[1,p]};
Step 2: each mean vector obtained in step 1 is subjected to the following processing using a linear rectification function (RELU):
w(MGi,k)=RELU(G0-MGi,k)=max(0,G0-MGi,k) (6)
here, the point w at which the gain is larger than G0 is 0, and the nonlinearity of the points corresponding to these amplitudes is considered to be small, and therefore the amplitude fingerprint value is classified as 0. According to the AM-AM characteristic of the power amplifier, the gain compression degree is in direct proportion to the amplitude of the input signal, and can be approximately regarded as monotone change. To eliminate noise interference, the data set w is subjected toi,k,i∈(0,max(A)],k∈[1,p]And (4) performing polynomial fitting treatment to obtain a continuous smooth gain compression curve g (w).
And step 3: setting a threshold value delta G, wherein the larger the threshold value is, the longer the length of a segmentation interval is, the fewer the number of segments is, and the lower the model precision is; the smaller the threshold value is, the shorter the length of the segmentation interval is, the more the number of segments is, and the higher the model precision is;
and 4, step 4: and (3) carrying out segmentation processing on the gain compression curve G (w) obtained in the step (2) according to a threshold value delta G to obtain an amplitude fingerprint value:
setting the starting point of the current amplitude fingerprint to correspond to the z (initial value is 1) th amplitude value AkLet j be z +1,
if g (w)j)-g(wz) If Δ G, z is j +1, and A is substitutedkSetting the next amplitude fingerprint value, and then repeating the step 4;
if g (w)j)-g(wz) If j is not larger than deltaG, j +1, and repeating the step 4;
and 5: after the step 4 is completed on the whole data set, an amplitude fingerprint set { alpha ] is obtainedi,i∈(0,max(A)]};
Step 6: repeating the steps 4 and 5 on the data set { MG } obtained by the formula (6) of all the p frequency bands to obtain a full-frequency-band amplitude fingerprint data set { alphai,k,i∈(0,max(A)],k∈[1,p]}。
The phase fingerprint value of the signal is uniformly quantized in [0, 2 pi ], and the phase interval delta theta is set, so that the length of the phase fingerprint set is as follows: r-INT (2 pi/Δ θ).
The smaller r, the lower the model accuracy; conversely, the higher the model accuracy. Quantized set of phase fingerprints [ theta ]j,j∈[1,r]}
The dynamic memory fingerprint technology can control the number of neurons of an input layer of the dynamic multi-core bandwidth generalized regression neural network and the dimensionality of a mode layer kernel function by changing the memory depth m of the amplitude and the carrier frequency of an input signal in the nonlinear fingerprint sample data.
The dynamic nonlinear fingerprint sample structure:
{[α(t),α(t-1),...,α(t-m)],[f(t),f(t-1),...,f(t-m)],θ}
referring to fig. 4, the dynamic multi-core bandwidth generalized regression neural network topology is composed of an input layer, a mode layer, a summation layer, and an output layer, referring to fig. 3.
The number of neurons in the input layer is v ═ 2 × m +1, and the number of neurons corresponds to an amplitude vector having a length of m and m frequency vectors in the fingerprint data, and 1 phase data, respectively;
the number of the neurons in the mode layer is jointly determined by the length p of the frequency fingerprint set, the length r of the amplitude fingerprint set, the length r of the phase fingerprint set and the memory depth m;
the summation layer comprises 4 neurons;
the output layer comprises 2 neurons which respectively correspond to the amplitude value and the phase value estimated by the model.
Training the generalized regression neural network by using a dynamic nonlinear fingerprint sample data set, wherein the trained target parameters are a Gaussian kernel bandwidth set { sigma (sigma) corresponding to pattern layer neuronsi,i∈[1,v]}。
The Gaussian kernel bandwidth parameter can be obtained by an optimization algorithm, such as a mathematical method of a gradient descent method, a Newton method, a particle swarm algorithm, a genetic algorithm, an annealing algorithm and the like. The modeling of the full-band nonlinear characteristics of the wide-band transmitter is realized.
The dynamic multi-core bandwidth generalized regression neural network model adopts an incremental learning mode, can further improve the model precision by inputting new training samples, can track the nonlinear change of the power amplifier in time, corrects parameters and improves the robustness of the model.
The invention can realize accurate modeling of the nonlinear characteristic of the wide-band radio transmitter, extracts and screens the nonlinear characteristic mode of the wide-band transmitter by a heuristic learning and incremental learning method based on the thought of machine learning, trains the model by utilizing a neural network algorithm, has flexible model structure, less parameters to be trained, good fitting performance and strong anti-noise capability, can accurately represent the nonlinear characteristic of the transmitter in the wide band, can track the nonlinear change of a power amplifier in time, corrects the parameters and improves the robustness of the model.
The above description is only for the best mode of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.
Claims (8)
1. A wide frequency band transmitter nonlinear modeling method based on a dynamic multi-core bandwidth generalized regression neural network algorithm comprises the following steps: building a test platform, measuring and recording the nonlinear characteristics of the full working frequency band of the wide-frequency-band transmitter, and collecting the input and output amplitudes and phase values of a test signal and the corresponding carrier frequency; carrying out segmentation processing on the signal sample according to carrier frequency, quantizing the amplitude of an input signal by using an amplitude non-uniform quantization technology, and realizing the construction of a dynamic non-linear fingerprint sample data set by using a dynamic memory fingerprint technology; training a generalized regression neural network by using a dynamic nonlinear fingerprint sample data set, and combining an optimization algorithm to obtain a dynamic multi-core bandwidth generalized regression neural network model so as to realize modeling of full-band nonlinear characteristics of a wideband transmitter;
the dynamic memory fingerprint technology controls the number of neurons of an input layer of the dynamic multi-core bandwidth generalized regression neural network and the dimensionality of a mode layer kernel function by changing the memory depth m of the amplitude and the carrier frequency of an input signal in nonlinear fingerprint sample data;
the dynamic nonlinear fingerprint sample structure:
{[α(t),α(t-1),...,α(t-m)],[f(t),f(t-1),...,f(t-m)],θ}
wherein m represents the memory depth value of the nonlinear model, the parameter alpha represents the carrier amplitude value, f represents the carrier frequency value, and theta is the phase;
the dynamic multi-core bandwidth generalized regression neural network topology consists of an input layer, a mode layer, a summation layer and an output layer;
the number of neurons in the input layer is v ═ 2 × m +1, and the number of neurons corresponds to the amplitude fingerprint and the frequency fingerprint with the memory depth of m and 1 phase data respectively;
the number of the neurons in the mode layer is jointly determined by the length p of the frequency fingerprint set, the length r of the amplitude fingerprint set, the length r of the phase fingerprint set and the memory depth m;
the summation layer comprises 4 neurons;
the output layer comprises 2 neurons which respectively correspond to the amplitude value and the phase value estimated by the model;
training the generalized regression neural network by using a dynamic nonlinear fingerprint sample data set, wherein the trained target parameters are a Gaussian kernel bandwidth set { sigma (sigma) corresponding to pattern layer neuronsi,i∈[1,v]}。
2. The method of claim 1, wherein the test platform is comprised of a vector signal generator, a vector signal analyzer, an attenuator, and a PC device.
3. The method of claim 1, wherein the test platform measures and records an input digital baseband signal from a signal source, such as a vector signal generator, input to the transmitter, a carrier frequency of the transmitter, and a digital baseband signal obtained by demodulating an output signal of the transmitter with a receiver, such as a vector signal analyzer.
4. The method of claim 1, wherein the set of dynamic nonlinear fingerprint sample data is a set consisting of an amplitude fingerprint set, a frequency fingerprint set, and a phase fingerprint.
5. The method of claim 1, wherein the frequency segmentation technique classifies the operating frequency bands according to transmitter frequency-nonlinearity, determines the magnitude and step of frequency fingerprint points, and controls the number and bandwidth of neurons in the pattern layer of the dynamic multi-core bandwidth generalized regression neural network.
6. The method of claim 1, wherein the amplitude non-uniform quantization technique performs a segmentation process on the amplitude of the input signal according to a signal amplitude gain compression ratio, determines the number and the step of the amplitude fingerprint points, and controls the number and the bandwidth of the neurons in the pattern layer of the dynamic multi-kernel bandwidth generalized regression neural network.
7. The method of claim 1, wherein the optimization algorithm is a gradient descent method, a newton method, a particle swarm algorithm, a genetic algorithm, or an annealing algorithm.
8. The method of claim 1, wherein the dynamic multi-core bandwidth generalized regression neural network model adopts an incremental learning mode, improves model accuracy by inputting new training samples, and can track changes of power amplifier nonlinearity, modify parameters, and improve model robustness.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810648237.9A CN108768550B (en) | 2018-06-21 | 2018-06-21 | Wide-band transmitter nonlinear modeling method based on dynamic multi-core bandwidth generalized regression neural network algorithm |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810648237.9A CN108768550B (en) | 2018-06-21 | 2018-06-21 | Wide-band transmitter nonlinear modeling method based on dynamic multi-core bandwidth generalized regression neural network algorithm |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108768550A CN108768550A (en) | 2018-11-06 |
CN108768550B true CN108768550B (en) | 2021-07-06 |
Family
ID=63976468
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810648237.9A Active CN108768550B (en) | 2018-06-21 | 2018-06-21 | Wide-band transmitter nonlinear modeling method based on dynamic multi-core bandwidth generalized regression neural network algorithm |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108768550B (en) |
Families Citing this family (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109657390A (en) * | 2018-12-28 | 2019-04-19 | 中国电子科技集团公司第二十九研究所 | A kind of technique IP statistical modeling method in radio frequency Integrated manufacture |
CN110035425B (en) * | 2019-04-04 | 2021-10-01 | 中国科学技术大学 | Physical fingerprint extraction method for wireless equipment based on wireless network card |
CN112231986B (en) * | 2020-11-04 | 2023-04-07 | 中国电子科技集团公司第二十九研究所 | Numerical control attenuator modeling method |
CN113553771B (en) * | 2021-07-30 | 2024-02-20 | 海宁利伊电子科技有限公司 | Dynamic X parameter accounting method based on RNN network |
CN115327373B (en) * | 2022-04-20 | 2023-04-18 | 岱特智能科技(上海)有限公司 | Hemodialysis equipment fault diagnosis method based on BP neural network and storage medium |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2005032189A1 (en) * | 2003-09-26 | 2005-04-07 | Universite Du Quebec En Abitibi-Temiscamingue (Uqat) | Method and system for indoor geolocation using an impulse response fingerprinting technique |
CN102437822A (en) * | 2011-11-30 | 2012-05-02 | 上海瑞和安琦通信科技有限公司 | Self-adaptive digital pre-distortion linear system of radio frequency power amplifier |
CN102983819A (en) * | 2012-11-08 | 2013-03-20 | 南京航空航天大学 | Imitating method of power amplifier and imitating device of power amplifier |
CN106130661A (en) * | 2016-06-13 | 2016-11-16 | 杭州电子科技大学 | Broadband wireless transmitter recognition methods based on Hammerstein Wiener model |
CN107104746A (en) * | 2017-04-26 | 2017-08-29 | 中央军委装备发展部第六十三研究所 | Frequency hopping radio set sender nonlinear characteristic modeling method |
-
2018
- 2018-06-21 CN CN201810648237.9A patent/CN108768550B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2005032189A1 (en) * | 2003-09-26 | 2005-04-07 | Universite Du Quebec En Abitibi-Temiscamingue (Uqat) | Method and system for indoor geolocation using an impulse response fingerprinting technique |
CN102437822A (en) * | 2011-11-30 | 2012-05-02 | 上海瑞和安琦通信科技有限公司 | Self-adaptive digital pre-distortion linear system of radio frequency power amplifier |
CN102983819A (en) * | 2012-11-08 | 2013-03-20 | 南京航空航天大学 | Imitating method of power amplifier and imitating device of power amplifier |
CN106130661A (en) * | 2016-06-13 | 2016-11-16 | 杭州电子科技大学 | Broadband wireless transmitter recognition methods based on Hammerstein Wiener model |
CN107104746A (en) * | 2017-04-26 | 2017-08-29 | 中央军委装备发展部第六十三研究所 | Frequency hopping radio set sender nonlinear characteristic modeling method |
Non-Patent Citations (1)
Title |
---|
广义记忆型神经网络射频功放数字预失真器;尹思源等;《微波学报》;20180430;第34卷(第2期);第1、2、3.1节,图1,图3 * |
Also Published As
Publication number | Publication date |
---|---|
CN108768550A (en) | 2018-11-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108768550B (en) | Wide-band transmitter nonlinear modeling method based on dynamic multi-core bandwidth generalized regression neural network algorithm | |
CN112640299A (en) | Digital predistortion under varying operating conditions | |
CN113809992A (en) | Machine learning based digital predistortion for power amplifiers | |
Li et al. | Wireless transmitter identification based on device imperfections | |
CN101853661A (en) | Noise spectrum estimation and voice mobility detection method based on unsupervised learning | |
CN107104746B (en) | Nonlinear characteristic modeling method for frequency hopping radio transmitter | |
US6892155B2 (en) | Method for the rapid estimation of figures of merit for multiple devices based on nonlinear modeling | |
WO2006069846A1 (en) | Improved blind snr estimation | |
CN113033457B (en) | Soft fault recognition model training and soft fault recognition method and device | |
CN112733477B (en) | GaN HEMT modeling method based on error function precision compensation | |
CN116760491A (en) | Signal-to-noise ratio estimation method based on deep learning | |
Wisell et al. | Behavioral power amplifier modeling using the LASSO | |
CN110716167A (en) | Amplitude frequency sweep self-adaptive resolution calibration method and system for arbitrary waveform generator | |
CN114584230B (en) | Predictive channel modeling method based on countermeasure network and long-term and short-term memory network | |
US20230136220A1 (en) | Quantifying Signal Purity by means of Machine Learning | |
Xie et al. | Data-driven RF transmit power calibration for wireless communication systems | |
CN112001115B (en) | Soft measurement modeling method of semi-supervised dynamic soft measurement network | |
RV et al. | Optimization of digital predistortion models for RF power amplifiers using a modified differential evolution algorithm | |
CN115378776A (en) | MFSK modulation identification method based on cyclic spectrum parameters | |
Makled et al. | Detection and identification of mobile network signals | |
CN108268700A (en) | A kind of RF power amplification temperature characterisitic modeling method based on BPNN | |
Yin et al. | Pattern recognition of RF power amplifier behaviors with multilayer perceptron | |
CN116596061B (en) | Power amplifier harmonic prediction method based on transfer learning | |
CN116595368B (en) | Nonlinear modeling-based power amplifier harmonic prediction method | |
Li et al. | Complex radial basis function networks trained by QR‐decomposition recursive least square algorithms applied in behavioral modeling of nonlinear power amplifiers |
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 |