5G radio frequency signal amplification method and system

Abstract

The invention relates to the technical field of radio-frequency signal processing, and discloses a 5G radio-frequency signal amplification method and system, and the method comprises the steps: decomposing an obtained radio-frequency signal into a high-frequency sub-band and a low-frequency sub-band through a multilayer wavelet packet, judging whether the radio-frequency signal is sparse in a wavelet packet domain, and carrying out the sparse reconstruction of an LI norm minimization problem if the radio-frequency signal is sparse; obtaining feature data of each sub-band, extracting a nonlinear transfer function of each sub-band through a CNN-GRU network, and carrying out transfer learning pre-training and dynamic adaptation updating; on the basis of dynamic adaptation update ending, layered memory polynomial calculation and layered model joint output construction, layered model parameter optimization is converted into a least square problem to obtain an error compensation function, and resource allocation is dynamically adjusted according to the peak-to-average power ratio of an input signal; and based on power amplifier state data, holographic sensing is performed on nonlinear parameters of the device through a fusion model, and dynamic changes of the device are compensated through multi-stage calibration and calibration delay is controlled.

Description

Technical Field

[0001] The present invention relates to the field of radio frequency signal processing technology, and in particular to a 5G radio frequency signal amplification method and system. Background Art

[0002] In the 5G era, the demand for communication transmission rates and spectrum resources is rapidly increasing. To achieve high-speed data transmission, low-latency communication, and large-scale device connectivity, 5G networks require stronger signal strength and better signal quality, placing higher demands on RF signal amplification. For example, 5G's high-density networking and full-spectrum access technology have led to an increase in the number of base stations and a further expansion of the spectrum, placing greater demands on the number and performance of terminal RF devices. In wireless communication transmission systems, RF power amplifiers are core functional modules. Their function is to amplify weak RF signals to sufficient power to be fed to the antenna for transmission, achieving effective wireless transmission. Their performance indicators, such as output power, efficiency, and linearity, directly determine the signal coverage, transmission quality, and energy efficiency of communication equipment.

[0003] In the existing technology, the efficiency of the Doherty architecture drops sharply at high back-off. At 39GHz, the power added efficiency of the Doherty architecture drops sharply at 6dB back-off, which does not meet high modulation requirements, such as 256QAM. The traditional Doherty structure has difficulty in impedance matching and aggravated nonlinear characteristics of transistors in the millimeter wave band, resulting in an uneven efficiency curve and high signal distortion. Therefore, the architecture needs to be improved, such as multi-level load modulation or dynamic bias technology. The broadband memory effect requires a high-order model for DPD, which is computationally complex. The existing DPD requires a high-order nonlinear model, which is too computationally intensive. The existing technology cannot reduce the computational complexity, resulting in high computational delay. It is necessary to combine new algorithms to reduce the complexity, such as machine learning or memory effect decomposition methods, or use frequency domain processing for block management.

[0004] In view of this, it is necessary to provide a 5G RF signal amplification method and system. Summary of the Invention

[0005] The object of the present invention is to provide a 5G radio frequency signal amplification method and system. To solve the above-mentioned problems in the prior art, the present invention is implemented through the following technical solutions:

[0006] In a first aspect, an embodiment of the present invention provides a 5G radio frequency signal amplification method, which specifically includes the following steps:

[0007] Step 1: Decompose the acquired RF signal into high-frequency sub-bands and low-frequency sub-bands using multi-layer wavelet packets, and determine whether the RF signal is sparse in the wavelet packet domain. If it is sparse, perform sparse reconstruction by minimizing the LI norm problem.

[0008] Step 2: Obtain the feature data of each sub-band, extract the nonlinear transfer function of each sub-band through a lightweight CNN-GRU network, and perform transfer learning pre-training and dynamic adaptive update;

[0009] Step 3: After the dynamic adaptive update is completed, the hierarchical memory polynomial is calculated, and the joint output of the hierarchical model is constructed. The hierarchical model parameter optimization is converted into a least squares problem to obtain the error compensation function, and resource allocation is dynamically adjusted according to the peak-to-average power ratio of the input signal;

[0010] Step 4: Collect power amplifier status data through multi-dimensional sensors, perform holographic perception of device nonlinear parameters through fusion models, compensate for device dynamic changes and control calibration delay through multi-level calibration.

[0011] Furthermore, the multi-layer wavelet packet decomposition method is:

[0012] Based on the normalization of the signal frequency signal, the DC offset and the RF carrier are eliminated, and finally the baseband equivalent signal x[n] is obtained, where n is the sampling point index;

[0013] The 800MHz bandwidth RF signal is decomposed using m-layer wavelet packets to obtain Sub-band, each layer decomposes the signal into high-frequency and low-frequency sub-bands through the orthogonal filter bank Daubechies wavelet, and finally obtains indivual MHz bandwidth sub-band signal ,in [0, 800MHz];

[0014] The filter bank of the first layer decomposition is a low-pass filter and high-pass filtering , the signal decomposition recursive formula is:

[0015] ;

[0016] Where l is the number of decomposition layers, k is the subband index, represents the convolution operation, Indicates the subband signal with the sequence number 2k decomposed at the lth layer, Indicates the subband signal numbered 2k+1 when decomposed at the lth layer;

[0017] Furthermore, the method for determining whether the radio frequency signal is sparse in the wavelet packet domain is:

[0018] If the baseband equivalent signal x[n] is represented by the coefficient vector in the wavelet packet basis Ψ In , there are only K non-zero elements, K is much smaller than N, N is the length of the baseband equivalent signal, which means that the baseband equivalent signal is K-sparse under the wavelet packet basis Ψ;

[0019] Draw the absolute value histogram of the wavelet packet coefficients to observe the coefficient distribution shape;

[0020] If the histogram shows a peak + long tail feature, with a few large coefficients and most small coefficients close to zero, then the sparsity is significant;

[0021] Furthermore, the sparse reconstruction method is:

[0022] Based on the judgment, the baseband equivalent signal is sparse in the wavelet packet domain, that is, x[n]=Ψα, where Ψ is the wavelet packet basis matrix and α is the sparse coefficient;

[0023] Using a partial Fourier matrix For signal compression sampling, the formula y= = =A Get the observed value y, where A= is the perception matrix;

[0024] Recover the sparse coefficients by solving the L1 norm minimization problem , through the formula

[0025] ;

[0026] Get the recovery sparse coefficient ;

[0027] in is the regularization parameter, is the square error between the observed value and the reconstructed signal, is the L1 norm of the sparse coefficient, The value of the variable when the function reaches its minimum value;

[0028] Furthermore, the method for extracting the nonlinear transfer function of each sub-band is:

[0029] Build a lightweight CNN-GRU network architecture,

[0030] The fully connected layer is mapped to output dimension 2, including: AM-AM gain coefficient and AM-PM phase coefficient ;

[0031] Through small convolution kernels, low number of channels and single-layer recurrent structure;

[0032] By simulating the 5GNR signal in MATLAB and using the memory polynomial model, we can generate Group labeled data ;

[0033] Furthermore, the dynamic adaptive update method is:

[0034] The mean square error is calculated based on the loss function using the formula , where is the batch size, is the predicted value of the gain coefficient, is the true value of the gain coefficient, is the predicted value of the phase coefficient, is the true value of the phase coefficient;

[0035] Fine-tune the lightweight CNN-GRU network architecture based on local real-time data from each RF front end;

[0036] The server aggregates the global model and updates the shared parameters;

[0037] Furthermore, the method for constructing the hierarchical model joint output is:

[0038] For the short-term memory layer, the instantaneous nonlinearity is compensated by the memoryless polynomial;

[0039] A fifth-order memoryless polynomial is used to describe the instantaneous AM-AM / PM nonlinear characteristics of the amplifier, and the output signal of the nonlinear system is obtained by the formula ;

[0040] For the long-term memory layer, the thermal memory effect is compensated by the delayed tap structure;

[0041] A memory polynomial with a third-order delay tap is used to capture the slow time-varying memory effect caused by thermal effects, and the output signal with memory effect is obtained through the formula ;

[0042] For the cross-coupling layer, interference compensation is performed by calculating the intermodulation products between sub-bands;

[0043] The inter-subband intermodulation products are calculated by the formula ;

[0044] The total predistortion signal is obtained by superimposing the output signals of the short-term memory layer, the long-term memory layer and the cross-coupling layer. ;

[0045] Furthermore, the method for dynamically adjusting resource allocation is:

[0046] By utilizing the hardware parallel characteristics of FPGA, pipeline parallel computing is performed on the three-level sub-problems;

[0047] After the distortion error compensation is completed, the power supply voltage V and the operating frequency f of the computing unit are dynamically adjusted according to the PAPR of the input signal;

[0048] Furthermore, the method for compensating for the dynamic changes of the device is:

[0049] Through the three-level calibration strategy of coarse, fine and micro, dynamic changes of the device can be compensated in real time;

[0050] For coarse calibration, if the fluctuation of the state variable output by the Bayesian network exceeds the threshold, baseline calibration is initiated. Based on the nonlinear parameters after sensor fusion, the initial predistortion parameters of the stage 2 hierarchical model are updated:

[0051] ;

[0052] in is the smoothing factor, t is the calibration period index, is the new distortion parameter;

[0053] For fine processing, an orthogonal pilot sequence is inserted into the transmitted signal ;

[0054] The receiving end extracts the pilot signal , the pilot signal obtained Orthogonal pilot sequence Subtract the subband error and minimize the mean square error; iterate the gradient descent method 5 times;

[0055] For micro-calibration, the decoding error rate (BER) and phase error estimate of the forward error correction code are obtained from the baseband processing unit. , is the signal after error correction, is the conjugate of the signal being sent;

[0056] By phase adjustment formula: Get the phase adjustment component ,in is the micro-calibration phase adjustment coefficient.

[0057] In a second aspect, an embodiment of the present invention provides a 5G radio frequency signal amplification system, which specifically includes the following modules:

[0058] Sparse reconstruction module: decomposes the acquired RF signal into high-frequency sub-bands and low-frequency sub-bands through multi-layer wavelet packets, determines whether the RF signal is sparse in the wavelet packet domain, and if so, performs sparse reconstruction by minimizing the LI norm problem;

[0059] Dynamic update module: obtains the feature data of each sub-band, extracts the nonlinear transfer function of each sub-band through a lightweight CNN-GRU network, and performs transfer learning pre-training and dynamic adaptive update;

[0060] Compensation allocation module: Based on the completion of dynamic adaptive update, it calculates the hierarchical memory polynomial, constructs the joint output of the hierarchical model, transforms the hierarchical model parameter optimization into a least squares problem to obtain the error compensation function, and dynamically adjusts resource allocation according to the peak-to-average power ratio of the input signal;

[0061] Control and calibration module: collects power amplifier status data through multi-dimensional sensors, performs holographic perception of device nonlinear parameters through fusion models, compensates for device dynamic changes and controls calibration delay through multi-level calibration.

[0062] Beneficial effects of the present invention:

[0063] 1. Utilize multi-layer wavelet packet decomposition to achieve frequency band segmentation of RF signals, decomposing the 800MHz bandwidth signal into multiple sub-bands. A sparse reconstruction algorithm constrained by the L1 norm effectively suppresses inter-sub-band intermodulation interference. A dynamic bandwidth adjustment mechanism automatically optimizes bandwidth allocation based on the sub-band peak-to-average ratio. A CNN-GRU network optimized for RF signal processing was designed, constructing an extremely lightweight network with less than 50k parameters, suitable for hardware integration. Combined with transfer learning and dynamic adaptive updates, it can quickly adapt to changes in device characteristics.

[0064] 2. A three-layer memory polynomial model is proposed. Each layer focuses on compensating for a specific type of nonlinear distortion to improve modeling accuracy. The parallel computing capabilities of FPGAs are utilized to achieve efficient parameter optimization and solution. The power supply voltage and operating frequency of the computing unit are dynamically adjusted based on the input signal PAPR. An intelligent power management mechanism is used to reduce energy consumption while ensuring processing performance. Multi-dimensional sensor data fusion is used to monitor the power amplifier status in real time. A Bayesian network is used to map sensor data to nonlinear parameters. A three-level calibration strategy is constructed to control latency while ensuring calibration accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0066] Figure 1 This is a flowchart of the steps of a 5G radio frequency signal amplification method and system provided in Example 1 of the present invention;

[0067] Figure 2 This is a structural diagram of a 5G radio frequency signal amplification method and system provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0068] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0069] Example 1

[0070] like Figure 1 As shown, a 5G radio frequency signal amplification method provided by an embodiment of the present invention specifically includes the following steps:

[0071] Step 1: Decompose the acquired RF signal into high-frequency sub-bands and low-frequency sub-bands using multi-layer wavelet packets, and determine whether the RF signal is sparse in the wavelet packet domain. If it is sparse, perform sparse reconstruction by minimizing the LI norm problem.

[0072] In a specific embodiment, the specific process of the multi-layer wavelet packet decomposition is:

[0073] Based on the input RF signal, obtain the 800MHz bandwidth RF signal , where n is the sampling point index, n=0,1,...,N-1, and the sampling rate satisfies the Nyquist criterion. For example, the sampling frequency ≥1.6GHz, corresponding to the number of sampling points N=1024;

[0074] Based on the acquired radio frequency signal , the transfer function of the designed first-order high-pass filter is ,in , , T is the sampling period, is the angular frequency;

[0075] The time domain recursive formula corresponding to the transfer function of the first-order high-pass filter Get the filtered output value;

[0076] It should be noted that the current filtered output value in the time domain recursive formula is composed of the weighted difference between the current input value and the filtered output value at the previous moment, which has low computational complexity and small storage requirements;

[0077] Based on the obtained filtered output value, the mean normalization formula is used Perform mean normalization to eliminate DC bias and suppress low-frequency components;

[0078] in, ;

[0079] Based on the obtained normalized RF signal, the RF signal is decomposed into a baseband complex signal using two carriers, one in phase and the other in quadrature.

[0080] The orthogonal carrier and RF signal will be obtained Multiply them to get the I / O components I(n) and Q(n), and remove the high-frequency components through the low-pass filter LPF. Calculate the retained baseband signal ;

[0081] Based on the normalization of the signal frequency signal, the DC offset and the RF carrier are eliminated, and finally the baseband equivalent signal x[n] is obtained;

[0082] The 800MHz bandwidth RF signal is decomposed using m-layer wavelet packets to obtain Sub-band, each layer decomposes the signal into high-frequency and low-frequency sub-bands through the orthogonal filter bank Daubechies wavelet, and finally obtains indivual MHz bandwidth sub-band signal ,in [0, 800MHz];

[0083] The filter bank of the first layer decomposition is a low-pass filter and high-pass filtering , the signal decomposition recursive formula is:

[0084] ;

[0085] Where l is the number of decomposition layers, k is the subband index, represents the convolution operation, Indicates the subband signal with the sequence number 2k decomposed at the lth layer, Indicates the subband signal numbered 2k+1 when decomposed at the lth layer;

[0086] Introducing a sparse reconstruction algorithm with L1 norm constraints to suppress inter-subband intermodulation interference;

[0087] If the baseband equivalent signal x[n] is represented by the coefficient vector in the wavelet packet basis Ψ In , there are only K non-zero elements, K is much smaller than N, N is the length of the baseband equivalent signal, which means that the baseband equivalent signal is K-sparse under the wavelet packet basis Ψ;

[0088] Draw the absolute value histogram of the wavelet packet coefficients to observe the coefficient distribution shape;

[0089] If the histogram shows a peak + long tail feature, with a few large coefficients and most small coefficients close to zero, then the sparsity is significant;

[0090] Compared with the non-sparse cases of uniform or Gaussian distribution, the histogram of sparse signals has a clear peak near zero;

[0091] A mean of zero or close to zero, combined with a large variance, indicates the presence of a few large coefficients and / or a majority of small coefficients.

[0092] Setting thresholds , the absolute value of the statistic is greater than The number of coefficients K, if K is much smaller than N, the signal is considered sparse;

[0093] Based on the judgment, the baseband equivalent signal is sparse in the wavelet packet domain, that is, x[n]=Ψα, where Ψ is the wavelet packet basis matrix, α is the sparse coefficient, and the proportion of non-zero elements s is much smaller than N;

[0094] Using a partial Fourier matrix For signal compression sampling, the formula y= = =A Get the observed value y, where A= is the perception matrix, satisfying the finite isometry property RIP;

[0095] Recover the sparse coefficients by solving the L1 norm minimization problem , through the formula

[0096] ;

[0097] Get the recovery sparse coefficient ;

[0098] in is the regularization parameter, balancing the reconstruction error and sparsity, is the square error between the observation value and the reconstructed signal, which measures the reconstruction accuracy. is the L1 norm of the sparse coefficients, which promotes sparse solutions and suppresses the non-sparse components corresponding to intermodulation interference. The value of the variable when the function reaches its minimum value;

[0099] The peak-to-average power ratio of the RF signal in each sub-band is calculated based on the formula

[0100] ;

[0101] Calculate the sub-band peak-to-average ratio of each sub-band signal ;

[0102] in is the subband signal, is the sub-band peak power, is the average power expectation;

[0103] The obtained sub-band peak-to-average ratio Compare with the set peak-to-average ratio threshold Thresh and dynamically allocate sub-band bandwidth;

[0104] Based on any sub-band, if the sub-band peak-to-average ratio If the peak-to-average ratio is greater than the set peak-to-average ratio threshold Thresh, the sub-band bandwidth is narrowed by truncating the frequency domain response through the filter;

[0105] For example, a 4-layer wavelet packet is used to decompose the 800MHz bandwidth radio frequency signal to obtain 16 sub-bands, and finally obtain For a 50MHz bandwidth sub-band signal, for any sub-band, if the sub-band peak-to-average ratio If the peak-to-average ratio is greater than the set peak-to-average ratio threshold Thresh, the sub-band bandwidth is narrowed from 50MHz to 30MHz by truncating the frequency domain response through the filter, reducing the nonlinear impact of high PAPR signals on the amplifier;

[0106] It should be noted that the ASIC chip integrates the Sparse Fourier Transform (SFT) module to complete 1024-point signal decomposition in a single cycle, and the multi-channel FIR filter bank realizes real-time processing of wavelet packet decomposition and compressed sensing sampling.

[0107] Step 2: Obtain the feature data of each sub-band, extract the nonlinear transfer function of each sub-band through a lightweight CNN-GRU network, and perform transfer learning pre-training and dynamic adaptive update;

[0108] It should be noted that the feature data includes: I / Q baseband signal, envelope signal and input tensor;

[0109] Based on the acquired RF signal, the following features are extracted for each sub-band signal:

[0110] I / Q baseband signal: ;

[0111] Envelope signal: ;

[0112] Based on the I / Q baseband signal and envelope signal, the input tensor [ ], the dimension is T×3, T is the time series length, for example 1024 points;

[0113] Build a lightweight CNN-GRU network architecture,

[0114] For the CNN layer, two convolutional layers are set, and a 3×1 convolution kernel is used to adapt to the one-dimensional characteristics of the time series signal. The number of channels in each layer is 8, and the activation function is ReLU to reduce the number of parameters;

[0115] Set the maximum pooling layer with a pooling kernel of 2×1 to reduce the dimension and output dimension to T / 2×8;

[0116] For the GRU layer, set 1 layer of GRU units, the number of hidden units is 16, capture the temporal dependency, and the output dimension is T / 2×16;

[0117] The fully connected layer is mapped to output dimension 2, including: AM-AM gain coefficient and AM-PM phase coefficient ;

[0118] Through small convolution kernel, low number of channels and single-layer loop structure, the total number of parameters is less than 50k;

[0119] By simulating the 5GNR signal in MATLAB and using the memory polynomial model, we can generate Group labeled data ;

[0120] Based on the loss function through the formula

[0121] ;

[0122] Calculate the mean square error , where is the batch size, is the predicted value of the gain coefficient, is the true value of the gain coefficient, is the predicted value of the phase coefficient, is the true value of the phase coefficient;

[0123] Perform multiple rounds of pre-training based on the Adam optimizer. If the mean square error calculated by the loss function is less than the preset mean square error threshold and the mean square error change rate is less than the preset change rate threshold, pre-training is stopped.

[0124] Based on each RF front end, the lightweight CNN-GRU network architecture is fine-tuned for 1-2 epochs using local real-time data, and only the gradients are uploaded to the central server.

[0125] The server aggregates the global model, updates shared parameters, and suppresses nonlinear characteristic changes caused by device aging and temperature drift;

[0126] Step 3: After the dynamic adaptive update is completed, the hierarchical memory polynomial is calculated, and the joint output of the hierarchical model is constructed. The hierarchical model parameter optimization is converted into a least squares problem to obtain the error compensation function, and resource allocation is dynamically adjusted according to the peak-to-average power ratio of the input signal;

[0127] It should be noted that the layered model parameters include: nonlinear system output signal, output signal with memory effect and inter-subband intermodulation products;

[0128] In a specific implementation, the method for calculating the hierarchical memory polynomial is:

[0129] For the short-term memory layer, the instantaneous nonlinearity is compensated by the memoryless polynomial;

[0130] Specifically, a fifth-order memoryless polynomial is used to describe the instantaneous AM-AM / PM nonlinear characteristics of the amplifier, using the formula

[0131] ;

[0132] Get the nonlinear system output signal ;

[0133] in are odd-order polynomial coefficients, It is a complex number containing amplitude and phase information, k=1 is a linear term, and k≥3 is a nonlinear term;

[0134] It should be noted that odd-order terms are taken to simplify the calculation, and even-order terms are eliminated by DC bias to reduce the complexity of the model;

[0135] For the long-term memory layer, the thermal memory effect is compensated by the delayed tap structure;

[0136] Specifically, a memory polynomial with a third-order delay tap is used to capture the slow time-varying memory effect caused by thermal effects.

[0137] ;

[0138] Get an output signal with memory effect ;

[0139] Where m=0,1,2 is the delay tap index, corresponding to the delay time =m* , is the sampling period, such as =1ns, the delay span is 0-2ns, corresponding to the actual thermal effect time span of 1μs-1ms, which is compressed to the digital domain through oversampling; are the coefficients of the polynomial with memory, describing the effect of the delayed signal on the current output; the order k=3 balances the modeling accuracy of the memory effect and the amount of computation;

[0140] For the cross-coupling layer, interference compensation is performed by calculating the intermodulation products between sub-bands;

[0141] Specifically, a 16×16 complex coefficient matrix C=[ ], compensate for the intermodulation interference between sub-band i and sub-band j, such as third-order intermodulation IM3 and fifth-order intermodulation IM5, by the formula

[0142] ;

[0143] Calculate the inter-subband intermodulation products ,in is the baseband signal of subband i, is the conjugate of the baseband signal of subband j;

[0144] It should be noted that Characterize the intermodulation coupling strength between sub-bands i and j, estimated by frequency domain cross-correlation analysis;

[0145] The total predistortion signal is obtained by superimposing the output signals of the short-term memory layer, the long-term memory layer and the cross-coupling layer. ;

[0146] The optimization of the layered model parameters is transformed into a least squares problem, and the error between the predistorted signal and the ideal linear signal is minimized through the error compensation function.

[0147] ;

[0148] Get the minimized error ,in is the parameter to be optimized, d is the ideal distortion-free signal;

[0149] The global optimization problem is divided into short-term, long-term, and cross-layer sub-problems by introducing auxiliary variables. Implement parallel solving:

[0150] ;

[0151] in is the penalty parameter, t is the number of iterations, and convergence is achieved by alternately updating the sub-layer parameters and the global variable z;

[0152] By utilizing the hardware parallel characteristics of FPGA, pipeline parallel computing is performed on the three-level sub-problems;

[0153] After the distortion error compensation is completed, the power supply voltage V and the operating frequency f of the computing unit are dynamically adjusted according to the PAPR of the input signal;

[0154] Specifically, the PAPR of the input signal is compared with a preset peak-to-average power ratio threshold;

[0155] If the PAPR of the input signal is greater than or equal to the peak-to-average power ratio threshold PAPR1, the system maintains full frequency and full voltage operation to ensure real-time performance.

[0156] If the PAPR of the input signal is less than the peak-to-average power ratio threshold PAPR1 and greater than the peak-to-average power ratio threshold PAPR2, the supply voltage V and the operating frequency f remain unchanged.

[0157] If the PAPR of the input signal is less than or equal to the peak-to-average power ratio threshold PAPR2, the peak-to-average power ratio threshold PAPR2 is subtracted from the PAPR of the input signal and then the result is compared with the peak-to-average power ratio threshold PAPR2 to obtain a power reduction ratio JH, based on which the power consumption is reduced to 1-JH times the original value;

[0158] Step 4: Collect power amplifier status data through multi-dimensional sensors, perform holographic perception of device nonlinear parameters through fusion models, and compensate for device dynamic changes and control calibration delay through multi-level calibration;

[0159] In a specific embodiment, the specific method of holographic perception is:

[0160] Real-time data collection of power amplifier status is achieved through thermocouples, current sensors, and phase detectors, and holographic perception of device nonlinear parameters is achieved using Bayesian networks.

[0161] Power amplifier status data includes: power amplifier status temperature, current fluctuation signal and output signal phase noise;

[0162] 16 micro thermocouples are integrated on the surface of the power amplifier chip, distributed in key areas such as the drain, gate, and heat sink, to monitor the temperature in real time. ;

[0163] A current sensor based on magnetoresistance effect is connected in series with the gate and drain circuits to capture the current fluctuation signal and obtain the gate current. and drain current ;

[0164] Compare the input / output signal phases through the phase-locked loop (PLL) and measure the output signal phase noise ;

[0165] A three-layer Bayesian network is constructed based on the obtained power amplifier status data, and the input layer is the power amplifier status data , the hidden layer is the power amplifier state variables, which include but are not limited to: output power, gain compression point and phase error;

[0166] The output layer is a nonlinear parameter ;

[0167] Based on the effect of temperature on gain, the formula Get the temperature influence parameters;

[0168] in , is the temperature coefficient, is the reference temperature;

[0169] Based on the influence of current fluctuation on the memory effect coefficient, the formula Get the parameters affecting current fluctuation;

[0170] in is the delay-related weight factor, describing the impact of current change on memory effect;

[0171] Variational inference is used to approximate the posterior distribution, and the parameters are estimated by maximizing the log-marginal likelihood, thus mapping the sensor data into the nonlinear parameter space.

[0172] Holographic perception of device nonlinear parameters through fusion model;

[0173] Exemplarily, the fusion output includes: temperature sensitive parameters, short-term polynomial coefficients , the nonlinear distortion increases by 0.5dB for every 10℃ increase in temperature; current related parameters, for long term memory coefficient , when the drain current fluctuation is >5%, the memory effect is enhanced by 30%; the phase noise parameter, for the cross-coupling matrix When the phase noise is >1dBc / Hz, the number of intermodulation subbands increases by 2;

[0174] Through the three-level calibration strategy of coarse, fine and micro, dynamic changes of devices are compensated in real time, signal accuracy is improved and calibration delay is controlled;

[0175] Specifically, for coarse calibration, if the fluctuation of the state variable output by the Bayesian network exceeds the threshold, baseline calibration is initiated, and the initial predistortion parameters of the hierarchical model in stage 2 are updated based on the nonlinear parameters after sensor fusion:

[0176] ;

[0177] in is the smoothing factor, which ranges from 0.1 to 0.3 to avoid sudden interference, and t is the calibration period index. is the new distortion parameter, which is calculated based on the nonlinear parameters after sensor fusion;

[0178] For fine processing, an orthogonal pilot sequence is inserted into the transmitted signal , the power is 3dB higher than the average power of the signal, and the frequency domain position is located at the sub-band center frequency;

[0179] The receiving end extracts the pilot signal , the pilot signal obtained Orthogonal pilot sequence Subtraction calculation is performed to obtain the subband error and minimize the mean square error; the gradient descent method is iterated 5 times to reduce the FPGA pipeline processing delay;

[0180] For micro-calibration, the decoding error rate (BER) and phase error estimate of the forward error correction code are obtained from the baseband processing unit. , is the signal after error correction, is the conjugate of the signal being sent;

[0181] By phase adjustment formula: Get the phase adjustment component ,in To fine-tune the phase adjustment coefficient, a value of 0.5 is used to avoid overcorrection. The register pipeline is used to achieve phase fine-tuning in each cycle, which directly acts on the phase component of the DPD output;

[0182] The pipeline architecture based on three-stage calibration compensates for device dynamic changes and controls calibration delay through multi-stage calibration;

[0183] The technical solution of the embodiments of the present invention is as follows: multi-layer wavelet packet decomposition is used to achieve frequency band segmentation of RF signals, decomposing the 800 MHz bandwidth signal into multiple subbands. An L1-norm constrained sparse reconstruction algorithm effectively suppresses inter-subband intermodulation interference. A dynamic bandwidth adjustment mechanism automatically optimizes bandwidth allocation based on the subband peak-to-average ratio (PAR) is used. A CNN-GRU network optimized for RF signal processing is designed to construct an extremely lightweight network with less than 50k parameters, suitable for hardware integration. Transfer learning and dynamic adaptive updates are combined to quickly adapt to changes in device characteristics. A three-layer memory polynomial model is proposed, with each layer focusing on compensating for a specific type of nonlinear distortion to improve modeling accuracy. The parallel computing capabilities of the FPGA are utilized to achieve efficient parameter optimization. The supply voltage and operating frequency of the computing unit are dynamically adjusted based on the input signal PAPR. An intelligent power management mechanism is used to reduce energy consumption while ensuring processing performance. Multi-dimensional sensor data fusion is used to monitor the power amplifier status in real time. A Bayesian network is used to map sensor data to nonlinear parameters. A three-level calibration strategy is constructed to control latency while ensuring calibration accuracy.

[0184] Example 2

[0185] like Figure 2 As shown, a 5G radio frequency signal amplification system provided by an embodiment of the present invention specifically includes the following modules:

[0186] Sparse reconstruction module: decomposes the acquired RF signal into high-frequency sub-bands and low-frequency sub-bands through multi-layer wavelet packets, determines whether the RF signal is sparse in the wavelet packet domain, and if so, performs sparse reconstruction by minimizing the LI norm problem;

[0187] Dynamic update module: obtains the feature data of each sub-band, extracts the nonlinear transfer function of each sub-band through a lightweight CNN-GRU network, and performs transfer learning pre-training and dynamic adaptive update;

[0188] Compensation allocation module: Based on the completion of dynamic adaptive update, it calculates the hierarchical memory polynomial, constructs the joint output of the hierarchical model, transforms the hierarchical model parameter optimization into a least squares problem to obtain the error compensation function, and dynamically adjusts resource allocation according to the peak-to-average power ratio of the input signal;

[0189] Control and calibration module: collects power amplifier status data through multi-dimensional sensors, performs holographic perception of device nonlinear parameters through fusion models, compensates for device dynamic changes and controls calibration delay through multi-level calibration.

[0190] An embodiment of the present invention is described in detail above, but the content described is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention; the above formulas are all dimensionless and numerical calculations, and the formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field based on actual conditions and historical experience, and can be adjusted according to actual conditions; the above description is only a preferred embodiment of the present invention and is not used to limit the present invention. All equal changes and improvements made according to the scope of application of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A 5G radio frequency signal amplification method, characterized in that: The following steps are involved: Step 1: Decompose the acquired RF signal into high-frequency sub-bands and low-frequency sub-bands using multi-layer wavelet packets, and determine whether the RF signal is sparse in the wavelet packet domain. If it is sparse, perform sparse reconstruction by minimizing the LI norm problem. Step 2: Obtain the feature data of each sub-band, extract the nonlinear transfer function of each sub-band through a lightweight CNN-GRU network, and perform transfer learning pre-training and dynamic adaptive update; Step 3: After the dynamic adaptive update is completed, the hierarchical memory polynomial is calculated, and the joint output of the hierarchical model is constructed. The hierarchical model parameter optimization is converted into a least squares problem to obtain the error compensation function, and resource allocation is dynamically adjusted according to the peak-to-average power ratio of the input signal; Step 4: Collect power amplifier status data through multi-dimensional sensors, perform holographic perception of device nonlinear parameters through fusion models, compensate for device dynamic changes and control calibration delay through multi-level calibration.

2. A 5G radio frequency signal amplification method according to claim 1, characterized in that: The method of multi-layer wavelet packet decomposition is: Based on the normalization of the signal frequency signal, the DC offset and the RF carrier are eliminated, and finally the baseband equivalent signal x[n] is obtained, where n is the sampling point index; The 800MHz bandwidth RF signal is decomposed using m-layer wavelet packets to obtain Sub-band, each layer decomposes the signal into high-frequency and low-frequency sub-bands through the orthogonal filter bank Daubechies wavelet, and finally obtains indivual MHz bandwidth sub-band signal ,in [0, 800MHz]; The filter bank of the first layer decomposition is a low-pass filter and high-pass filtering , the signal decomposition recursive formula is: ； Where l is the number of decomposition layers, k is the subband index, represents the convolution operation, Indicates the subband signal with the sequence number 2k decomposed at the lth layer, It represents the subband signal numbered 2k+1 when decomposed at the lth layer.

3. A 5G radio frequency signal amplification method according to claim 1, characterized in that: The method for determining whether the radio frequency signal is sparse in the wavelet packet domain is: If the baseband equivalent signal x[n] is represented by the coefficient vector in the wavelet packet basis Ψ In , there are only K non-zero elements, K is much smaller than N, N is the length of the baseband equivalent signal, which means that the baseband equivalent signal is K-sparse under the wavelet packet basis Ψ; Draw the absolute value histogram of the wavelet packet coefficients to observe the coefficient distribution shape; If the histogram shows a peak + long tail feature, with a few large coefficients and most small coefficients close to zero, then the sparsity is significant.

4. A 5G radio frequency signal amplification method according to claim 1, characterized in that: The sparse reconstruction method is: Based on the judgment, the baseband equivalent signal is sparse in the wavelet packet domain, that is, x[n]=Ψα, where Ψ is the wavelet packet basis matrix and α is the sparse coefficient; Using a partial Fourier matrix For signal compression sampling, the formula y= = =A Get the observed value y, where A= is the perception matrix; Recover the sparse coefficients by solving the L1 norm minimization problem , through the formula ； Get the recovery sparse coefficient ; in is the regularization parameter, is the square error between the observed value and the reconstructed signal, is the L1 norm of the sparse coefficient, The value of the variable when the function reaches its minimum value.

5. The 5G radio frequency signal amplification method according to claim 1, wherein: The method for extracting the nonlinear transfer function of each sub-band is: Build a lightweight CNN-GRU network architecture, The fully connected layer is mapped to output dimension 2, including: AM-AM gain coefficient and AM-PM phase coefficient ; Through small convolution kernels, low number of channels and single-layer recurrent structure; By simulating the 5GNR signal in MATLAB and using the memory polynomial model, we can generate Group labeled data .

6. A 5G radio frequency signal amplification method according to claim 1, characterized in that: The method of dynamic adaptive updating is: The mean square error is calculated based on the loss function using the formula , where is the batch size, is the predicted value of the gain coefficient, is the true value of the gain coefficient, is the predicted value of the phase coefficient, is the true value of the phase coefficient; Fine-tune the lightweight CNN-GRU network architecture based on local real-time data from each RF front end; The server aggregates the global model and updates the shared parameters.

7. The 5G radio frequency signal amplification method according to claim 1, wherein: The method for constructing the joint output of the hierarchical model is: For the short-term memory layer, the instantaneous nonlinearity is compensated by the memoryless polynomial; A fifth-order memoryless polynomial is used to describe the instantaneous AM-AM / PM nonlinear characteristics of the amplifier, and the output signal of the nonlinear system is obtained by the formula ; For the long-term memory layer, the thermal memory effect is compensated by the delayed tap structure; A memory polynomial with a third-order delay tap is used to capture the slow time-varying memory effect caused by thermal effects, and the output signal with memory effect is obtained through the formula ; For the cross-coupling layer, interference compensation is performed by calculating the intermodulation products between sub-bands; The inter-subband intermodulation products are calculated by the formula ; The total predistortion signal is obtained by superimposing the output signals of the short-term memory layer, the long-term memory layer and the cross-coupling layer. .

8. The 5G radio frequency signal amplification method according to claim 1, wherein: The method for dynamically adjusting resource allocation is: By utilizing the hardware parallel characteristics of FPGA, pipeline parallel computing is performed on the three-level sub-problems; After the distortion error compensation is completed, the supply voltage V and the operating frequency f of the computing unit are dynamically adjusted according to the PAPR of the input signal.

9. The 5G radio frequency signal amplification method according to claim 1, wherein: The method for compensating for dynamic changes of the device is: Through the three-level calibration strategy of coarse, fine and micro, dynamic changes of the device can be compensated in real time; For coarse calibration, if the fluctuation of the state variable output by the Bayesian network exceeds the threshold, baseline calibration is initiated. Based on the nonlinear parameters after sensor fusion, the initial predistortion parameters of the stage 2 hierarchical model are updated: ； in is the smoothing factor, t is the calibration period index, is the new distortion parameter; For fine processing, an orthogonal pilot sequence is inserted into the transmitted signal ; The receiving end extracts the pilot signal , the pilot signal obtained Orthogonal pilot sequence Subtract the subband error and minimize the mean square error; iterate the gradient descent method 5 times; For micro-calibration, the decoding error rate (BER) and phase error estimate of the forward error correction code are obtained from the baseband processing unit. , is the signal after error correction, is the conjugate of the signal being sent; By phase adjustment formula: Get the phase adjustment component ,in is the micro-calibration phase adjustment coefficient.

10. A 5G radio frequency signal amplification system, applied to a 5G radio frequency signal amplification method according to any one of claims 1 to 9, characterized in that: include: Sparse reconstruction module: decomposes the acquired RF signal into high-frequency sub-bands and low-frequency sub-bands through multi-layer wavelet packets, determines whether the RF signal is sparse in the wavelet packet domain, and if so, performs sparse reconstruction by minimizing the LI norm problem; Dynamic update module: obtains the feature data of each sub-band, extracts the nonlinear transfer function of each sub-band through a lightweight CNN-GRU network, and performs transfer learning pre-training and dynamic adaptive update; Compensation allocation module: Based on the completion of dynamic adaptive update, it calculates the hierarchical memory polynomial, constructs the joint output of the hierarchical model, transforms the hierarchical model parameter optimization into a least squares problem to obtain the error compensation function, and dynamically adjusts resource allocation according to the peak-to-average power ratio of the input signal; Control and calibration module: collects power amplifier status data through multi-dimensional sensors, performs holographic perception of device nonlinear parameters through fusion models, compensates for device dynamic changes and controls calibration delay through multi-level calibration.

Images