CN115225041B

CN115225041B - A digital enhancement method for power amplifiers suitable for 5G signals

Info

Publication number: CN115225041B
Application number: CN202210872489.6A
Authority: CN
Inventors: 胡欣; 常旭明; 刘志军; 冀昕; 邱翊; 谢树宾; 王卫东
Original assignee: Beijing University of Posts and Telecommunications
Current assignee: Beijing University of Posts and Telecommunications
Priority date: 2022-07-20
Filing date: 2022-07-20
Publication date: 2025-07-04
Anticipated expiration: 2042-07-20
Also published as: CN115225041A

Abstract

The present invention provides a digital enhancement method for a power amplifier suitable for 5G signals, and belongs to the field of wireless communication technology. The method of the present invention sets a digital enhancement module to process the generated OFDM sequence source signal, and trims the source signal in the time domain to reduce the peak-to-average power ratio of the signal; the Volterra model and the linear time-invariant system LTI are cascaded to realize PA modeling; the short-memory GMP model and LTI are selected for the nonlinear part through a filter; a low-complexity compensation module is constructed to reduce the memory effect of the diagonal terms, and the GMP model coefficients are calculated using the input signal and feedback signal of the PA and stored in a lookup table; the final pre-distortion signal is obtained by multiplying the input signal with the corresponding coefficient of the lookup table and then summing them. The present invention can realize dynamic configuration compensation of multi-phase branches based on the lookup table in hardware, which reduces the complexity of broadband compensation while greatly improving the flexibility of compensation, and has practical high efficiency and high precision.

Description

Digital enhancement method suitable for power amplifier of 5G signal

Technical Field

The invention belongs to the technical field of wireless communication, and particularly relates to a digital enhancement method of a 5G signal for a hundred-megalevel broadband amplifier.

Background

With the development of wireless communication systems, the demand for high-throughput data transmission has led to a need for using complex modem technologies to improve spectrum utilization, for example, in the Fifth Generation (5G) mobile communication system, orthogonal frequency division multiplexing (Orthogonal Frequency Division Multiplexing, OFDM) technology is adopted. However, while improving the spectrum utilization, the signals in the system have a very high Peak-to-Average Power Ratio (PAPR) ratio, and nonlinear distortion and adjacent channel leakage can be generated when the signals pass through a radio frequency Power Amplifier (PA), so that the linearity of the system is reduced and interference is caused.

A common method is to back off the PA input in order to allow normal transmission when the signal transmission power does not meet the channel requirement, resulting in reduced power amplification efficiency. Digital predistortion compensates for the inverse distortion characteristics by preprocessing the input signal. After the signal is input into the power amplifier, the two distortions cancel each other, so that the linearization output of the power amplifier is realized. Therefore, the linearization technology of the broadband power amplifier is an important technology in the wireless communication system at present, and has important practical significance.

In order to better study the linearization technique of the broadband power amplifier, accurate modeling of the power amplifier is required. The compact memory polynomial model has good balance between performance and complexity, and is most widely applied to behavior modeling of a power amplifier. However, the large bandwidth peak-to-peak power amplifier under signal excitation presents more complex nonlinear characteristics, the traditional memory polynomial model only contains diagonal term coefficients, the modeling accuracy for representing complex memory effects is limited, and the traditional generalized memory polynomial model has a large number of coefficients after cross terms are introduced, and the high-complexity model needs more hardware resources.

In practical applications, compensation in the trunk is usually implemented by displaying Look-up tables (LUTs). Therefore, to ensure good linearity of the power amplifier under wideband signal excitation and low resource consumption of hardware, it is necessary to implement LUT-based low complexity digital enhancement techniques under high precision compensation for PA nonlinearities.

Disclosure of Invention

Aiming at the problem of large occupied resources in a compensation scheme which is commonly used in a broadband communication system and is realized based on a lookup table, in order to reduce the complexity of broadband predistortion compensation and simultaneously maintain the good linearity and efficiency of the power amplifier, the invention provides a digital enhancement method applicable to a 5G signal and facing to a hundred-megalevel broadband amplifier.

The invention provides a digital enhancement method of a power amplifier suitable for a 5G signal, which comprises the following steps:

A signal source module generates an OFDM signal x _n; the digital enhancement module is used for setting the crest factor clipping rate, adjusting the peak-to-average power ratio of the OFDM signal and generating clipping signal v _k, interpolating v _k to generate multiphase signal transmission, and sampling feedback signals through a power amplifier;

the structure of the equivalent baseband output of the PA model is composed of two parts, namely a nonlinear system related to the current signal and the previous signal and a linear time-invariant system after low-pass filtering and downsampling;

And (III) constructing a compensation module in the digital enhancement module, which comprises the following steps:

(1) GMP (generalized memory polynomial) model representation is adopted for a nonlinear system, diagonal terms, lag terms and lead term basis function matrixes of the model are respectively phi _a、Φ_b、Φ_c, and corresponding model coefficient matrixes are respectively W _a、W_b、W_c;W_a diagonal term model coefficients a _kl,W_b lag term model coefficients b _klm,W_c lead term model coefficients c _klm;

(2) Taking a low memory part of a diagonal basis function matrix phi _a to form a matrix phi _a ', and forming a new basis function matrix psi, psi= [ psi _aΦ_bΦ_c ] by a convolution output matrix psi _a of phi _a' and a convolution kernel H and phi _b、Φ_c;

(3) Calculating a model coefficient matrix W, W= (ψ ^ΗΨ)^-1Ψ^ΗX_n; wherein W= [ W _a W_b W_c]^T; the upper corner mark T represents transposition, H represents a conjugate matrix, -1 represents an inverse matrix, and X _n represents an input multiphase signal;

(4) Building a lookup table of model coefficients and realizing the lookup table in hardware, wherein the lookup table comprises the following steps:

The output of the GMP model is expressed as follows:

Wherein K, L _a represents the nonlinear order and the memory depth of the diagonal terms respectively, M _b、L_b represents the lag depth and the memory depth of the lag term respectively, M _c、L_c represents the lead depth and the memory depth of the lead term respectively, |x (n-l) | represents the envelope of the input signal x (n-l) | represents the convolution operation, h _k (n) represents the unit impulse response of the (k+1) th filter in the configured FIR filter bank, and the model coefficients in the above formula Calculating model coefficients in advance to establish a lookup table;

the input signal x (n-l) is multiplied by the model coefficient obtained by the corresponding lookup table index and summed to obtain the final predistortion signal y (n).

The method has the advantages and positive effects that the digital enhancement method for the hundred-megalevel broadband amplifier is realized, the multi-phase cascade structure is adopted, the memory depth of diagonal terms in the existing power amplifier model is effectively reduced through cascading a short memory nonlinear system and a long memory LTI (linear time invariant system), the dynamic configuration compensation of multi-phase branches based on a lookup table can be realized in hardware, and the compensation flexibility is greatly improved while the broadband compensation complexity is reduced. The method can reduce the hardware resources required by realizing wideband power amplification, has higher configuration flexibility, maintains good linearity and efficiency of the power amplifier while reducing the realization cost in actual engineering, and has practical, efficient and high precision.

Drawings

FIG. 1 is a flow chart of a specific implementation of a digital enhancement method of a power amplifier of the present invention;

fig. 2 is a model inversion DPD (digital predistortion) structure;

FIG. 3 is a graph of the corresponding normalized power spectrum of the PA modeling output signal and the PA actual output signal of the present invention;

FIG. 4 is a schematic diagram of an implementation structure of the present invention for constructing a LUT-based cascade structure model;

Fig. 5 is a normalized power spectrum of the digital enhanced signal and the original PA output signal according to the method of the present invention.

Detailed Description

Embodiments of the present invention will now be described in detail and with reference to the accompanying drawings.

The embodiment of the invention realizes a digital enhancement method applicable to a 5G signal and facing to a power amplifier, as shown in fig. 1, and the structure of a cascade unit in fig. 1 is shown in fig. 4. The implementation of the steps of the method of the invention will be described separately.

Step one, the signal source module modulates the data symbols X _k by Quadrature Phase Shift Keying (QPSK) or Quadrature Amplitude Modulation (QAM), and generates an OFDM signal X _n by Inverse Fast Fourier Transform (IFFT) and oversampling.

Where X _k is a Quadrature Phase Shift Keying (QPSK) or Quadrature Amplitude Modulation (QAM) modulated data symbol for the kth subcarrier, N is the number of subcarriers, J is the oversampling rate, and J is the imaginary unit.

Step two, the digital enhancement module sets clipping rate CR of Crest Factor Reduction (CFR), x _n is obtained by clipping the time domain of the OFDM signal x _n, PAPR of the OFDM signal is reduced by Clipping and Filtering (CF) in the frequency domain, v _k is generated, and the proportionality coefficient beta is pre-calculated and stored in a table.

Where F _n is time domain clipping noise, F _n is frequency domain clipping noise obtained by Fast Fourier Transform (FFT) F _k,F_k,For the filtered frequency domain clipping noise, β is a scaling factor, V _k is a frequency domain clipping signal, and V _k is subjected to inverse fourier transform (IFFT) to obtain V _k.

Step three, the following parameters can be dynamically configured according to the requirement:

(1) The clipping rate CR of the signal can be configured, and the PAPR of the OFDM signal can be adjusted according to the requirement;

(2) According to the bandwidth BW of the source signal and the corresponding rate SR, the proper multiphase branch number M is set, the delay parameters corresponding to each phase can be configured, the post-addition of the parameters corresponding to each phase delay is realized, and the value of M is as follows:

Wherein, p is the number of half-band filters in the interpolation module, SR is the signal rate, f _clk is the hardware clock frequency, ceil (·) is the upward rounding.

(3) Predistortion (DPD) coefficients are calculated from the source signal and the feedback signal, and stored predistortion (DPD) coefficients corresponding to the primary look-up table and the backup look-up table may be updated.

Step four, v _k generates multiphase signal transmission through interpolation, and performs sampling through a Power Amplifier (PA), and uses a source signal v _k and a sampling signal u _k, as shown in fig. 2, to extract coefficients by adopting a model inversion structure, construct a PA model based on a low-complexity cascade model and perform training, any nonlinearity in the transmission process can be represented by Volterra series, a nonlinear system thereof represents the whole radio frequency signal chain, and the PA is modeled as a second-order nonlinear system y (t) with memory for convenience of explanation. In fig. 2, y (n) is a discrete signal obtained by sampling y (t), y (t) is an output of PA, x (n) is a discrete signal corresponding to x (t), and x (t) is an input of PA.

Where g (T) is a continuous signal modulated by v _k, p (T) is a zero-order keeper, T is a sampling duration, T, τ ₁,τ₂ each represent a corresponding time, 0< τ ₁≤τ₂≤T,w_c represents a carrier frequency, and k ₂(τ₁,τ₂) represents a Volterra series model second-order kernel.

Where z (T) is a component generated corresponding to the second-order nonlinear system, p ₁ (T) is a pulse from time τ ₂ to time t+τ ₁, and p ₂ (T) is a pulse from time τ ₁ to time τ ₂. Z (jw) represents the frequency domain transform of Z (t), w represents frequency, delta represents impulse response, H _1k(jw)、H_2k (jw) respectively corresponds to the frequency response characteristics of the nonlinear term, LTI represents a linear time-invariant system, and NL represents a nonlinear system. The structure of the equivalent baseband output of the PA second-order nonlinear system, after low-pass filtering and downsampling, is shown by the above equation to be composed of two parts, one involving nonlinear conversion of current and previous samples, denoted by v _k ²,v_kv_k+1, denoted by NL, and the other by the linear time-invariant system, denoted by LTI.

By cascading the Volterra model and the LTI system, the PA nonlinearity can be accurately modeled, the power spectrums of the modeled output signal and the PA actual output signal are shown in fig. 3, NMSE (normalized mean square error) performance can reach-36 dB, and high-precision modeling of the PA can be realized.

Step five, the nonlinear part of the PA model selects a GMP model with short memory, long memory LTI is realized by a filter, the extraction process of a coefficient matrix W of the model is explained below, in addition, the cascade model is also adopted to construct a low-complexity compensation module, the model used by the compensation module is the same as the PA model, the parameter extraction process needs to exchange input and output, namely, the output of the PA is used as input, the source signal is used as output, the rest deduction processes are consistent, the specific hardware structure is shown in figure 4, the GMP model coefficients of short memories are stored in the lookup table, and the long memory LTI is realized by the polyphase filter bank.

The GMP model is expressed as follows:

Wherein x (n) represents an input signal, y (n) represents an output signal, a _kl、b_klm、c_klm represents a diagonal term, a lag term and a lead term model coefficient respectively, K _a、L_a represents a nonlinear order and a memory depth of the diagonal term respectively, K _b、M_b、L_b represents a nonlinear order, a lag depth and a memory depth of the lag term respectively, K _c、M_c、L_c represents a nonlinear order, a lead depth and a memory depth of the lead term respectively, and |x (n-l) | represents a signal envelope.

The above is rewritten into a matrix form: Wherein:

Wherein, phi _a、Φ_b、Φ_c is a diagonal term, a lag term and a lead term basis function matrix respectively, W _a、W_b、W_c represents model parameters, the memory depth of the cross term is considered to be lower, the memory effect of the diagonal term is only reduced, the low memory part of the phi _a matrix is written as phi _a', and the dimension is far lower than phi _a.

Ψ_a＝H*Φ_a'

Wherein Φ _a 'is a low-memory basis function column matrix of a diagonal term, H is a convolution kernel of M ₁×N₁, M ₁ is the filter order, N ₁ is the number of filters, the implementation of the FIR filter bank can be configured, and the matrix ψ _a is a convolution output matrix of the Φ _a' column matrix and the convolution kernel H. And combining the phi _a and the phi _b、Φ_c into a new basis function matrix, wherein the phi= [ psi _aΦ_bΦ_c ].

Model coefficient matrixCalculated by LS (least squares) algorithm:

W=(Ψ^ΗΨ)^-1Ψ^ΗX_n

Wherein, the upper corner mark H represents a conjugate matrix, the upper corner mark-1 represents an inverse matrix, and X _n represents an input multiphase signal.

As shown in fig. 1, the generated multiphase signal X _n-L…X_n is input into the cascade unit, where L is the number of cascade units, and may represent multiphase signal X _n = [ X [ nm+1], X [ nm+2]. X [ (n+1) M ] ]. As described in step three, the number of branches M of the multi-phase signal may be set according to the bandwidth and the rate. As shown in fig. 4, the cascade unit of the present invention is provided with an index configuration and a cross index configuration, which are both configurable signal paths, and can input the amplitude of an input signal as an index into a corresponding lookup table query model coefficient, wherein the coefficient index of a diagonal term only relates to the current moment, the amplitude of the current signal indexes the lookup table corresponding to the moment signal through the signal path of the index configuration, the coefficient indexes of a lag term and a lead term relate to the signal intersection at different moments, and the lookup tables at different moments are indexed through the cross signal path of the cross index configuration. The configurable delay is to delay the multi-phase signal such that the multi-phase signal and the compensation coefficient are multiplied at corresponding times.

The multi-phase input signals search model coefficients in the corresponding lookup tables according to signal amplitude, find the corresponding model coefficients from the lookup tables in the index configuration, multiply the signals with the model coefficients at the configured delay time, then filter and output the signals through a multi-phase filter bank, wherein the multi-phase filter bank consists of filters used for reducing the order of the diagonal basis functions, in the cross index configuration, the input multi-phase signals find the corresponding model coefficients from the corresponding lookup tables, multiply the signals with the model coefficients at the configured delay time, output the signals with the corresponding phases after fixed delay and the output of the index configuration are added to obtain compensation signals Y _n = [ Y nM+1], Y nM+2..Y [ (n+1) M ] ].

And step six, calculating a model coefficient according to the input and output signals of the power amplifier, and establishing an LUT in hardware by the model coefficient.

The model coefficients are first transformed by the following equation:

Where K represents the order of the diagonal terms after the reduction of the order.

The product and the memory of the nonlinear power part are used as a table, a table lookup operation is utilized to avoid complex multiplication operation, DSP resources are saved, the implementation difficulty of hardware is effectively reduced compared with direct calculation, the amplitude of a signal is calculated only by rotation of a cordic algorithm (coordinate rotation digital calculation algorithm), and corresponding model parameters are indexed in the memory. Meanwhile, after the model parameters are extracted, in order to reduce the calculated amount of the convolution process in the concrete implementation, considering the linear transformation, ψ _a*W_a is equivalent to Φ _a'*W_a, and the final implementation expression is obtained through convolution:

Where h _k (n) represents the unit impulse response of the k+1th filter in the polyphase filter bank, k=0, 1, 2. According to the formula, the input signal x (n-l) is multiplied with a table value obtained by a corresponding lookup table index and then summed to obtain a final predistortion signal, so that the low-complexity digital enhancement module is realized.

As shown in fig. 5, the effectiveness of the method of the invention is proved by an experimental platform, wherein the center frequency of a 5g micro base station power amplifier is 3.5GHz, the saturated power is 24dbm, 3000 groups of 100-MHz OFDM signals are collected by the method, the model coefficient can be reduced under the condition of ensuring the model precision, and the performance of NMSE (normalized mean square error) can reach-36 dB. When the signal bandwidth is 100M, the ACLR (adjacent channel leakage ratio) of the power amplifier can be improved by more than 15dB, and the ACPR (adjacent channel power ratio) can reach below-45 dBc, so that the practical, feasible, efficient and high-precision in practical engineering can be met.

Claims

1. A digital enhancement method for a power amplifier suitable for 5G signals, characterized in that a digital enhancement module is provided to process an OFDM signal _xn generated by a signal source module; the method comprises the following steps:

(i) The digital enhancement module pre-sets the crest factor clipping rate, adjusts the peak-to-average power ratio of the OFDM signal, generates a clipping signal v _k , interpolates v _k to generate a multi-phase signal for transmission, passes through a power amplifier and performs feedback signal sampling; wherein OFDM stands for orthogonal frequency division multiplexing;

(ii) Constructing a PA model and training the model, using Volterra series to represent the nonlinearity in the signal transmission process, the structure of the PA model equivalent baseband output consists of two parts: a nonlinear system involving the current signal and the previous signal, and a linear time-invariant system; PA stands for power amplifier;

(III) Constructing a compensation module in the digital enhancement module, including:

(1) The nonlinear system is represented by the GMP model. The diagonal term, lag term and lead term basis function matrices of the model are _Φa , _Φb and _Φc respectively, and the corresponding model coefficient matrices are Wa, _Wb _{and Wc} _respectively ; _Wa is the diagonal term model coefficient _akl , _Wb is the lag term model coefficient _bklm , and _Wc is the lead term model coefficient _cklm ;

(2) taking the low memory part of the diagonal basis function matrix _Φa to form a matrix _Φa ', and convolving _Φa ' with the convolution kernel H to form a new basis function matrix Ψ with _Φb and _Φc _; the convolution kernel H is implemented by configuring an FIR filter bank;

(3) Calculate the model coefficient matrix W by the least square method, W = (Ψ ^Ｈ Ψ) ^-1 Ψ ^Ｈ X _n ; wherein W = [W _a W _b W _c ] ^T , the superscript T represents the transpose, H represents the conjugate matrix, -1 represents the inverse matrix, and X _n represents the input multi-phase signal;

(4) Create a lookup table for the model coefficients and implement it in hardware, including:

The output of the GMP model is expressed as follows:

Wherein, K and _La represent the nonlinear order and memory depth of the diagonal term, _Mb and _Lb represent the lag depth and memory depth of the lag term, _Mc and _Lc represent the lead depth and memory depth of the lead term, respectively; |x(nl)| represents the envelope of the input signal x(nl); * represents the convolution operation, and _hk (n) represents the unit impulse response of the k+1th filter in the FIR filter bank;

The model coefficients in the above formula are Calculate the model coefficients in advance to establish a lookup table;

The input signal x(n-l) is multiplied by the model coefficient obtained by the corresponding lookup table index and then summed to obtain the final predistortion signal y(n).

2. The method according to claim 1, characterized in that in the step (i), a crest factor clipping rate CR is set to clip the OFDM signal _xn in the time domain to obtain The peak-to-average power ratio of the OFDM signal is reduced by frequency domain clipping and filtering as follows:

Among them, f _n is the time domain clipping noise, and the frequency domain clipping noise F _k is obtained by fast Fourier transform FFT on f _n . is the frequency domain clipping noise after filtering; N is the number of subcarriers, J is the oversampling rate; X _k is the data symbol modulated by the kth subcarrier; β is the proportional coefficient; V _k is the frequency domain clipping signal, and V _k is obtained by performing inverse Fourier transform IFFT on V _k .

3. The method according to claim 1, characterized in that, in the step (iii), the diagonal terms, lag terms and lead terms basis function matrices of the GMP model are as follows:

Among them, _Ka and _La represent the nonlinear order and memory depth of the diagonal terms, respectively; _Kb , _Mb and _Lb represent the nonlinear order, lag depth and memory depth of the lag terms, respectively; _Kc , _Mc and _Lc represent the nonlinear order, lead depth and memory depth of the lead terms, respectively.

4. The method according to claim 1, 2 or 3, characterized in that the digital enhancement module sets the number M of multi-phase branches according to the bandwidth and signal rate of the source signal, and configures the delay parameters corresponding to each phase.

5. The method according to claim 1, 2 or 3, characterized in that the digital enhancement module calculates the pre-distortion coefficient according to the source signal _vk input to the PA and the sampled feedback signal, and updates the lookup table.

6. The method according to claim 1 or 3 is characterized in that in the step (iii), the generated multi-phase signal is input into a cascade unit, an index configuration and a cross-index configuration are set in the cascade unit, in which each phase signal finds the corresponding model coefficient from a lookup table in the index configuration and the cross-index configuration, and the signal is multiplied by the model coefficient at the configured delay time; the output signal multiplied by the index configuration is input into a FIR filter group, and the output signal multiplied by the cross-index configuration is summed with the signal output by the FIR filter group after a fixed delay to obtain a compensation signal.