A power amplifier predistortion processing method based on a simplified GMP variant model

Abstract

This invention discloses a power amplifier predistortion processing method based on a simplified GMP variant model, comprising: S1: constructing a neural network model for basis function selection and selecting basis functions; S2: constructing a simplified GMP variant model as a power amplifier distortion model based on the selected basis functions; S3: solving the power amplifier distortion model coefficients to obtain the global optimal solution of the power amplifier distortion model coefficients; S4: performing predistortion processing on the baseband signal u according to the obtained global optimal solution of the power amplifier distortion model coefficients, wherein the input signal for predistortion processing is the baseband signal u, the baseband signal u is predistorted to obtain a predistorted signal x as the input of the power amplifier, and the predistorted signal x is processed by the power amplifier to obtain a linearly amplified output signal. This invention reduces the number of coefficients used in the model, reduces computational complexity, and improves the model's adaptability.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a power amplifier predistortion processing method based on a simplified GMP variant model. Background Technology

[0002] With the rapid development of wireless communication, many input signals for power amplifier modules now exhibit characteristics such as high peak-to-average power ratio (PAPR), wide bandwidth, and high-order modulation. This makes power amplifiers more prone to entering the nonlinear region and causing severe memory effects, further impacting their gain. For power amplifier linearization techniques, digital predistortion methods offer advantages such as simplicity, high model accuracy, good stability, and adaptability, and have been widely adopted.

[0003] The core of digital predistortion methods is fitting various distortion characteristics of the power amplifier. Due to the memory effect and nonlinear changes of the power amplifier, memory terms and higher-order terms should be added to achieve high-performance fitting of the distortion model, provided that implementation conditions permit. Among these, the Generalized Memory Polynomial (GMP) model, which incorporates cross-terms of memory lead and lag and higher-order terms, is widely accepted. In recent years, neural networks have also gradually gained attention as an emerging predistortion technology. Theoretically, neural network (NN) models can fit any nonlinear model. Through iterative operation with undetermined coefficients and fixed inputs and outputs, parameters can be calculated and updated, exhibiting strong learning and adaptive capabilities.

[0004] In practice, models that achieve high fitting performance, such as the GMP model (a Waterlasic series VS model) and neural network models, often suffer from drawbacks such as numerous coefficients and complex structures. To increase their applicability in various scenarios and mitigate these shortcomings, many adaptive methods have been proposed, primarily focusing on coefficient adaptation and structure adaptation. The former refers to iteratively updating coefficients until the model reaches an acceptable fitting error; it also includes adaptive parameter modifications during convergence, such as the convergence step size, to improve fitting accuracy. The latter refers to adaptively adjusting the memory depth and polynomial order to change the model structure; or making the model sparser based on input and output signals, thus eliminating the need to calculate many coefficients; or reducing the calculation of less important weight parameters, ultimately reducing model complexity.

[0005] In the future, the main direction for improving predistortion is to maintain fitting accuracy while effectively reducing coefficients, making the hardware structure more streamlined, achieving lower costs, and making the model more adaptive.

[0006] The shortcomings of existing technologies are that the coefficient simplification schemes are not flexible enough, the GMP model is prone to having a large number of coefficients when fitting nonlinearity, the neural network model is limited by the amount of data, training time and complexity, and the overall structure is prone to failure or has high modification costs when the power amplifier performance changes, and the adaptive capability is insufficient. Summary of the Invention

[0007] To address the problem of large number of model coefficients and high computational complexity in existing technologies, this invention proposes a power amplifier predistortion processing method based on a simplified GMP variant model, which reduces the use of redundant coefficients and computational complexity.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A power amplifier predistortion processing method based on a simplified GMP variant model includes the following steps:

[0010] S1: Build a neural network model for basis function selection and select the basis functions;

[0011] S2: Construct a simplified GMP variant model based on the selected basis functions as the power amplifier distortion model;

[0012] S3: Solve for the power amplifier distortion model coefficients to obtain the global optimal solution for the power amplifier distortion model coefficients;

[0013] S4: Based on the globally optimal solution of the power amplifier distortion model coefficients, predistort the baseband signal u. The input signal for the predistortion processing is the baseband signal u. After predistortion processing, the baseband signal u yields a predistorted signal x, which serves as the input to the power amplifier. After processing by the power amplifier, the predistorted signal x yields a linearly amplified output signal.

[0014] The working principle of this invention is as follows:

[0015] The basis function is selected by constructing a neural network model. A simplified GMP variant model is constructed based on the selected basis function as the power amplifier distortion model. The global optimal solution of the power amplifier distortion model coefficients is solved. The power amplifier distortion model is used to obtain the input of the actual power amplifier and obtain the linearly amplified output signal. By changing the basis function, the power amplifier distortion model can be adapted to different inputs or power amplifier changes, realizing the adaptive iteration of the power amplifier distortion model.

[0016] Preferably, the neural network model is a 3-CNN neural network model; the 3-CNN neural network model includes one convolutional layer, one pooling layer, and one fully connected layer.

[0017] Furthermore, the process of constructing a neural network model is as follows:

[0018] The neural network model is trained using a training set and evaluated using a test set; the training set and test set are derived from input and output signals of different durations generated by different power amplifiers at different times.

[0019] Features are obtained by using the input and output sample sets of the power amplifier; the features are then used to calculate feature vectors, and the labels corresponding to the feature vectors are the types of basis functions.

[0020] The features of the training set and their corresponding labels are fed into the neural network model for model training. The features of the test set are fed into the neural network model for model evaluation. The model fitting error is calculated by comparing the results of the neural network model's basis function selection with the correct labels. The structural parameters of the neural network model are then adjusted to make the model fitting error less than the fitting error threshold.

[0021] Furthermore, the aforementioned feature is the time-frequency characteristic of the sequence. Let the baseband signal and the original power amplifier output signal both have N data points, the baseband signal be u(n), n = 1, ..., N, and the original power amplifier output signal be y(n), n = 1, ..., N.

[0022] Suppose there are p time-frequency features calculated from a sequence, then a feature vector of length 2p can be calculated based on a set of u and y; suppose there are g sets of signals, then a feature matrix of size 2p*q can be calculated; for any sequence s, the time-frequency features representing that arbitrary sequence s and the calculation formula are as follows:

[0023] The time-domain features are s_mean, s_std, s_max, and s_min.

[0024] s_mean is the average of the magnitudes of vector s, expressed as:

[0025] s_std is the variance of the magnitude of vector s, expressed as:

[0026] s_max is the maximum value of the magnitude of vector s, expressed as s_max = max(|s|).

[0027] s_min is the minimum value of the magnitude of vector s, expressed as s_min=min(|s|).

[0028] Frequency domain characteristics include p_max and f_pmax; p_max is the maximum power amplitude, obtained through... -π<ω≤π is used to obtain the power spectrum, where j is an imaginary number and ω is the frequency; the maximum power amplitude is calculated from the power spectrum.

[0029] f_pmax is the frequency value corresponding to the maximum power amplitude in the power spectrum.

[0030] Furthermore, the types of basis functions include memory polynomial fitting basis functions, Fourier polynomial fitting basis functions, and Euler polynomial fitting basis functions; the basis function representations and corresponding fitting polynomials are as follows:

[0031] The expression for the basis functions of the memory polynomial fitting is:

[0032] u(nl)|u(nl)| k

[0033] Where n is the data point index of u, l is the memory depth, and k is the order.

[0034] The polynomial fitting of the basis functions using the memory polynomial is as follows:

[0035]

[0036] Where a k，l u represents the basis function coefficients for different memory depths and orders.

[0037] The expression for the basis functions of the Fourier polynomial fitting is:

[0038] coS[u(nl)]|cos[u(nl)]| k

[0039] The Fourier polynomial fitting of the basis functions is as follows:

[0040]

[0041] The expression for the Euler polynomial fitting basis functions is:

[0042]

[0043]

[0044]

[0045] In the formula, A n-l θ represents the amplitude coefficients of the basis functions at different memory depths. n-l These are the frequency coefficients for different memory depths.

[0046] The Euler polynomial fitting of the basis functions is as follows:

[0047]

[0048] In the formula, K is the order and L is the memory depth.

[0049] The baseband signal u corresponding to the eigenvector and the original power amplifier output signal y are fitted using polynomials with different basis functions. The memory depth and order are set, and the corresponding coefficients are calculated using the least squares method. The fitting result is... Fitting error through calculate.

[0050] Set a fitting error threshold. If the fitting error is not less than the fitting error threshold, increase the memory depth or order until the fitting error is less than the threshold. Compare the coefficients used by the polynomials of different basis functions. The basis function with the fewest coefficients best matches the nonlinear characteristics of the power amplifier model. This type of basis function will be labeled as the feature vector.

[0051] Furthermore, the steps to build a simplified GMP variant model are as follows:

[0052] The traditional GMP model is as follows:

[0053]

[0054] Where u is the baseband signal, y GMP For the output signal, a k，l b k，l，m C k，l，m L represents the coefficients of the distinct terms. a L b L c For memory depth, K a K b K c M is the order. b M c The lead or lag length is used; the number of coefficients in the traditional GMP model is calculated to be K. a L a +K b L b M b +K c L c M c One; Substitute the original power amplifier output signal y into y GMP Then, the coefficients are calculated.

[0055] When choosing memory polynomial fitting basis functions and Fourier polynomial fitting basis functions respectively, the traditional GMP model undergoes the following variations:

[0056] y GMP1 (n)=y GMP (n)

[0057]

[0058]

[0059] Let f(n,l,k)=f1(n,l)·|f2(n,l)| k If the basis functions are used, then the variant GMP model and its corresponding values ​​are as follows:

[0060] When f1(n, l) is u(nl) and f2(n, l) is u(nl), the corresponding GMP variant model is represented as y GMP1 ;

[0061] When f1(n, l) is cos[u(nl)] and f2(n, l) is cos[u(nl)], the corresponding GMP variant model is represented as y GMP2 ;

[0062] When f1(n, l) is A n-l f2(n, l) is When, the corresponding GMP variant model is represented as y GMP3 ;

[0063] The coefficients of the GMP variant model are simplified, and the summation terms of the simplified GMP variant model are further simplified. The matrix form of the simplified GMP variant model is as follows:

[0064]

[0065] In the formula, u′ is the baseband signal data point u(n) based on the set memory depth L. a Lead length M b hysteresis length M c The calculated vector has a length of 2(L) a +L b M b +L c M c w′ corresponds to y GMP′ The coefficient vector of (n).

[0066] Furthermore, the steps for simplifying the coefficients of the GMP variant model are as follows:

[0067] When it exists In formal cases, the domain can be any value.

[0068]

[0069]

[0070] The coefficients of the two Taylor formulas are transformed, and the expression is:

[0071]

[0072] When z l =|f2(n,l)|, the formula is expressed as

[0073]

[0074] like q is an even number from 0 to K-1, and

[0075] q is an odd number between 0 and K-1;

[0076]

[0077] Furthermore, the steps for simplifying the summation terms of the GMP variant model after coefficient simplification are as follows:

[0078]

[0079]

[0080]

[0081] The above expressions can be written in the following matrix form:

[0082]

[0083]

[0084]

[0085]

[0086]

[0087]

[0088] Finally, the simplified matrix form of the GMP variant model is obtained.

[0089] Furthermore, the steps to solve for the global optimal solution of the power amplifier distortion model coefficients are as follows:

[0090] Based on the simplified GMP variant model, the corresponding predistortion coefficients are calculated using the power amplifier's input and output; that is, u′ is calculated from the actual power amplifier's distortion output y(n) according to the set memory depth L. a Lead length M b hysteresis length M c The vector obtained by calculation is w′, which is the coefficient vector of u(n), i.e., u(n) = u′·w′;

[0091] All u(n) use the same set of coefficients w′. If the data points for u(n) and y(n) are N, then

[0092] u=Y·ω′,u∈R N×1 ,

[0093] The global optimal solution is calculated using the least squares method: w′=(Y H Y) -1 Y H u.

[0094] The parameter is set to L a =L b =L c =3,M b =M c =2, used to adapt to power amplifier distortion.

[0095] During the solution process, a lookup table is used to assist in the calculation of the sine basis function or cosine basis function, as follows:

[0096] All results from the calculation logic circuit are compiled into a table indexed by x, and the table is pre-written into RAM. During the calculation, the corresponding value in the table is selected as the result based on x and output.

[0097] Furthermore, the input baseband signal u is calculated using the following formula:

[0098] x = U·w′, x∈R N×1 ,

[0099] The predistorted signal x is obtained as the input to the actual power amplifier, and finally the linearly amplified output signal is obtained. The power amplifier distortion model is adaptively iterated, and S1 to S4 are repeated to complete the adaptive adjustment.

[0100] Compared with the prior art, the beneficial effects of the present invention are:

[0101] 1. By constructing a neural network model, the most suitable basis function is selected, and a simplified GMP variant model is constructed to reduce the use of redundant coefficients, reduce computational complexity, and speed up the model's operation.

[0102] 2. By using lookup tables to assist in calculations, the flexibility of coefficient simplification schemes can be improved.

[0103] 3. By observing changes in the power amplifier, the power amplifier distortion model can be adaptively adjusted, thereby improving the adaptive capability of the power amplifier distortion model.

[0104] 4. By using a 3-CNN neural network model, the training time and cost before use are reduced, the distortion relationship of the actual power amplifier is quickly approximated, and the cost is reduced. Attached Figure Description

[0105] Figure 1 This is a flowchart illustrating a power amplifier predistortion processing method based on a simplified GMP variant model.

[0106] Figure 2This is a flowchart illustrating an embodiment of the present invention. Detailed Implementation

[0107] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0108] Example 1

[0109] In this embodiment, as Figure 1 , Figure 2 As shown, a power amplifier predistortion processing method based on a simplified GMP variant model includes the following steps:

[0110] S1: Build a neural network model for basis function selection and select the basis functions.

[0111] S2: Construct a simplified GMP variant model based on the selected basis functions as the power amplifier distortion model.

[0112] S3: Solve for the power amplifier distortion model coefficients to obtain the global optimal solution for the power amplifier distortion model coefficients.

[0113] S4: Based on the globally optimal solution of the power amplifier distortion model coefficients, predistort the baseband signal u. The input signal for the predistortion processing is the baseband signal u. After predistortion processing, the baseband signal u yields a predistorted signal x, which serves as the input to the power amplifier. After processing by the power amplifier, the predistorted signal x yields a linearly amplified output signal.

[0114] The working principle of this invention is as follows:

[0115] The basis function is selected by constructing a neural network model. A simplified GMP variant model is constructed based on the selected basis function as the power amplifier distortion model. The global optimal solution of the power amplifier distortion model coefficients is solved. The power amplifier distortion model is used to obtain the input of the actual power amplifier and obtain the linearly amplified output signal. By changing the basis function, the power amplifier distortion model can be adapted to different inputs or power amplifier changes, realizing the adaptive iteration of the power amplifier distortion model.

[0116] In this embodiment, the 3-CNN neural network model includes one convolutional layer, one pooling layer, and one fully connected layer.

[0117] More specifically, the process of building a neural network model is as follows:

[0118] The neural network model is trained using a training set and evaluated using a test set; the training set and test set are derived from input and output signals of different durations generated by different power amplifiers at different times; the ratio of the training set to the test set is 4:1.

[0119] Features are obtained by using the input and output sample sets of the power amplifier; the features are then used to calculate feature vectors, and the labels corresponding to the feature vectors are the types of basis functions;

[0120] The features of the training set and their corresponding labels are fed into the neural network model for model training. The features of the test set are fed into the neural network model for model evaluation. The model fitting error is calculated by comparing the results of the neural network model's basis function selection with the correct labels. The structural parameters of the neural network model are then adjusted to make the model fitting error less than the fitting error threshold.

[0121] More specifically, the aforementioned feature is the time-frequency characteristic of the sequence. Let the baseband signal and the original power amplifier output signal both have N data points, the baseband signal be u(n), = 1,..., and the original power amplifier output signal be y(n), = 1,...,.

[0122] Suppose there are p time-frequency features calculated from a sequence, then a feature vector of length 2p can be calculated based on a set of u and y; suppose there are q sets of signals, then a feature matrix of size 2p*q can be calculated; for any sequence s, the time-frequency features representing that arbitrary sequence s are calculated as shown in Table 1.

[0123]

[0124] Table 1

[0125] The labels corresponding to the feature vectors are the types of basis functions; the types of basis functions include memory polynomial fitting basis functions, Fourier polynomial fitting basis functions, and Euler polynomial fitting basis functions; the representations of the basis functions and the corresponding fitting polynomials are shown in Table 2:

[0126]

[0127]

[0128] Table 2

[0129] a in the table k,l For the basis function coefficients corresponding to different memory depths and orders; A n-l θ represents the amplitude coefficients of the basis functions at different memory depths. n-l These are the frequency coefficients for different memory depths.

[0130] The baseband signal u corresponding to the eigenvector and the original power amplifier output signal y are fitted using polynomials with different basis functions. The memory depth and order are set, and the corresponding coefficients are calculated using the least squares method. The fitting result is... Fitting error through calculate.

[0131] A fitting error threshold is set. If the fitting error is not less than the threshold, the memory depth or order is increased until the fitting error is less than the threshold. The coefficients used by the polynomials of different basis functions are compared. The basis function with the fewest coefficients best matches the nonlinear characteristics of the power amplifier model, and this type of basis function is labeled as the eigenvector. For the selection of basis functions, the polynomial fitting model here does not need to be the traditional GMP model, i.e., only higher-order terms and simple memory terms are required. After selecting basis functions that match the characteristics of the current power amplifier, the GMP form model can further enhance its nonlinear fitting effect and solve the nonlinear problem in broadband cases.

[0132] More specifically, we choose the general three-layer convolutional neural network model 3-CNN, use the feature vector as the input of the neural network model, and select suitable basis functions for modeling the power amplifier distortion model.

[0133] More specifically, the steps to build a simplified GMP variant model are as follows:

[0134] The traditional GMP model is as follows:

[0135]

[0136] Where u is the baseband signal, y GMP For the output signal, a k，l b k，l，m c k，l，m L represents the coefficients of the distinct terms. a L b L c For memory depth, K a K b K c M is the order. b M c The leading or lagging length is used; therefore, the number of coefficients in this model can be calculated as K. a L a +K b L b M b +K c L c M c In actual calculations, the original power amplifier output y is directly substituted into y; GMP Then, the coefficients are calculated.

[0137] When choosing memory polynomial fitting basis functions and Fourier polynomial fitting basis functions respectively, the traditional GMP model undergoes the following variations:

[0138] y GMP1 (n)=y GMP (n)

[0139]

[0140]

[0141] Let f(n, l, k) = f1(n, l)·f2(n, l, k) be the basis functions. Then the variant GMP model and its corresponding values ​​are shown in Table 3:

[0142]

[0143] Table 3

[0144] The coefficients of the GMP variant model are simplified, and the summation terms of the simplified GMP variant model are further simplified. The matrix form of the simplified GMP variant model is as follows:

[0145]

[0146] In the formula, u′ is the baseband signal data point u(n) based on the set memory depth L. a Lead length M b hysteresis length M c The calculated vector has a length of 2(L) a +L b M b +L c M c w′ corresponds to y GMP′ The coefficient vector of (n).

[0147] More specifically, since the GMP model is essentially a VS-class model in the form of a Taylor series generalization, the main calculation module of the GMP variant formula can use the Taylor formula for auxiliary calculations. This effectively extends higher-order terms to infinite-order terms while merging coefficients at the same memory depth, avoiding the hardware implementation difficulties caused by excessive redundancy. The steps for simplifying the coefficients of the GMP variant model are as follows:

[0148] When it exists In formal cases, the domain can be any value.

[0149]

[0150]

[0151] The coefficients of the two Taylor formulas are transformed, and the expression is:

[0152]

[0153] When z l =|f2(n,l)|, the formula is expressed as

[0154]

[0155] like q is an even number from 0 to K-1, and

[0156] q is an odd number between 0 and K-1;

[0157]

[0158] The above equation is a simplified expression for the coefficients of the GMP variant model. The above coefficient merging process is called the "coefficient simplification" process. Based on the proper matching of basis functions and the condition of infinite-order fitting, the fitting error caused by this coefficient simplification is acceptable. That is, the simplification result of this module can be widely applied in the GMP variant model.

[0159] More specifically, based on the simplified expression of the coefficients of the GMP variant model, the steps for simplifying each summation term of the simplified GMP variant model are as follows:

[0160]

[0161]

[0162]

[0163] The above expressions can be written in the following matrix form:

[0164]

[0165]

[0166]

[0167]

[0168]

[0169]

[0170]

[0171] Finally, the simplified matrix form of the GMP variant model is obtained.

[0172] More specifically, the steps to solve for the global optimal solution of the power amplifier distortion model coefficients are as follows:

[0173] Based on the simplified GMP variant model, the corresponding predistortion coefficients are calculated using the power amplifier's input and output; that is, u′ is calculated from the actual power amplifier's distortion output y(n) according to the set memory depth L. a Lead length Mb hysteresis length M c The vector obtained by calculation is w′, which is the coefficient vector of u(n), i.e., u(n) = u′·w′.

[0174] All u(n) use the same set of coefficients w′. If the data points for u(n) and y(n) are N, then

[0175] u=Y·ω′,u∈R N×1 ,

[0176] The global optimal solution is calculated using the least squares method: w′=(Y H Y) -1 Y H u.

[0177] The parameter is set to L a =L b =L c =3,M b =M c =2, used to adapt to power amplifier distortion.

[0178] During the solution process, a lookup table is used to assist in the calculation of the sine basis function or cosine basis function, as follows:

[0179] Using PLD / FPGA development software, all results of the logic circuit are calculated and created into a table with x as the index value. The table is pre-written into RAM, and during the calculation, the corresponding value in the table is selected as the result and output according to x.

[0180] More specifically, the input baseband signal u is calculated using the following formula:

[0181] x = U·ω′, x∈R N×1 ,

[0182] The predistorted signal x is obtained as the input to the actual power amplifier, and finally the linearly amplified output signal is obtained. The power amplifier distortion model is adaptively iterated, and S1 to S4 are repeated to complete the adaptive adjustment.

[0183] Adaptive Iteration of Power Amplifier Distortion Model: During the continuous operation of the power amplifier, the distortion characteristics of the power amplifier may change due to changes in component temperature or baseband signal. Therefore, the basis functions and coefficients used for predistortion will be updated in real time according to the latest power amplifier output. Since the basis functions are also updated at the same time, the number of coefficients will not increase significantly.

[0184] Example 2

[0185] In this embodiment, a power amplifier predistortion processing method based on a simplified GMP variant model includes the following steps:

[0186] S1: Build a neural network model for basis function selection and select the basis functions;

[0187] S2: Construct a simplified GMP variant model based on the selected basis functions as the power amplifier distortion model;

[0188] S3: Solve for the power amplifier distortion model coefficients to obtain the global optimal solution for the power amplifier distortion model coefficients;

[0189] S4: Based on the globally optimal solution of the power amplifier distortion model coefficients, predistort the baseband signal u. The input signal for the predistortion processing is the baseband signal u. After predistortion processing, the baseband signal u yields a predistorted signal x, which serves as the input to the power amplifier. After processing by the power amplifier, the predistorted signal x yields a linearly amplified output signal.

[0190] Example 3

[0191] In this embodiment, a power amplifier predistortion processing method based on a simplified GMP variant model includes the following steps:

[0192] Build a neural network model for basis function selection; that is, let the neural network only perform the function of basis function selection, without building the coefficients or calculating the parameters of the power amplifier distortion model. Therefore, the neural network will not heavily depend on the various data points of the power amplifier's input and output, and can be widely used; a ready-made neural network model can be used directly to effectively reduce the time cost before using the hardware circuit, such as training time; the classic 3-layer convolutional neural network 3-CNN model can be used here.

[0193] This neural network model performs basis function selection; by using adapted basis functions, it quickly approximates the actual power amplifier distortion relationship.

[0194] A simplified GMP variant model is constructed as a power amplifier distortion model. Based on the selected basis functions, a corresponding GMP variant model is designed, modifying the traditional GMP model. Taylor series is used to simplify the model and reduce the number of coefficients. The computational cost of coefficients is reduced while retaining necessary polynomials that maintain nonlinear fitting performance, such as higher-order terms and memory cross terms.

[0195] The global optimal solution for calculating the power amplifier distortion model coefficients is obtained, and a lookup table will be used to assist in the calculation of special basis functions.

[0196] Predistortion processing of the original power amplifier input, i.e., the baseband signal u, involves using a GMP variant model and directly using the coefficients calculated in the previous step, while setting the power amplifier gain. This yields a predistorted signal x, which serves as the input to the actual power amplifier, ultimately resulting in a linearly amplified output signal. Here, lookup tables will also be used to assist in operations on special basis functions.

[0197] The power amplifier model is adaptive and iterative; the basis functions can be changed at any time to adapt to different inputs or power amplifier changes without adding a large number of coefficients for adaptive adjustment.

[0198] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A power amplifier predistortion processing method based on a simplified GMP variant model, characterized in that, Includes the following steps: S1: Build a neural network model for basis function selection; the process of building the neural network model is as follows: The neural network model is trained using a training set and evaluated using a test set; the training set and test set are derived from input and output signals of different durations generated by different power amplifiers at different times. Features are obtained by using the input and output sample sets of the power amplifier; feature The feature vectors are obtained through calculation, and the labels corresponding to the feature vectors are the types of basis functions; The features of the training set and their corresponding labels are fed into the neural network model for training. The features of the test set are then fed into the neural network model for evaluation. The model fitting error is calculated by comparing the selected basis functions with the correct labels. The structural parameters of the neural network model are adjusted to ensure the fitting error is less than a threshold. Types of basis functions include memory polynomial fitting basis functions, Fourier polynomial fitting basis functions, and Euler polynomial fitting basis functions. The representations of the basis functions and their corresponding fitting polynomials are as follows: The expression for the basis functions of the memory polynomial fitting is: in for Data point sequence number, For memory depth, It is the order; The polynomial fitting of the basis functions using the memory polynomial is as follows: in These are the basis function coefficients corresponding to different memory depths and orders; The expression for the basis functions of the Fourier polynomial fitting is: The Fourier polynomial fitting of the basis functions is as follows: ； The expression for the Euler polynomial fitting basis functions is: In the formula, These are the amplitude coefficients of the basis functions at different memory depths. These are the frequency coefficients for different memory depths; The Euler polynomial fitting of the basis functions is as follows: ； In the formula, For order, For memory depth; The baseband signal u corresponding to the eigenvector and the original power amplifier output signal y are fitted using polynomials with different basis functions. The memory depth and order are set, and the corresponding coefficients are calculated using the least squares method. The fitting result is... Fitting error pass Calculations are performed, assuming the baseband signal and the original power amplifier output signal data points are both... ,in For baseband signals, , This is the original power amplifier output signal. ; Set a fitting error threshold. If the fitting error is not less than the fitting error threshold, increase the memory depth or order until the fitting error is less than the threshold. Compare the coefficients used by the polynomials of different basis functions. The basis function with the fewest coefficients best matches the nonlinear characteristics of the power amplifier model. This type of basis function will be labeled as the feature vector. S2: Construct a simplified GMP variant model based on the selected basis functions as the power amplifier distortion model; S3: Solve for the power amplifier distortion model coefficients to obtain the global optimal solution for the power amplifier distortion model coefficients; S4: Based on the globally optimal solution of the obtained power amplifier distortion model coefficients, the baseband signal... u Predistortion processing is performed, and the input signal for predistortion processing is the baseband signal. u， Baseband signal u The predistorted signal is obtained after predistortion processing. x As the input to the power amplifier, the predistortion signal x The output signal is linearly amplified after processing by the power amplifier. .

2. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 1, characterized in that, The neural network model described uses a 3-CNN neural network model; the 3-CNN neural network model includes one convolutional layer, one pooling layer, and one fully connected layer.

3. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 2, characterized in that, The aforementioned feature is the time-frequency characteristic of the sequence; Suppose that the time-frequency characteristics calculated from a sequence are: p One, then according to a set u, y Calculate 2 p Long feature vectors; It has q Given a set of signals, calculate 2. p*q The size of the feature matrix; for any sequence s , representing the arbitrary sequence s The time-frequency characteristics and calculation formulas are as follows: Time-domain characteristics include , , , ; For vectors The average value of the modulus is expressed as: ; For vectors The variance of the modulus is expressed as: ; For vectors The maximum value of the modulus is expressed as follows: ; For vectors The minimum value of the modulus is expressed as follows: ; Frequency domain characteristics include , ; For the maximum power amplitude, through To obtain the power spectrum, in the formula It is an imaginary number. It is the frequency; the maximum power amplitude is calculated using the power spectrum. This is the frequency value corresponding to the maximum power amplitude in the power spectrum.

4. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 3, characterized in that, The steps to build a simplified GMP variant model are as follows: The traditional GMP model is as follows: in u For baseband signals, For output signal, , For the coefficients of different terms, , , For memory depth, , , For order, , The lead or lag length is used; the number of coefficients in the traditional GMP model is calculated as follows: One; the original power amplifier output signal y Substitute Then, the coefficients are calculated. When choosing memory polynomial fitting basis functions and Fourier polynomial fitting basis functions respectively, the traditional GMP model undergoes the following variations: set up If the basis functions are used, then the variant GMP model and its corresponding values ​​are as follows: when for , for When, the corresponding GMP variant model is represented as ; when for , for When, the corresponding GMP variant model is represented as ; when for , for When, the corresponding GMP variant model is represented as ; The coefficients of the GMP variant model are simplified, and the summation terms of the simplified GMP variant model are further simplified. The matrix form of the simplified GMP variant model is as follows: In the formula, It consists of baseband signal data points Based on the set memory depth Advanced length lag length The calculated vector has a length of ; It corresponds The coefficient vector.

5. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 4, characterized in that, The steps for simplifying the coefficients of the GMP variant model are as follows: When it exists In formal cases, the domain can be any value. The coefficients of the two Taylor formulas are transformed, and the expression is: like , It is an even number between 0 and K-1, and , Odd numbers ranging from 0 to K-1; 6. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 5, characterized in that, The steps for simplifying the summation terms of the GMP variant model with simplified coefficients are as follows: The above expressions can be written in the following matrix form: Finally, the simplified matrix form of the GMP variant model is obtained.

7. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 6, characterized in that, The steps to find the global optimal solution for the power amplifier distortion model coefficients are as follows: Based on the simplified GMP variant model, the corresponding predistortion coefficients are calculated using the power amplifier's input and output; that is... It is caused by the distortion output of the actual power amplifier. Based on the set memory depth Advanced length lag length The vector obtained by calculation in the middle, It corresponds The coefficient vector, i.e. ; All They all use the same set of coefficients ,like and If the number of data points is N, then Calculate the global optimal solution using the least squares method: ; Parameters are set to , This is used to adapt to power amplifier distortion. During the solution process, a lookup table is used to assist in the calculation of the sine basis function or cosine basis function, as follows: Calculate all the results of the logic circuit and create a sheet of... A table of indexed values ​​is created and pre-written into RAM for calculation. Select the corresponding value from the table as the result and output it.

8. The power amplifier predistortion processing method based on a simplified GMP variant model according to claim 7, characterized in that, Input baseband signal u Calculated using the following formula: Obtaining predistortion signal x As the input to the actual power amplifier, the final output signal is obtained after linear amplification. Adaptive iteration is performed on the power amplifier distortion model, repeating S1~S4 to complete the adaptive adjustment.

Images