A memory polynomial digital predistortion method based on multi-dimensional amplitude switching

Abstract

The application discloses a wideband digital pre-distortion method based on a multi-dimensional amplitude switching memory polynomial, and proposes a switching behavior model, which divides the input signal sequence of the model into different classes by using the amplitudes of the current input signal and the amplitudes of the memory input signal, and uses a separate model to operate the input data of each class. Each class uses a memory polynomial model as a sub-model, combines the multi-dimensional amplitude classification algorithm with the MP model, proposes a memory polynomial model based on multi-dimensional amplitude switching, and adopts a direct learning structure to iteratively complete nonlinear system identification and model coefficient calculation. The method finally restores data closer to the real power amplifier output signal at the same sampling rate, achieves better pre-distortion effect, and has certain improvement in performance compared with the digital pre-distortion based on the MP model.

Description

Technical Field

[0001] This invention relates to the field of digital signal processing, and more specifically to a memory polynomial digital predistortion method based on multidimensional amplitude switching. Background Technology

[0002] In recent years, wireless communication technology has developed rapidly, especially with the advent of the 5G era, where the data transmission rate of mobile communication systems has exceeded 1Gbps. In the 5G era, the goal is to achieve even higher transmission rates, greater capacity, and lower latency. For designers, achieving these new functions while maintaining high energy efficiency has become a challenging yet crucial task in the design of future wireless communication systems. With the pursuit of higher spectral efficiency and data rates, more and more complex modulation schemes and wider signal bandwidths will be introduced. However, realizing complex modulation schemes and larger transmission bandwidths requires more complex power amplifier architectures, which places higher demands on the transmitter's linearization mechanism. Meanwhile, power amplifiers are widely considered the most power-consuming devices in the RF front-end. To ensure power amplifier efficiency, they are usually pushed to the saturation region; however, this introduces nonlinearity issues. Because the power amplifier operates in the nonlinear region, the large output signal of the power amplifier will experience nonlinear distortion. The resulting intermodulation distortion, spectrum spread, and strong channel interference will all significantly impact communication quality. In the past, the most widely used method to address the nonlinearity problem of power amplifiers was power back-off. As the name suggests, this involves reducing the output power of the power amplifier to operate in the linear region, thereby resolving the nonlinear distortion. However, backing up power reduces the efficiency of the power amplifier, resulting in significant energy waste. For RF system designers, striking a balance between efficiency and linearity is a challenging problem. Simultaneously, with the development of wireless communication technology, increasingly higher information transmission rates are required. This makes an inherent problem of power amplifiers more prominent: the large memory problem. That is, the output of the current power amplifier is not only controlled by the current input signal but also influenced by previous output signals, which can significantly affect the output of the power amplifier.

[0003] Distortion caused by nonlinearity and memory effects must be compensated for using effective linearization methods. Digital predistortion (DPD) is currently the most commonly used linearization technique, allowing wireless LANs to operate at higher power levels without sacrificing linearity, thus achieving higher efficiency and enabling more efficient wireless systems. This invention considers the different behaviors of power amplifiers under different signal inputs and classifies and models them according to different memory term amplitudes and different input signal amplitudes. This allows for accurate modeling and compensation of the nonlinearity and memory properties of the power amplifier. Summary of the Invention

[0004] The purpose of this invention is to address the problem that current methods using a single behavioral model for power amplifiers cannot accurately model power amplifiers under different input conditions.

[0005] A memory polynomial digital predistortion method based on multidimensional amplitude switching is proposed, and the specific technical solution is as follows:

[0006] A memory polynomial digital predistortion method based on multidimensional amplitude switching, characterized in that:

[0007] Includes the following steps:

[0008] S1. Input signal x(n), classify the input signal x(n) according to the multidimensional amplitude switching method;

[0009] S2. Based on the classification results, model multiple sub-amplifiers;

[0010] S3. There are multiple predistorter sub-models. The coefficients of each predistorter sub-model are initialized. The input signal is classified by the threshold value of each input delay term. The signal is handed over to the corresponding predistorter sub-model for processing according to the classification result to obtain y(n) and the output signal z(n) of the hardware system is obtained.

[0011] e(n) = z(n) - x(n), based on the classification results of the input signal, determine the class to which each e(n) belongs, and then use the following model coefficient update formula to update the coefficients;

[0012]

[0013] in, This is the iteration step value. For positive modeling of linear terms Let K be the sub-kernel matrix of the k-th sub-model. This represents the error data for the k-th sub-model;

[0014] S4. Repeat S1-S3 to complete the model coefficient update for all predistorters;

[0015] S5. Input signal x(n), select the corresponding digital predistorter according to the multidimensional amplitude classification algorithm to process it, and obtain the predistorted signal z(n). Then pass z(n) through the power amplifier model to obtain the power amplifier model output v(n).

[0016] S6. The effect of the predistorter is judged based on the error amplitude of the input signal x(n) and the output signal v(n), i.e., the magnitude of |x(n)-v(n)|. The predistorter has the best effect when the error amplitude is the smallest.

[0017] To better realize the present invention, it can be further configured as follows:

[0018] The multidimensional amplitude classification algorithm includes the following steps:

[0019] S1-1. Write the input signal of the MP model in the form of a matrix, where each row represents the input value at each time step, including the current input term x(n) and the memory input terms x(n-1),...,x(nM), and the columns represent the model input information at different time steps.

[0020]

[0021] S1-2. Determine the threshold value for each item. First, determine the number of segments for each item. Based on the number of segments for each item, determine the total number of categories of the output signal, which is the product of the number of segments for each item.

[0022] S1-3. Based on the threshold values ​​determined in S1-2, obtain the classification results of the input signal;

[0023]

[0024] in, , This represents the total number of categories.

[0025] It can be further described as follows:

[0026] The determination of the sub-model coefficients for the forward modeling of the power amplifier satisfies the following steps:

[0027] S2-1, Data set of input signal The input sequence at each time step is classified according to its amplitude to determine the sub-model to be used;

[0028] S2-2. Find the kernel matrix of the MP model. The mathematical expression of the MP model is:

[0029] It can be written in matrix form as follows: ,Will Defined as the input data kernel matrix, its expression is as follows:

[0030]

[0031] W represents the model coefficients, expressed as follows:

[0032]

[0033] S2-3, Kernel matrix block partitioning given dataset The dataset is analyzed using the mathematical model of the MP model. The kernel matrix is ​​obtained through calculation. ;

[0034] Based on the classification results of S1, the kernel matrix is ​​divided into blocks, and the row vectors of the kernel matrix formed by datasets belonging to the same class are divided into the same sub-matrix.

[0035] S2-4, Input Signal Sequence With power amplifier output data It is a set of correspondences;

[0036] Based on the classification results of the input signal sequence, the output signals of the power amplifier are classified accordingly to obtain the power amplifier output signals of various types;

[0037] S2-5. Solve the coefficients of the sub-model using the least squares method.

[0038] It can be further described as follows:

[0039] The iterative update process of the predistortion model coefficients satisfies the following steps:

[0040] S3-1: Initialize the coefficients of each sub-model Make the first term of the coefficient All other terms are set to zero;

[0041] S3-2: For a given dataset Using the method described above for obtaining the kernel matrix of the MP model, the total kernel matrix is ​​obtained. ;

[0042] The kernel matrix is ​​classified according to the determined segmentation thresholds to obtain the classified sub-kernel matrices. and use Mark the classification result for each input;

[0043] S3-3: For each sub-model, after classifying the input data, based on the classification results... For each input sequence, a corresponding predistorter sub-model is selected, utilizing... The model output for each iteration is obtained as follows: ;

[0044] S3-4: Obtain the output signal z(n) of the power amplifier and get the error. ;

[0045] Based on the classification results of the input data The error data is classified to obtain the error data after DPD for each sub-model, i.e. , used for iterative updates of sub-model coefficients;

[0046] S3-5: Based on the subkernel matrix The error data for each sub-model were obtained;

[0047] The iteration coefficients of each sub-model are obtained using Newton's iteration formula, as shown in the following equation.

[0048]

[0049] in, This is the iteration step value. For positive modeling of linear terms Let K be the sub-kernel matrix of the k-th sub-model. This represents the error data for the k-th sub-model;

[0050] S3-6: Iterate through S1-2 to S1-5, continuously optimizing the coefficients of the sub-model until the iteration count is complete.

[0051] The beneficial effects of this invention are as follows:

[0052] This invention proposes a memory polynomial model based on multidimensional amplitude switching. By dividing the amplitude of the current input signal and the amplitude of the memory input, the input sequence of the model is classified, and a separate model is used to pre-distort each type of data.

[0053] Simultaneously, employing a direct learning structure, a coefficient update formula for the sub-model based on Newton's iteration method is proposed. Compared to generalized memory polynomials, this model significantly improves both its power amplifier forward modeling capability and predistortion performance. Furthermore, compared to generalized memory polynomials, it achieves comparable performance with minimal model complexity and computational cost.

[0054] 1. This invention greatly improves the accuracy of modeling by classifying and modeling signals under different input conditions. The invented MASMP model improves the NMSE by 7.19dB compared to the MP model in terms of forward modeling capability.

[0055] 2. This invention overcomes the problem of training discrete data models for data classification and modeling. Based on Newton's iteration method, it proposes a training formula for the sub-models after classification. In terms of predistortion effect, the ACPR of this invention is 1.81dB higher than that of the MP model.

[0056] 3. The present invention proposes a memory polynomial digital predistortion system based on multidimensional amplitude switching, which improves the adjacent channel power ratio (ACPR) of the Doherty class power amplifier output signal by 12.77 dB. Attached Figure Description

[0057] Figure 1 This is a block diagram illustrating the implementation principle of the predistortion technology in the memory polynomial digital predistortion system based on multidimensional amplitude switching of the present invention.

[0058] Figure 2 It is the power spectrum of the power amplifier output and the power spectrum of the MP model output;

[0059] Figure 3 It is the power spectrum of the power amplifier output and the power spectrum of the MASMP model output;

[0060] Figure 4 It is the power amplifier output power spectrum and the GMP model output power spectrum;

[0061] Figure 5 It refers to the MP model, MASMP model, and GMP model applied to the output power spectrum of DPD and PA. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] like Figures 1 to 5 As shown:

[0064] S1: Establish a complete digital predistortion hardware platform, including a computer (PC), vector signal generator (VSG), power amplifier (PA), spectrum analyzer (VSA), digital power supply, attenuator, etc. The power amplifier used is a broadband Doherty class power amplifier with a working frequency of 1.8GHz.

[0065] S2: Test using a 32QAM signal with a bandwidth of 40MHz;

[0066] S3: The transmitter of a digital predistortion system is formed by connecting a vector signal generator to a linear driver, and then connecting the linear driver to a power amplifier.

[0067] S4: The power amplifier generates an output signal that is connected to the spectrum analyzer via an attenuator to form a digital predistortion feedback channel;

[0068] S5: The vector signal generator is connected to the PC via a router. The MATLAB on the PC controls the vector signal generator to generate a baseband signal and upconverts the baseband signal to the carrier frequency of the power amplifier. Then, the upconverted RF signal is amplified by a linear drive amplifier and sent to a Doherty class power amplifier that needs to be linearized. The power amplifier output is attenuated by an attenuator and then downconverted by a spectrum analyzer and captured and sent to the PC. The transmitted and received signal data are synchronized and aligned.

[0069] S6: Perform forward modeling of the MP model, MASMP model, and GMP model to obtain the linear gain of the model, i.e., the first term of the model coefficients.

[0070] The power amplifier model modeled by the MP model has a NSME of -37.22dB between the predicted output and the actual output of the power amplifier. The modeling accuracy of the MASMP model is -44.41dB, and the modeling accuracy of the GMP model is -44.75dB.

[0071] S7: Perform signal predistortion iteration. If it is the first iteration, initialize the coefficients of each sub-model of the MASMPMP model, keep the first coefficient of the model coefficient as 1 and the rest as 0, so that the signal remains unchanged after passing through the predistortion module. If it is not the first iteration, then obtain the predistorted signal.

[0072] S8: The pre-distorted signal obtained in S7 is generated by a vector signal generator, and then sent to the Doherty power amplifier that requires pre-distortion via a linear driver. The power amplifier output is attenuated by an attenuator and then sent to a signal analyzer to obtain a digital signal, which is finally transmitted back to the PC.

[0073] S9: Align the acquired power amplifier output signal with the predistortion processed signal, and use the obtained error signal to update the coefficients of all sub-models to obtain a better predistortion effect.

[0074] S10: Repeat S7-S9 to iteratively update the model coefficients. Here, the number of iterations for the MASMP model is chosen to be 10, and finally the coefficients of the predistortion model are obtained.

[0075] S11: Perform pre-distortion processing on the signal, and use the coefficients obtained in S10 to acquire data through the above-mentioned transmission channel and feedback channel.

[0076] S12: The ACPR of the power amplifier output signal without digital predistortion is -33.34dB, the ACPR of the power amplifier output signal after predistortion using the MP model is -44.23dB, the ACPR of the power amplifier output signal after predistortion using the MASMP model is -46.01dB, and the ACPR of the power amplifier output signal after predistortion using the GMP model is -46.84dB.

[0077] The multidimensional amplitude classification algorithm satisfies the following steps:

[0078] Step 1. Write the input signal of the MP model in the form of a matrix, where each row represents the input value at each time step, including the current input term x(n) and the memory input terms x(n-1),...,x(nM), and the columns represent the model input information at different time steps;

[0079]

[0080] Step 2. Determine the threshold values ​​for each item. First, determine the number of segments for each item. For example, if the current input item is divided into three segments, two threshold values ​​are needed, which are taken from the values ​​at 1 / 3 and 2 / 3 of the signal amplitude, respectively. If the memory delay item is divided into two segments, one threshold value is needed, which is taken from the value at 1 / 2 of the signal amplitude. Based on the number of segments for each item, determine the total number of categories of the output signal, which is the product of the number of segments for each item.

[0081] Step 3. Based on the threshold values ​​determined in Step 2, obtain the classification results of the input signal;

[0082]

[0083] in, , This represents the total number of categories.

[0084] The determination of the sub-model coefficients for the forward modeling of the power amplifier satisfies the following steps:

[0085] S6-1. Input Signal Classification: For input signal datasets... The input sequence at each time step is classified according to its amplitude to determine the sub-model to be used;

[0086] S6-2. Find the kernel matrix of the MP model. The mathematical expression of the MP model is:

[0087] The matrix form is shown below: ,Will Defined as the input data kernel matrix, its expression is as follows:

[0088]

[0089] W represents the model coefficients, expressed as follows:

[0090]

[0091] S6-3. Kernel Matrix Blocking Given a Dataset The kernel matrix is ​​obtained by operating the dataset using the mathematical model of the MP model. Based on the classification results of S1, the kernel matrix is ​​divided into blocks, with the row vectors of the kernel matrices formed by datasets belonging to the same class grouped into the same sub-matrix, such as...

[0092]

[0093] S6-4. Power Amplifier Output Signal Classification For the model input and power amplifier output signals, the input signal sequence... With power amplifier output data It is a set of correspondences. Based on the classification results of the input signal sequence, the output signals of the power amplifier are classified accordingly to obtain the power amplifier output signals of various categories, such as... .

[0094] S6-5. Sub-model Coefficient Calculation For each sub-model, its input kernel matrix has been obtained in step 3, and its output is obtained in step 5. The coefficients of each forward sub-model can be obtained using the least squares method. For example, the coefficients of the first model can be obtained using the following formula... .

[0095] S6-6. Initialize the coefficients of each sub-model of the predistortion, classify the input signal of the model through the threshold value of each input delay term, and give the signal to the corresponding sub-model for processing according to the classification result to obtain y(n). Transmit the test data y(n) into the hardware system and obtain the output signal z(n) of the power amplifier in the signal acquisition device.

[0096] S6-7. The output signal z(n) of the power amplifier is sampled and normalized. The class of each e(n) is determined by e(n) = z(n) - x(n) based on the classification result of the input signal. Then, the coefficients are updated using the model coefficient update formula.

[0097] The iterative update process of the predistortion model coefficients satisfies the following steps:

[0098] S9-1. Initialize the coefficients of each sub-model Make the first term of the coefficient All other terms are set to zero.

[0099] S9-2. Input Data Classification for a Given Dataset Using the method described above for obtaining the kernel matrix of the MP model, the total kernel matrix is ​​obtained. The kernel matrix is ​​classified according to the determined segmentation thresholds to obtain the classified sub-kernel matrices. and use Mark the classification result for each input.

[0100] S9-3. Data Predistortion Processing: For each sub-model, after classifying the input data, based on the classification results... For each input sequence, a corresponding sub-model is selected, utilizing... The model output for each iteration is obtained as follows: .

[0101] S9-4. Error extraction and classification yield the output of all models. Then, the test data y(n) is transmitted into the hardware system, and the output signal z(n) of the power amplifier is obtained using a signal acquisition device. (Error) Meanwhile, based on the classification results of the input data The error data is classified to obtain the error data after DPD for each sub-model, i.e. It is used for iterative updates of the sub-model coefficients.

[0102] S9-5. Sub-model coefficient update based on sub-kernel matrix The error data of each sub-model is obtained, and the iteration coefficients of each sub-model are calculated using Newton's iteration formula, as shown in the following formula.

[0103] ,in, This is the iteration step value. For positive modeling of linear terms Let K be the sub-kernel matrix of the k-th sub-model. This represents the error data for the k-th sub-model.

[0104] S9-6. Iterate through S9-2 to S9-5, continuously optimizing the coefficients of the sub-model until the iteration count is complete.

[0105] S9-7. Input signal x(n), select the corresponding digital predistorter according to the multidimensional amplitude classification algorithm to process it, and obtain the predistorted signal z(n). Then pass z(n) through the power amplifier model to obtain the power amplifier model output v(n), and enter S7.

[0106] S9-8. Determine the effect of the predistorter based on the error magnitude of the input signal x(n) and the output signal v(n), i.e., the magnitude of |x(n)-v(n)|.

[0107] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0108] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A memory polynomial digital predistortion method based on multidimensional amplitude switching, characterized in that: Includes the following steps: S1. Input signal x(n), classify the input signal x(n) according to the multidimensional amplitude switching method; S2. Based on the classification results, model multiple sub-amplifiers; S3. There are multiple predistorter sub-models. The coefficients of each predistorter sub-model are initialized. The input signal is classified by the threshold value of each input delay term. The signal is handed over to the corresponding predistorter sub-model for processing according to the classification result to obtain y(n) and the output signal z(n) of the hardware system is obtained. e(n) = z(n) - x(n), based on the classification results of the input signal, determine the class to which each e(n) belongs, and then use the following model coefficient update formula to update the coefficients; in, This is the iteration step value. To model linear terms in a positive direction, Let K be the sub-kernel matrix of the k-th sub-model. This represents the error data for the k-th sub-model; S4. Repeat S1-S3 to complete the model coefficient update for all predistorters; S5. Input signal x(n), select the corresponding digital predistorter according to the multidimensional amplitude classification algorithm to process it, and obtain the predistorted signal z(n). Then pass z(n) through the power amplifier model to obtain the power amplifier model output v(n). S6. Determine the effect of the predistorter based on the error amplitude of the input signal x(n) and the output signal v(n), i.e., the magnitude of |x(n)-v(n)|. When the error amplitude value converges, obtain the predistorter model coefficients. The multidimensional amplitude classification algorithm includes the following steps: S1-1. Write the input signal of the MP model in the form of a matrix, where each row represents the input value at each time step, including the current input term x(n) and the memory input terms x(n-1),...,x(nM), and the columns represent the model input information at different time steps. S1-2. Determine the threshold value for each item. First, determine the number of segments for each item. Based on the number of segments for each item, determine the total number of categories of the output signal, which is the product of the number of segments for each item. S1-3. Based on the threshold values ​​determined in S1-2, obtain the classification results of the input signal; in, , This represents the total number of categories.

2. The memory polynomial digital predistortion method based on multidimensional amplitude switching according to claim 1, characterized in that: The determination of the sub-model coefficients for the forward modeling of the power amplifier satisfies the following steps: S2-1, Data set of input signal The input sequence at each time step is classified according to its amplitude to determine the sub-model to be used; S2-2. Find the kernel matrix of the MP model. The mathematical expression of the MP model is: ； It can be written in matrix form as follows: ,Will Defined as the input data kernel matrix, its expression is as follows: W represents the model coefficients, expressed as follows: S2-3, Kernel matrix block partitioning given dataset The dataset is analyzed using the mathematical model of the MP model. The kernel matrix is ​​obtained through calculation. ; Based on the classification results of S1, the kernel matrix is ​​divided into blocks, and the row vectors of the kernel matrix formed by datasets belonging to the same class are divided into the same sub-matrix. S2-4, Input Signal Sequence , and power amplifier output data It is a set of correspondences; Based on the classification results of the input signal sequence, the output signals of the power amplifier are classified accordingly to obtain the power amplifier output signals of various types; S2-5. Solve the coefficients of the sub-model using the least squares method.

Images