Digital pre-distortion method and apparatus, and device, storage medium and program product

By initializing the GMP model as a neural network architecture in digital predistortion technology, and utilizing adaptive filtering and neural network iterative algorithms, the problems of slow convergence speed and poor accuracy in high bandwidth scenarios are solved, achieving a more efficient signal correction effect.

WO2026233149A1PCT designated stage Publication Date: 2026-11-12CHINA MOBILE COMM LTD RES INST +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/CN2026/092965
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2026-04-24
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Existing digital predistortion techniques suffer from slow convergence speed and poor convergence accuracy in high-bandwidth scenarios. In particular, the simplified Volterra series model cannot accurately fit the thermal memory effect caused by broadband signals, resulting in residual predistortion errors and difficulties in parameter iteration.

Method used

The GMP model is initialized using an adaptive filtering parameter iterative algorithm, then converted into a neural network architecture. The system is stabilized using an adaptive filtering algorithm, and the parameters are optimized using a neural network iterative algorithm to achieve architecture reconstruction, thereby improving convergence speed and accuracy.

Benefits of technology

It significantly improves the convergence speed and accuracy of digital predistortion, saves circuit area, handles complex input-output relationships through neural network architecture, supports diverse calculations, and avoids the maintenance of GMP models and neural network parallel circuits.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN2026092965_12112026_PF_FP_ABST
    Figure CN2026092965_12112026_PF_FP_ABST
Patent Text Reader

Abstract

Disclosed in the present application are a digital pre-distortion method and apparatus, and a device, a storage medium and a program product. In an initial stage of a system, a first parameter iteration algorithm is used to perform an initialization operation on a basic processing unit; in a stable stage of the system, a second parameter iteration algorithm is used to perform parameter iteration on the basic processing unit; and when an iteration parameter reaches a first parameter threshold, the basic processing unit is reconstructed, such that the basic processing unit is reconstructed from a lookup table structure to a neural network architecture. The reconstructed neural network architecture can handle more complex input-output relationships and support diverse neural network computations, thereby significantly improving the convergence accuracy and convergence speed of digital pre-distortion. In addition, by means of a reconfigurable architecture, dynamic switching between two architectures is realized at a hardware level, which avoids simultaneously maintaining parallel circuits for a GMP model and a neural network, thereby reducing a circuit area.
Need to check novelty before this filing date? Find Prior Art

Description

Digital predistortion methods, apparatuses, devices, storage media, and program products

[0001] Cross-references to related applications

[0002] This disclosure is based on and claims priority to Chinese Patent Application No. 202510590464.0, filed on May 8, 2025, the entire contents of which are incorporated herein by way of inclusion. Technical Field

[0003] This disclosure relates to the field of wireless communications, and more particularly to a digital predistortion method, apparatus, device, storage medium, and program product. Background Technology

[0004] As the communications industry evolves, when signal bandwidth increases to the 10-200MHz range, the distortion characteristics of power amplifiers are no longer simple static nonlinearities, but exhibit time-dependent characteristics. That is, the output at the current moment depends not only on the current input but also on historical input signals; this phenomenon is called the memory effect. The superposition of the memory effect and nonlinearity leads to frequency-selective distortion, time delay, and phase distortion problems in the transmitted signal, destroying the signal phase integrity.

[0005] Digital pre-distortion (DPD) is a technique used to improve the linearity of power amplifiers. Among related DPD techniques, the integrated modeling of memory effects and static nonlinearity, represented by Volterra series and its simplified model, has gradually become the main solution for engineering implementation. However, the complexity of Volterra series increases with the memory depth and nonlinear order, making it impractical for real-time computation in high-bandwidth scenarios. Therefore, in practical engineering, the practicality of this behavioral model is often improved by simplification. The Volterra behavioral model is simplified using generalized memory polynomial (GMP), and complexity is reduced by removing "minor terms" (such as higher-order cross terms and long memory depth terms) in the Volterra series. However, when the power amplifier has a strong memory effect (such as the thermal memory effect caused by broadband signals), the lack of cross terms may result in inaccurate fitting, leading to residual pre-distortion errors and decreased convergence accuracy. Furthermore, when the model structure does not match the actual power amplifier characteristics, parameter iteration is prone to getting trapped in local optima, resulting in slower convergence speed. Summary of the Invention

[0006] The purpose of this disclosure is to provide a digital predistortion method, apparatus, device, storage medium, and program product to improve the convergence speed and convergence accuracy in the digital predistortion process.

[0007] To achieve the above objectives, embodiments of this disclosure provide a digital predistortion method, comprising:

[0008] In the initial stage of the system, the basic processing unit is initialized using the first parameter iterative algorithm; wherein, the basic processing unit is configured as a lookup table structure of the GMP model, and the basic processing unit is used to output the predistortion signal;

[0009] During the system stabilization phase, the initial iteration parameters after the initialization operation are obtained;

[0010] Based on the initial iteration parameters, the basic processing unit is iterated using the second parameter iteration algorithm;

[0011] When the parameters of an iteration reach the first parameter threshold, the basic processing unit is reconstructed to convert the basic processing unit from the lookup table structure into a neural network architecture.

[0012] As an improvement to the above scheme, the lookup table structure is converted into the neural network architecture after at least two reconstructions; wherein, after each reconstruction, parameter iteration is performed, and when the parameters of the second iteration reach the second parameter threshold, the next reconstruction is performed.

[0013] As an improvement to the above scheme, the first parameter iteration algorithm is an adaptive filter parameter iteration algorithm, and the second parameter iteration algorithm is a neural network parameter iteration algorithm.

[0014] As an improvement to the above scheme, the initialization operation of the basic processing unit using the first parameter iterative algorithm includes:

[0015] Obtain the data to be processed;

[0016] The first step value is calculated using the first parameter iterative algorithm and the data to be processed, and the initial iteration parameters are determined based on the first step value;

[0017] Based on the initial iteration parameters and the combination of basis function vectors of the initial GMP model, the data to be processed is quantized to obtain the first input data;

[0018] The first input data is input into the basic processing unit.

[0019] As an improvement to the above scheme, the step of iterating the parameters of the basic processing unit using a second parameter iteration algorithm based on the initial iteration parameters includes:

[0020] Obtain the data to be processed;

[0021] The second step value is calculated using the second parameter iterative algorithm and the data to be processed, and the reference iterative parameters are determined based on the initial iterative parameters and the second step value;

[0022] Based on the reference iteration parameters and the combination of basis function vectors of the GMP model after initialization, the data to be processed is quantized to obtain the second input data.

[0023] The second input data is input into the basic processing unit after the initialization operation is completed.

[0024] As an improvement to the above solution, the acquisition of the data to be processed includes:

[0025] r sets of continuous input and output data are acquired through on-chip storage, where r is a positive integer and r≥2;

[0026] Extract g sets of input and output data from r sets of continuous input and output data. Use g sets of input and output data as the data to be processed, where g is a positive integer and r>g≥1.

[0027] As an improvement to the above solution, the quantization processing of the data to be processed includes:

[0028] The modulus of the data to be processed is quantized starting from the most significant bit to obtain the quantization depth.

[0029] For each basis function vector in the basis function vector combination, multiply it by the corresponding iteration parameter to obtain a two-dimensional array with a length equal to the length of the basis function vector and a depth equal to the quantization depth, and use the two-dimensional array as input data.

[0030] As an improvement to the above scheme, after obtaining the two-dimensional array, the method further includes:

[0031] Merge the basis function terms in the two-dimensional array that have the same cross term but different orders.

[0032] As an improvement to the above scheme, the basic processing unit includes a first input terminal, a second input terminal, an output terminal, a lookup table unit, a multiplier, and an adder; wherein,

[0033] The lookup table unit includes a lookup table structure and a trigger. The output of the lookup table unit is multiplied by the input of the first input terminal using the multiplier. The output after the multiplication operation is added by the input of the second input terminal using the adder.

[0034] As an improvement to the above scheme, the reconstruction of the basic processing unit to change the basic processing unit from the lookup table structure to a neural network architecture includes:

[0035] Multi-level time-delay latching is performed on the second input terminal and the modulo input terminal of the lookup table structure;

[0036] Each level of latched data is input into the corresponding parallel unit to obtain the initial array unit; wherein, the parallel unit includes a basic processing unit and a group delay module;

[0037] Based on the initial array unit, the latched data pairs of each level are input into the three parallel initial array units, and the processing results of each initial array unit are added together to obtain the target array unit.

[0038] By changing the connection relationship of each array unit in the target array unit, a three-layer fully connected neural network architecture is obtained.

[0039] As an improvement to the above scheme, the latched data pair includes:

[0040] The raw data input at the second input terminal;

[0041] The modulus of the original data input at the modulus input terminal.

[0042] As an improvement to the above scheme, after inputting the latched data pairs of each stage into the three parallel initial array units, the method further includes:

[0043] The results are concatenated within the initial array unit of each level.

[0044] As an improvement to the above scheme, the connection relationship of each array unit in the target array unit is changed to obtain a three-layer fully connected neural network architecture, including:

[0045] Disconnect the output cascade connection of array units belonging to the same layer in the target array unit;

[0046] By fully connecting the outputs of the array units in the previous layer to the inputs of all array units in the next layer, a three-layer fully connected neural network architecture is obtained.

[0047] To achieve the above objectives, embodiments of this disclosure also provide a digital predistortion apparatus, comprising:

[0048] The first parameter iteration module is used to initialize the basic processing unit using the first parameter iteration algorithm in the initial stage of the system; wherein, the basic processing unit is configured as a lookup table structure of the GMP model, and the basic processing unit is used to output the predistortion signal;

[0049] The second parameter iteration module is used to obtain the initial iteration parameters after the initialization operation during the system stabilization phase; and to perform parameter iteration on the basic processing unit using the second parameter iteration algorithm based on the initial iteration parameters.

[0050] The predistortion correction module is used to reconstruct the basic processing unit when the parameters of a single iteration reach a first parameter threshold, so as to reconstruct the basic processing unit from the lookup table structure into a neural network architecture.

[0051] To achieve the above objectives, this disclosure also provides a digital predistortion device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the digital predistortion method as described in any of the above embodiments.

[0052] To achieve the above objectives, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the digital predistortion method as described in any of the above embodiments.

[0053] To achieve the above objectives, this disclosure also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the digital predistortion method as described in any of the above embodiments.

[0054] Compared to related technologies, the digital predistortion method, apparatus, device, storage medium, and program products disclosed in this disclosure significantly improve the convergence accuracy and speed of digital predistortion because the reconstructed neural network architecture can handle more complex input-output relationships and support diverse neural network computations. Furthermore, the reconfigurable architecture enables dynamic switching between the two architectures at the hardware level, avoiding the need to maintain parallel circuits for both the GMP model and the neural network simultaneously, thus saving circuit area. Attached Figure Description

[0055] Figure 1 is a flowchart of a digital predistortion method provided in an embodiment of this disclosure;

[0056] Figure 2 is a schematic diagram of a digital predistortion system provided in an embodiment of this disclosure;

[0057] Figure 3 is a schematic diagram of the structure of the lookup table unit provided in an embodiment of this disclosure;

[0058] Figure 4 is a schematic diagram of the structure of the basic processing unit provided in the embodiment of this disclosure;

[0059] Figure 5 is a schematic diagram of a first structure of the initial array unit provided in an embodiment of this disclosure;

[0060] Figure 6 is a schematic diagram of a second structure of the initial array unit provided in an embodiment of this disclosure;

[0061] Figure 7 is a schematic diagram of the structure of the target array unit provided in an embodiment of this disclosure;

[0062] Figure 8 is a schematic diagram of the three-layer fully connected neural network architecture provided in an embodiment of this disclosure;

[0063] Figure 9 is a structural block diagram of a digital predistortion device provided in an embodiment of this disclosure;

[0064] Figure 10 is a structural block diagram of a digital predistortion device provided in an embodiment of this disclosure. Embodiments of the present invention

[0065] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0066] Referring to Figure 1, which is a flowchart of a digital predistortion method provided in an embodiment of this disclosure, the digital predistortion method includes:

[0067] S1. In the initial stage of the system, the basic processing unit is initialized using the first parameter iterative algorithm; wherein, the basic processing unit is configured as a lookup table structure of the GMP model, and the basic processing unit is used to output the predistortion signal;

[0068] S2. During the system stabilization phase, obtain the initial iteration parameters after the initialization operation;

[0069] S3. Based on the initial iteration parameters, the basic processing unit is iterated using the second parameter iteration algorithm;

[0070] S4. When the parameters of a single iteration reach the first parameter threshold, the basic processing unit is reconstructed to convert the basic processing unit from the lookup table structure into a neural network architecture.

[0071] Referring to Figure 2, which is a schematic diagram of a digital predistortion system provided in an embodiment of this disclosure, the digital predistortion method described in this embodiment is applied to the digital predistortion system shown in Figure 2. This system includes a main control processor, SRAM (Static Random-Access Memory), a neural network parameter iteration module, an adaptive filter parameter iteration module, and a digital predistortion correction execution module. In this embodiment, these modules are used to reconstruct the GMP model architecture into a neural network architecture, thereby outputting a predistortion signal using the neural network architecture. It should be noted that the working process of these modules is explained in detail below with reference to specific operations.

[0072] For example, the first parameter iteration algorithm is an adaptive filtering parameter iteration algorithm, and the second parameter iteration algorithm is a neural network parameter iteration algorithm. During the initialization phase, the basic processing unit is configured as a lookup table structure of the GMP model. The input signal is quantized by modulus and indexed into the lookup table to output a pre-distortion signal. At this time, the basic processing unit does not have neural network functionality enabled; it is only used as a hardware implementation of the GMP model. During the initialization phase, since the power amplifier characteristics are not yet stable (e.g., the temperature has not reached steady state), directly introducing a neural network may cause training divergence due to high input data noise (e.g., dominated by thermal noise). Therefore, it is necessary to reduce the EVM (Error Vector Magnitude) to a certain value (stable correction state) through adaptive filtering. At this point, the data signal-to-noise ratio is high, and the neural network can learn distortion features more accurately. The stabilized GMP coefficients can be used as the initial weight reference for the neural network (transfer learning). After the system stabilizes, based on the initial iteration parameters of the GMP model, the second parameter iteration algorithm is used to iterate the parameters of the basic processing unit, gradually reconstructing the basic processing unit from the lookup table structure into a neural network architecture.

[0073] For example, the initial stage of the system refers to a period of time after the system is started. During this stage, the system needs to perform a series of initialization operations. This disclosure provides an example of determining whether the system is in a stable stage: observe the changes in the iterative parameters (such as the basis function coefficients of the GMP model) in the adaptive filtering parameter iterative algorithm. When, after multiple iterations, the update amount (step value) of these parameters is less than a certain set threshold, it indicates that the parameters have stabilized and no longer change significantly. At this point, the system can be considered to have entered a stable stage. This means that the adaptive filtering algorithm has brought the model parameters close to the optimal solution and can relatively stably compensate for the nonlinearity of the power amplifier.

[0074] In this embodiment, the reconstructed neural network architecture can handle more complex input-output relationships and supports diverse neural network computations, significantly improving the convergence accuracy and speed of digital predistortion. Furthermore, the reconstructable architecture enables dynamic switching between the two architectures at the hardware level, avoiding the need to maintain parallel circuits for both the GMP model and the neural network simultaneously, thus saving circuit area.

[0075] In some embodiments, in operation S1, the initialization operation of the basic processing unit using the first parameter iteration algorithm includes:

[0076] S11. Obtain the data to be processed;

[0077] S12. Calculate the first step value using the first parameter iteration algorithm and the data to be processed, and determine the initial iteration parameters based on the first step value;

[0078] S13. Based on the initial iteration parameters and the basis function vector combination of the initial GMP model, the data to be processed is quantized to obtain the first input data;

[0079] S14. Input the first input data into the basic processing unit.

[0080] It should be noted that operation S1 is executed by the adaptive filter parameter iteration module. The input of the adaptive filter parameter iteration module comes from the feedback loop of the RF transceiver chip and the real-time downlink data before digital predistortion. This module uses adaptive filtering algorithms such as LS (Least Squares) / LMS (Least Mean Squares) to iterate the data, using the data error between the feedback loop and the synchronous downlink as the error function. Through operations such as tabulation, the updated values ​​of each lookup table are output to the digital predistortion correction execution module.

[0081] In some embodiments, in operation S11, acquiring the data to be processed includes: acquiring r sets of continuous input and output data through on-chip storage (such as SRAM), where r is a positive integer and r≥2; extracting g sets of input and output data from the r sets of continuous input and output data, using the g sets of input and output data as the data to be processed, where g is a positive integer and r>g≥1.

[0082] For example, the data to be processed is I / Q data, that is, the quadrature components (in-phase and quadrature-phase) of the baseband signal, which is the core input / feedback signal of digital predistortion processing (such as the original signal before predistortion). and the feedback signal of the power amplifier output , This represents the discrete-time index, used to identify the sequence number of the signal sampling point and the step number in the iteration process. Assuming r = 16384 and g = 4096, 16384 sets of continuous I / Q data are acquired through on-chip storage. Through decimation selection, these 16384 sets of continuous data are extracted into 4096 sets of I / Q data, which are then used as the data to be processed. For example, at fixed intervals (e.g., one data point every four points), 4096 sets can be extracted from the 16384 sets of data (decimation factor = 4), preserving the main characteristics of the signal while reducing the data volume. Furthermore, the decimation operation reduces the computational load.

[0083] In some embodiments, in step S12, the first step value is calculated using the first parameter iteration algorithm and the data to be processed, and the initial iteration parameters are determined based on the first step value.

[0084] For example, this disclosure provides an example of calculating the first step value and initial iteration parameters (this is just an example and is not limited thereto): the extracted 4096 sets of data are divided into forward link signals. (The original signal before predistortion) and the feedback link signal (The power amplifier output is a sampled distorted signal), set the desired signal. , Assuming The manager envisions a linear amplification version after power amplification (or derives it through the inverse model of the predistorter); then the error function satisfies: This error function reflects the compensation error of the current predistorter for power amplifier distortion. Taking LMS as an example for the first iteration algorithm, the first step input... ,in, Let be the step size factor. The parameter update process satisfies: That is, the original iteration parameters (Model coefficients from the previous period) and the first step value Add them together to get the updated initial iteration parameters. If the initial iteration parameters are stable and convergent... .

[0085] In some embodiments, in operation S13, the quantization processing of the data to be processed includes: quantizing the modulus of the data to be processed from the most significant bit to obtain a quantization depth; multiplying each basis function vector in the basis function vector combination by the corresponding iteration parameter to obtain a two-dimensional array with a length equal to the length of the basis function vector and a depth equal to the quantization depth, and using the two-dimensional array as input data.

[0086] For example, calculate the input signal modulus The data is then quantized 8-bit from the most significant bit (all 1s) to the least significant bit (minimum non-zero value), resulting in 255 quantization intervals (0000-0001 to 1111-1111, excluding all 0s to avoid division by zero). The basis function form of the GMP model is as follows: ,in, For memory depth, It is a nonlinear order. For timing offsets, each quantization interval corresponds to a set of basis functions; each basis function requires an independent initial iteration parameter for weighting to adjust its effect on the output. The contribution of the basis functions is directly determined by the number (length) of the initial iteration parameters, and the two must correspond one-to-one. The resulting two-dimensional array structure satisfies: number of rows = 255 (the range of quantized moduli), number of columns = length of the basis function vector (determined by the product of memory depth and order). Each element in the two-dimensional array stores the calculation result of the basis function under the corresponding quantization range, which is used for fast indexing in the subsequent lookup table.

[0087] Furthermore, after obtaining the two-dimensional array, the method further includes merging the basis function terms in the two-dimensional array that have the same cross term but different orders.

[0088] For example, identical cross-items refer to memory depth. The same basis function terms. For example, , , The memory depth of these three basis function terms =1, belonging to the basis functions with "same cross term". Furthermore, the nonlinear order of these three basis function terms is... The values ​​are 1, 2, and 3 respectively, which means they satisfy the condition of having the same memory depth. Next, order There are differences. At this point, merging these three basis function terms into a single term is... The related combined terms reduce the length of the basis function vector and the number of rows in the two-dimensional array, thus lowering hardware storage costs and reducing the amount of table lookups and computations required for real-time calculations. Furthermore, terms of different orders at the same memory depth all reflect the influence of the signal on the current output, ensuring that the combined terms fully preserve the nonlinearity and memory characteristics of the signal, simplifying the model structure only through parameter integration.

[0089] In some embodiments, during operation S14, the first input data is input to the basic processing unit.

[0090] For example, the final obtained first input data (a two-dimensional array) is assigned column by column to the lookup table of the basic processing unit to complete the parameter update. It should be noted that the structure and principle of the lookup table are described in detail later.

[0091] In this embodiment, the traditional GMP model has a relatively simple and regular mathematical structure. It describes the nonlinear characteristics of the power amplifier using polynomial basis functions, which have explicit forms and low computational complexity. At system startup, this simple structure allows the model to quickly establish an initial match with the power amplifier characteristics without complex calculations or extensive data training, significantly shortening the initialization time and enabling the system to quickly enter operational mode. Furthermore, the adaptive filtering parameter iterative algorithm features fast convergence. These algorithms can quickly adjust the model parameters based on the error between the input and output signals. In the initial stage of the system, this algorithm can rapidly iteratively update the parameters of the GMP model, making the model's output as close as possible to the ideal predistortion signal, thereby quickly achieving initial compensation for the power amplifier's nonlinearity.

[0092] In some embodiments, in operation S3, the step of iterating the parameters of the basic processing unit using a second parameter iteration algorithm based on the initial iteration parameters includes:

[0093] S31. Obtain the data to be processed;

[0094] S32. Calculate the second step value using the second parameter iterative algorithm and the data to be processed, and determine the reference iterative parameters based on the initial iterative parameters and the second step value;

[0095] S33. Based on the reference iteration parameters and the combination of basis function vectors of the GMP model after the initialization operation, the data to be processed is quantized to obtain the second input data;

[0096] S34. Input the second input data into the basic processing unit after the initialization operation is completed.

[0097] It should be noted that operations S2-S3 in this embodiment are executed by the neural network parameter iteration module, and the input and output data of the neural network parameter iteration module are similar to those of the "adaptive filtering parameter iteration module". The difference is that the neural network parameter iteration module treats the lookup table structure built with the GMP model as a single-layer neural network, and uses the ADAM (Adaptive Moment Estimation) parameter iteration algorithm (i.e., the first parameter iteration algorithm) of the iterative neural network to iteratively update the existing parameters. The error function, which is used as the intermediate variable for correction iteration, is changed to the cost function of the neural network. Therefore, compared with the adaptive filtering parameter iteration module, the input data of the neural network parameter iteration module adds an existing iteration parameter (the iteration parameter initialized by the adaptive filtering algorithm).

[0098] In some embodiments, in operation S31, the same as in operation S11, 16384 sets of continuous I / Q data are acquired through on-chip storage; through decimation selection, the 16384 sets of continuous data are decimated into 4096 sets of I / Q data, and these 4096 sets of I / Q data are used as data to be processed.

[0099] In some embodiments, during operation S32, a second step value is calculated using a second parameter iteration algorithm and the data to be processed, and a reference iteration parameter is determined based on the initial iteration parameter and the second step value.

[0100] For example, this disclosure provides an example of calculating the second step value and reference iteration parameters (this is only an example and not a limitation): in addition to the 4096 sets of I / Q data (input signals) of operation S31 and feedback signals This also includes existing iterative parameters (i.e., parameters initialized using an adaptive filtering algorithm, serving as the initial weights of the neural network). The GMP model is viewed as a single-layer neural network (the input is...). The basis function vectors are used to output the predistorted signal estimate. The cost function is defined as the loss function commonly used in neural networks (such as mean squared error, MSE). ,in, =1,2,…, ; This represents the total number of samples involved in the calculation; For model output, (For feedback signals). Combined with the ADAM algorithm, the cost function is calculated with respect to the iterative parameters (i.e., the coefficient vector of the GMP model). gradient of ) Update the second step value by adding the original initial iteration parameter to the second step value to obtain the reference iteration parameter.

[0101] In some embodiments, during operation S33, the data to be processed is quantized according to the reference iteration parameters and the basis function vector combination of the GMP model after the initialization operation is completed, to obtain the second input data.

[0102] For example, the quantization process refers to operation S13 above, where the input signal is calculated. modulus The data is quantized 8-bit from the most significant bit (all 1s) to the least significant bit (minimum non-zero value), resulting in 255 quantization intervals (0000-0001 to 1111-1111, excluding all 0s to avoid division by zero). Based on the GMP model basis function vector determined in the initialization phase, the basis function combination corresponding to each quantization level is multiplied by the updated reference iteration parameters to generate a two-dimensional array. At this point, the two-dimensional array has the following dimensions: length equal to the length of the basis function vector, depth 255, and each element represents the weighting coefficient of the basis function at the corresponding quantization level.

[0103] Furthermore, after obtaining the two-dimensional array, the basis function terms in the two-dimensional array that have the same cross term but different orders can be merged.

[0104] In some embodiments, during operation S34, the second input data is input into the basic processing unit after the initialization operation has been performed.

[0105] For example, the final second input data (two-dimensional array) is assigned column by column to the lookup table of the basic processing unit to complete the parameter update.

[0106] In this embodiment, the neural network possesses highly flexible nonlinear mapping capabilities. Compared to adaptive filtering algorithms combined with GMP models, it can more accurately fit the complex nonlinear characteristics and memory effects of power amplifiers. Adaptive filtering algorithms iterate parameters based on relatively regular mathematical models, which may be inadequate for complex and variable high-order nonlinear and strong memory effect scenarios. In contrast, the neural network, through the nonlinear transformation of a large number of neurons, can capture the subtle and complex nonlinear relationships in the power amplifier characteristics, thereby generating a more accurate predistortion signal, effectively reducing signal distortion and improving signal quality. Furthermore, in digital predistortion scenarios, the distortion characteristics of the power amplifier may dynamically change with factors such as time, temperature, and input signal strength. The neural network can learn these complex features and patterns by studying a large amount of historical and real-time data. For example, it can learn the changing patterns of the power amplifier's nonlinear characteristics under different operating conditions, adaptively adjust parameters, and continuously maintain high-precision predistortion correction, which is difficult for adaptive filtering methods to achieve.

[0107] In some embodiments, during operation S4, when the parameters of a single iteration reach a first parameter threshold, the basic processing unit is reconstructed to convert the basic processing unit from the lookup table structure into a neural network architecture. Further, the lookup table structure is converted into the neural network architecture through at least two reconstructions; wherein parameter iteration is performed after each reconstruction, and the next reconstruction is performed when the parameters of the second iteration reach a second parameter threshold.

[0108] It should be noted that operation S4 is executed by the digital predistortion correction execution module. The digital predistortion correction execution module has a reconfigurable function, which can realize the architecture transformation between the GMP behavioral level model and the three-layer neural network by switching the interconnection relationship inside the module according to the command of the main control processor.

[0109] For example, after N_iter_1 iterations (the first parameter threshold), the reconfigurable function of digital predistortion correction is started, and the array unit is gradually reconstructed into a three-layer neural network. After each reconstruction, at least N_iter_2 iterations (the second parameter threshold) are performed before starting the next reconstruction, until the digital predistortion correction execution module is completely converted from the GMP model architecture to the three-layer neural network architecture.

[0110] It should be noted that digital predistortion modules in communication equipment are often implemented based on FPGAs (Field-Programmable Gate Arrays) / ASICs (Application Specific Integrated Circuits). Directly reconstructing the entire architecture may result in excessively long hardware resource reconfiguration times, affecting the continuity of communication links. Phased reconstruction (adjusting only a portion of the array units each time) can complete the architecture migration without interrupting service. The purpose of setting N_iter_1 is to allow the neural network to learn basic distortion features (such as AM / AM conversion) under the "guidance" of GMP. At this time, the GMP coefficients are fixed, and only the forward connections of the neural network are trained to avoid parameter oscillations. Each time a new layer is added during reconstruction, at least N_iter_2 iterations are required to ensure that the weights of the new layer converge to a local optimum.

[0111] In this embodiment, the three-layer structure achieves an optimal balance between computational complexity and fitting accuracy. Because the reconstructed three-layer neural network architecture can handle more complex input-output relationships and supports diverse neural network computations, it significantly improves the convergence accuracy and speed of digital predistortion. Furthermore, the reconfigurable architecture enables dynamic switching between the two architectures at the hardware level, avoiding the need to maintain parallel circuits for both the GMP model and the neural network simultaneously, thus saving circuit area.

[0112] In some embodiments, the basic processing unit includes a first input terminal, a second input terminal, an output terminal, a lookup table unit, a multiplier, and an adder; wherein, the lookup table unit includes a lookup table structure and a trigger, the output of the lookup table unit and the input of the first input terminal are multiplied by the multiplier, and the output after multiplication is added by the input of the second input terminal using the adder.

[0113] For example, referring to Figure 3, which is a schematic diagram of the structure of a lookup table unit provided in an embodiment of this disclosure, the lookup table unit LutE (LUT element) includes a lookup table structure LUT_8_8. LUT_8_8 generally refers to a lookup table structure, where "8_8" indicates that the address width is 8 bits (depth is...). =256), and the data output is 8 bits. In fields such as digital signal processing and hardware design, LUTs are used to store pre-calculated data and quickly look up the output value through the input address to improve real-time processing efficiency. In digital predistortion (DPD) systems, predistortion parameters can be stored in LUTs, and correction parameters can be quickly obtained by indexing the lookup table according to the characteristics of the input signal to achieve real-time signal correction. The input end of the lookup table unit LutE consists of two ports, input11 and input12. input11 corresponds to the input data (i.e., the lookup table address), and input12 corresponds to the multiplexing control signal before the unit delay unit. Assuming the input signal is ADDR[7:0], i.e., an 8-bit address, it is indexed in LUT_8_8. These 8 bits can represent 256 different address values, which correspond to finding 256 storage units in the lookup table, and then outputting the corresponding 8-bit data DATA[7:0]. The data output after the lookup table is buffered in the flip-flop D at the output end.

[0114] For example, the multiplexing control signal implements the hysteresis sub-model in the GMP behavioral-level model. That is, the multiplexing signal enters the unit delay unit in the hysteresis sub-model of the GMP model. In other scenarios (i.e., when the characteristics of the hysteresis sub-model in the GMP model are not needed), the bypass mode is selected. The lookup table unit LutE has a bypass control option, implemented through a selector and a unit delay unit (D flip-flop). When bypass mode is selected, the signal directly bypasses the unit delay unit into the lookup table (i.e., the "unit delay unit" circuit is removed); when bypass is not selected, the signal first passes through the unit delay unit, delaying by one clock cycle before entering the lookup table. This design allows for the introduction of signal delay based on actual needs. For example, when dealing with systems with memory effects (where the current output signal is related to historical inputs), the delay unit can be used to retain historical signal information, while in simple static scenarios, bypass can be selected to avoid unnecessary delay.

[0115] For example, referring to Figure 4, which is a schematic diagram of the structure of the basic processing unit provided in this embodiment of the present disclosure, the lookup table unit LutE, the multiplier, and the adder constitute the MALE unit (multiply-add LUT element). The unit has four inputs, two of which are the original inputs of the LutE unit; the other two are data input ports, namely the first input port input21 and the second input port input22. The input of input21 is multiplied by the output of LutE, and the input of input22 is added by the result of the multiplication.

[0116] In this embodiment of the disclosure, the MALE unit integrates multiplication, addition, and lookup table operation functions through this combination, providing the digital predistortion correction execution module with more complex processing capabilities. It can realize comprehensive calculation and processing of input data and meet the specific calculation requirements in the digital predistortion correction process.

[0117] In some embodiments, in operation S4, reconstructing the basic processing unit to reconstruct the basic processing unit from the lookup table structure into a neural network architecture includes:

[0118] S41. Perform multi-level time-delay latching processing on the second input terminal and the modulo input terminal of the lookup table structure;

[0119] S42. Input the latched data pairs of each stage into the corresponding parallel unit to obtain the initial array unit; wherein, the parallel unit includes a basic processing unit and a group delay module;

[0120] S43. Based on the initial array unit, the latched data pairs of each level are input into the three parallel initial array units, and the processing results of each initial array unit are added together to obtain the target array unit.

[0121] S44. Change the connection relationship of each array unit in the target array unit to obtain a three-layer fully connected neural network architecture.

[0122] In some embodiments, during operation S41, the latched data pair includes: the original data input at the second input terminal; and the modulus of the original data input at the modulus input terminal.

[0123] For example, referring to Figure 5, which is a schematic diagram of the first structure of the initial array unit provided in the embodiment of this disclosure, the downlink data of the digital front end (second input end) and the downlink data modulus value (input data of the modulus input end of the lookup table structure) synchronized with it are subjected to multi-level time delay latching processing through multiple flip-flops D. The latched data pair (original data and data modulus value) of each level becomes the basis for subsequent processing.

[0124] It should be noted that each branch in Figure 5 contains a group delay module, indicating that the model is designed for scenarios with memory effects, simulating the time correlation of signals through group delay. Since the lookup table unit LutE contains a multi-path sorting control signal, a bypass can be used to select whether to enable a path containing group delay. Group delay is the signal processing function within a path, while bypass is the selection mechanism between paths; the two can coexist (e.g., some branches take the group delay path, and some branches take the bypass). Figure 5 is obtained by cascading the structure in Figure 4. In this case, the input of input21 in Figure 4 is the output of the next multiplication of "group delay output + lookup table output," while input22 in Figure 4 serves as the input of the original data.

[0125] In some embodiments, during operation S42, the latched data of each stage is input to a parallel unit consisting of a set of MALE units and a group delay module. The data modulus is input to the data input port of the LutE unit for lookup table operations to retrieve pre-stored relevant parameters or mapping values. For example, the LutE outputs the corresponding weighted value based on the pre-stored parameters (such as the basis function coefficients of the GMP model) indexed by the modulus. The raw data is input to the group delay module, which adjusts the signal delay to match the power amplifier's memory depth, ensuring timing alignment between historical and current signals. After delay processing, the data is connected to the multiplier processing input port of the MALE unit. For example, the MALE unit (multiplication and accumulation unit) multiplies the weighted value output by the LutE with the raw data after group delay processing, and performs multi-stage accumulation through the addition port. The group delay module ensures the synchronization and coordination of data in timing, providing data with appropriate delay for subsequent multiplication operations. The adder is cascaded through multiple sets of parallel units (composed of MALE and group delay modules) to obtain the initial array unit; among them, the addition data port of the first set of MALE units is connected to data 0.

[0126] It should be noted that different dashed boxes in the diagram divide the processing groups into different groups. Each group contains multiple parallel processing paths as described above. These groups process the signal separately, and finally, the processing results are aggregated and output to the correction path to achieve pre-distortion correction of the input signal. The positions of the triggers D in different processing groups are different, and the positions of the triggers D correspond to different delay levels, representing different depths of tracing the historical information of the signal. For example, triggers D closer to the input have shorter delays, capturing the signal state at a more recent moment; triggers D farther from the input have longer delays, capturing signal information from a more distant moment. In this way, the model can model different degrees of memory effects in power amplifiers. Shallow memory effects (such as short-term signal correlation) can be handled with shorter delays, while strong memory effects (such as characteristic changes caused by long-term heat accumulation) require longer delays to capture, meeting the needs of accurate modeling of complex memory effect scenarios.

[0127] Further, referring to Figure 6, which is a second structural schematic diagram of the initial array unit provided in the embodiments of this disclosure, Figure 6 is a simplified version of the initial array unit. `input` represents the input of the original data, and `|input|` represents the input of the data modulus, i.e., the modulus of the original data. In Figure 6, each basic processing unit MALE and the group delay module form parallel units that are sequentially connected in series, forming a cascaded series structure.

[0128] In some embodiments, during operation S43, based on the initial array unit, the latched data pairs of each level are input into the three parallel initial array units, and the processing results of each initial array unit are added together to obtain the target array unit.

[0129] For example, referring to Figure 7, which is a schematic diagram of the target array unit provided in an embodiment of this disclosure, based on the initial array unit shown in Figure 6, each set of latched data pairs (original data and data modulus) is input into three parallel initial array units. Each initial array unit performs result cascading processing internally, and the three cascading processing results are finally added together to obtain the target array unit, thus realizing the GMP behavioral-level model correction execution circuit function. An example of the target array unit processing procedure is as follows:

[0130] 1. First level: The input data is initially processed. The timing of each unit is adjusted by the group delay module. The MALE unit performs multiplication-accumulation and lookup table operations. The processing results are passed to the second level through serial transmission.

[0131] 2. Second stage: Receives the output of the first stage, performs group delay adjustment and MALE unit operation again, further processes the data, and the results are passed in series to the third stage;

[0132] 3. Third stage: Perform final processing on the output of the second stage to complete the group delay and MALE operation of this stage.

[0133] Each level of parallel unit performs cascaded processing of results. Finally, the results of the three levels of cascaded processing are added together by an adder. The combined results of the three levels of processing realize the function of GMP behavioral model correction execution circuit, ensuring that the output signal has undergone pre-distortion correction to compensate for distortion problems in signal transmission.

[0134] In some embodiments, during operation S44, changing the connection relationship of each array unit in the target array unit to obtain a three-layer fully connected neural network architecture includes: disconnecting the output cascade connection of array units belonging to the same layer in the target array unit; and fully connecting the output of the array unit in the upper layer to the input of all array units in the lower layer to obtain a three-layer fully connected neural network architecture.

[0135] For example, referring to Figure 8, which is a schematic diagram of the three-layer fully connected neural network architecture provided in this embodiment of the present disclosure, a three-layer fully connected neural network processing array can be realized by changing the interconnection relationship between the output and input data of the array units through circuit topology mechanisms such as multiplexers. The array units with the original cascade relationship are regarded as array units in the same layer (such as the combination of "MALE + group delay" is regarded as array units in the same layer), and the original output cascade connection of the array units in the same layer is disconnected. This means breaking the previous limitation that data can only be transmitted layer by layer in a specific order, thus preparing for the construction of a fully connected structure. Then, the output of the array unit of the previous layer is fully connected to the input of all array units in the next layer (which can be input to the LutE data input terminal, multiplier input terminal, or adder input terminal according to the network function requirements), forming a three-layer fully connected neural network architecture, thereby providing complex and flexible neural network functional scalability. The output of each array unit in the upper layer is no longer limited to a specific connection, but is connected to the input of all array units in the lower layer. This connection method provides high flexibility. For example, if the output of the upper layer is connected to the data input of the lower layer LutE, the lookup table function can be used to map the data or adjust the parameters; if it is connected to the input of the multiplier, the multiplication operation of data with other parameters can be realized, enhancing the feature extraction capability; if it is connected to the input of the adder, data accumulation or fusion can be performed, enriching the computational logic of the network.

[0136] In this embodiment of the disclosure, the structure originally having a specific cascade relationship is transformed into a three-layer fully connected neural network processing array through the above operations. This fully connected approach enables the network to handle more complex input-output relationships, supports diverse neural network computations, provides an architectural foundation for realizing complex and flexible neural network functions (such as feature learning, pattern recognition, etc.), and significantly improves the convergence accuracy and convergence speed in the digital predistortion process.

[0137] Referring to Figure 9, which is a structural block diagram of a digital predistortion device 100 provided in an embodiment of this disclosure, the digital predistortion device 100 includes:

[0138] The first parameter iteration module 11 is configured to initialize the basic processing unit using the first parameter iteration algorithm during the initial stage of the system; wherein, the basic processing unit is configured as a lookup table structure of the GMP model, and the basic processing unit is used to output the predistortion signal;

[0139] The second parameter iteration module 12 is configured to, during the system stabilization phase, obtain the initial iteration parameters after the initialization operation; and perform parameter iteration on the basic processing unit using the second parameter iteration algorithm based on the initial iteration parameters.

[0140] The predistortion correction module 13 is configured to reconstruct the basic processing unit when the parameters of a single iteration reach a first parameter threshold, so as to reconstruct the basic processing unit from the lookup table structure into a neural network architecture.

[0141] In some embodiments, the first parameter iteration module 11 is configured to: acquire data to be processed; calculate a first step value using a first parameter iteration algorithm and the data to be processed, and determine initial iteration parameters based on the first step value; quantize the data to be processed according to the initial iteration parameters and the basis function vector combination of the initial GMP model to obtain first input data; and input the first input data into the basic processing unit.

[0142] In some embodiments, the second parameter iteration module 12 is configured to: acquire data to be processed; calculate a second step value using the second parameter iteration algorithm and the data to be processed, and determine reference iteration parameters based on the initial iteration parameters and the second step value; quantize the data to be processed based on the reference iteration parameters and the basis function vector combination of the GMP model after the initialization operation is completed, to obtain second input data; and input the second input data into the basic processing unit after the initialization operation is completed.

[0143] In some embodiments, the predistortion correction module 13 is configured to: perform multi-level time delay latching processing on the second input terminal and the modulus input terminal of the lookup table structure; input the latched data pairs of each level into the corresponding parallel unit to obtain an initial array unit; wherein, the parallel unit includes a basic processing unit and a group delay module; based on the initial array unit, input the latched data pairs of each level into a three-level parallel initial array unit, and add the processing results of each initial array unit to obtain a target array unit; change the connection relationship of each array unit in the target array unit to obtain a three-layer fully connected neural network architecture.

[0144] It is worth noting that the working process of each module in the digital predistortion device 100 described in this embodiment is similar to the working process of the digital predistortion method described in the above embodiment, and will not be repeated here.

[0145] Referring to Figure 10, which is a structural block diagram of a digital predistortion device 200 provided in an embodiment of this disclosure, the digital predistortion device 200 includes a processor 21, a memory 22, and a computer program stored in the memory 22 and executable on the processor 21. When the processor 21 executes the computer program, it implements the operations described in the various digital predistortion method embodiments above.

[0146] For example, the computer program may be divided into one or more modules / units, which are stored in the memory 22 and executed by the processor 21 to complete the present disclosure. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the digital predistortion device 200.

[0147] The digital predistortion device 200 may include, but is not limited to, a processor 21 and a memory 22. Those skilled in the art will understand that the schematic diagram is merely an example of the digital predistortion device 200 and does not constitute a limitation on the digital predistortion device 200. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the digital predistortion device 200 may also include input / output devices, network access devices, buses, etc.

[0148] The processor 21 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor 21 is the control center of the digital predistortion device 200, connecting all parts of the digital predistortion device 200 via various interfaces and lines.

[0149] The memory 22 can be used to store the computer programs and / or modules. The processor 21 implements various functions of the digital predistortion device 200 by running or executing the computer programs and / or modules stored in the memory 22 and calling the data stored in the memory 22. The memory 22 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0150] The modules / units integrated into the digital predistortion device 200, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above-described embodiments can also be implemented by a computer program instructing related hardware. This computer program can be stored in a computer-readable storage medium, and when executed by the processor 21, it can implement the operations of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0151] Furthermore, this disclosure also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the digital predistortion method as described in any of the above embodiments.

[0152] The above description represents the preferred embodiments of this disclosure. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this disclosure, and these improvements and modifications are also considered to be within the scope of protection of this disclosure.

Claims

1. A digital predistortion method, wherein, include: In the initial stage of the system, the basic processing unit is initialized using the first parameter iterative algorithm; wherein, the basic processing unit is configured as a lookup table structure of the GMP model, and the basic processing unit is configured to output a predistortion signal; During the system stabilization phase, the initial iteration parameters after the initialization operation are obtained; Based on the initial iteration parameters, the basic processing unit is iterated using the second parameter iteration algorithm; When the parameters of an iteration reach the first parameter threshold, the basic processing unit is reconstructed to convert the basic processing unit from the lookup table structure into a neural network architecture.

2. The digital predistortion method as described in claim 1, wherein, The lookup table structure is converted into the neural network architecture through at least two reconstructions; wherein, after each reconstruction, parameter iteration is performed, and when the parameters of the second iteration reach the second parameter threshold, the next reconstruction is performed.

3. The digital predistortion method as described in claim 1, wherein, The first parameter iteration algorithm is an adaptive filtering parameter iteration algorithm, and the second parameter iteration algorithm is a neural network parameter iteration algorithm.

4. The digital predistortion method as described in claim 1, wherein, The initialization operation of the basic processing unit using the first parameter iterative algorithm includes: Obtain the data to be processed; The first step value is calculated using the first parameter iteration algorithm and the data to be processed, and the initial iteration parameters are determined based on the first step value; Based on the initial iteration parameters and the combination of basis function vectors of the initial GMP model, the data to be processed is quantized to obtain the first input data; The first input data is input into the basic processing unit.

5. The digital predistortion method as described in claim 1, wherein, The step of iterating the parameters of the basic processing unit using a second parameter iteration algorithm based on the initial iteration parameters includes: Obtain the data to be processed; The second step value is calculated using the second parameter iterative algorithm and the data to be processed, and the reference iterative parameters are determined based on the initial iterative parameters and the second step value. Based on the reference iteration parameters and the combination of basis function vectors of the GMP model after initialization, the data to be processed is quantized to obtain the second input data. The second input data is input into the basic processing unit after the initialization operation is completed.

6. The digital predistortion method as described in claim 4 or 5, wherein, The acquisition of data to be processed includes: r sets of continuous input and output data are acquired through on-chip storage, where r is a positive integer and r≥2; Extract g sets of input and output data from r sets of continuous input and output data, and use g sets of input and output data as the data to be processed, where g is a positive integer and r>g≥1.

7. The digital predistortion method as described in claim 4 or 5, wherein, The quantization process of the data to be processed includes: The modulus of the data to be processed is quantized starting from the most significant bit to obtain the quantization depth; For each basis function vector in the basis function vector combination, multiply it by the corresponding iteration parameter to obtain a two-dimensional array with a length equal to the length of the basis function vector and a depth equal to the quantization depth, and use the two-dimensional array as input data.

8. The digital predistortion method as described in claim 7, wherein, After obtaining the two-dimensional array, the method further includes: Merge the basis function terms in the two-dimensional array that have the same cross term but different orders.

9. The digital predistortion method as described in claim 1, wherein, The basic processing unit includes a first input terminal, a second input terminal, an output terminal, a lookup table unit, a multiplier, and an adder; wherein, The lookup table unit includes a lookup table structure and a trigger. The output of the lookup table unit is multiplied with the input of the first input terminal using the multiplier. The output after the multiplication operation is added with the input of the second input terminal using the adder.

10. The digital predistortion method as described in claim 9, wherein, The reconstructing of the basic processing unit to convert the lookup table structure into a neural network architecture includes: Multi-level time-delay latching is performed on the second input terminal and the modulo input terminal of the lookup table structure; Each stage of latched data is input into the corresponding parallel unit to obtain the initial array unit; wherein, the parallel unit includes the basic processing unit and the group delay module; Based on the initial array unit, the latched data pairs of each level are input into the three parallel initial array units, and the processing results of each initial array unit are added together to obtain the target array unit. By changing the connection relationship of each array unit in the target array unit, a three-layer fully connected neural network architecture is obtained.

11. The digital predistortion method as described in claim 10, wherein, The latched data pair includes: The raw data input at the second input terminal; The modulus of the original data input at the modulus input terminal.

12. The digital predistortion method as described in claim 10, wherein, After inputting the latched data pairs of each stage into the initial array cells of the three parallel stages, the method further includes: The results are concatenated within the initial array unit of each level.

13. The digital predistortion method as described in claim 10, wherein, The modification of the connection relationship of each array unit in the target array unit to obtain a three-layer fully connected neural network architecture includes: Disconnect the output cascade connection of array units belonging to the same layer in the target array unit; By fully connecting the outputs of the array units in the previous layer to the inputs of all array units in the next layer, a three-layer fully connected neural network architecture is obtained.

14. The digital predistortion method as described in claim 9, wherein, The lookup table unit also includes bypass control options, which include a selector and a unit delay unit; When bypass mode is selected, the input signal bypasses the unit delay unit and enters the lookup table unit; When bypass mode is not selected, the input signal first passes through the unit delay unit, and enters the lookup table unit after a delay of one clock cycle.

15. A digital predistortion device, wherein, include: The first parameter iteration module is configured to initialize the basic processing unit using the first parameter iteration algorithm during the initial stage of the system; wherein, the basic processing unit is configured as a lookup table structure of the GMP model, and the basic processing unit is configured to output a predistortion signal; The second parameter iteration module is configured to, during the system stabilization phase, obtain the initial iteration parameters after the initialization operation; and, based on the initial iteration parameters, use the second parameter iteration algorithm to perform parameter iteration on the basic processing unit. The predistortion correction module is configured to reconstruct the basic processing unit when the parameters of a single iteration reach a first parameter threshold, so as to reconstruct the basic processing unit from the lookup table structure into a neural network architecture.

16. A digital predistortion device, wherein, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the digital predistortion method as described in any one of claims 1 to 14.

17. A computer-readable storage medium, wherein, The computer-readable storage medium stores a computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the digital predistortion method as described in any one of claims 1 to 14.

18. A computer program product, wherein, Includes a computer program / instruction that, when executed by a processor, implements the digital predistortion method as described in any one of claims 1 to 14.