Power adaptive digital pre-distortion method, system and device based on sparse GRU and medium

Abstract

The invention discloses a power adaptive digital pre-distortion method, system and device based on sparse GRU and a medium. The method comprises the steps that input and output signals under different powers are collected, a composite feature vector is constructed through amplitude normalization and phase alignment, and a pre-distortion training target is generated in combination with indirect learning; constructing a sparse GRU neural network, introducing L1 regularization constraint, and obtaining a sparse pre-distortion model through joint optimization of mean square error and sparse constraint; partial weight updating is realized based on power grading, and weight fusion output is carried out during power interval switching so as to complete power adaptive optimization; pruning and compressing the sparse GRU model and then deploying the sparse GRU model in a pre-distortion module to compensate the nonlinearity of the power amplifier; according to the method, the model complexity is reduced through rarefaction, the dynamic working condition stability is improved through a power adaptive mechanism, pruning compression adapts to low-power-consumption hardware, high precision, low complexity and adaptive capacity are considered, and the method is suitable for the low-power-consumption and high-performance requirements of a broadband wireless communication system.

Description

Technical Field

[0001] This invention relates to the field of digital predistortion technology, and in particular to a power adaptive digital predistortion method, system, device and medium based on sparse GRU. Background Technology

[0002] Broadband wireless communication systems are crucial infrastructure for meeting the demands of high speed, large bandwidth, and low latency. The performance of their transmit link directly determines the system's stability and energy efficiency. As a key module in the transmit link, the linearity of the radio frequency power amplifier (PA) has a decisive impact on the overall system performance. However, to improve energy efficiency, modern PAs often operate in the saturation region. The resulting nonlinearity and memory effects cause severe amplitude and phase distortion, leading to spectral leakage and increased bit error rate, becoming a major bottleneck restricting system performance improvement.

[0003] Digital predistortion (DPD) technology, as an effective means of suppressing power amplifier nonlinearity, has become an important method for improving PA linearity due to its adaptability to wide bandwidth, strong compensation capability, and low implementation cost. Traditional DPD methods are mainly based on memory polynomials (MP), generalized polynomials (GMP), and Volterra structures, which are relatively mature in narrowband scenarios, but perform poorly in wideband systems due to the difficulty in effectively describing complex memory effects. In recent years, neural network models have gradually become a hot topic in DPD research due to their powerful nonlinear modeling capabilities, and can adapt to complex signal environments and various types of power amplifiers.

[0004] Existing research has shown that neural network-based digital pre-distortion (DPD) has made progress in different directions. The 1D-CNN DPD proposed by De Silva et al. achieves real-time processing in a 100 MHz broadband environment, but the model has high complexity and large hardware overhead (De Silva U, Koike-Akino T, Ma R, et al. Amodular 1D-CNN architecture for real-time digital pre-distortion[C] / / 2022IEEE Topical Conference on RF / Microwave PowerAmplifiers for Radio and Wireless Applications (PAWR).IEEE,2022:79-81.). The MP-DPD proposed by Wu et al., combined with mixed-precision quantization, maintains good linearization performance while reducing energy consumption, but does not involve structural sparsity (Wu Y, Li A, Beikmirza M, et al. Mp-dpd: Low-complexity mixed-precisionneural networks for energy-efficient digital predistortion of wideband poweramplifiers[J]. IEEE Microwave and Wireless Technology Letters, 2024, 34(6): 817-820.). Subsequently, DeltaDPD proposed by Wu et al. introduced a dynamic temporal sparsity mechanism, achieving a balance between performance and energy efficiency at 200MHz (Wu Y, Zhu Y, Qian K, et al. DeltaDPD: Exploiting Dynamic Temporal Sparsity in Recurrent Neural Networks for Energy-Efficient Wideband Digital Predistortion[J]. IEEE Microwave and Wireless Technology Letters, 2025.).Versluis et al.'s SparseDPD achieves high-ratio sparsity through spatial pruning, demonstrating hardware efficiency advantages on FPGAs (Versluis M, Wu Y, Gao C. SparseDPD: A Sparse Neural Network-based Digital Predistortion FPGA Accelerator for RF Power Amplifier Linearization[J]. arXivpreprint arXiv:2506.16591,2025.). Li et al.'s DPD-NeuralEngine ASIC accelerator, based on a 22nm chip, achieves an energy efficiency of 6.6 TOPS / W / mm² and a -45.3dBcACPR, demonstrating the practicality of neural DPD at the chip level (Li A, Wu H, Wu Y, et al. DPD-NeuralEngine: A 22-nm 6.6-TOPS / W / mm² Recurrent Neural Network Accelerator for Wideband Power Amplifier Digital Pre-Distortion[C] / / 2025 IEEE International Symposium on Circuits and Systems (ISCAS). IEEE, 2025: 1-5.). Furthermore, the DPD model based on the TCN structure proposed by Duan et al. achieves -51.6dB cACPR and -44.6dB NMSE through structural design, demonstrating the potential of convolutional models in non-recurrent structures (Duan H, Versluis M, Chen Q, et al. TCN-DPD: Parameter-Efficient Temporal Convolutional Networks for Wideband Digital Predistortion[J]. arXiv preprint arXiv:2506.12165,2025.).

[0005] However, existing technologies still have shortcomings: First, deep neural networks have complex structures, making them difficult to deploy efficiently on low-power hardware platforms; second, they lack adaptive mechanisms for dynamic changes in input power, resulting in insufficient model stability under different power conditions; and third, most research focuses on accuracy improvement or quantization compression, without systematically combining sparsity and power adaptive optimization. Therefore, developing a digital predistortion method that balances high accuracy, low complexity, and power adaptability has become a key technological requirement for the development of next-generation broadband wireless communication systems. Summary of the Invention

[0006] To overcome the problems existing in the prior art, the present invention aims to provide a power adaptive digital predistortion method, system, device, and medium based on sparse GRU. The method acquires input and output signals under different power conditions, performs amplitude normalization and phase alignment, constructs a composite feature vector, and generates a target output signal using an indirect learning structure. A sparse GRU neural network is constructed based on the input features and target output. L1 regularization is introduced during training to achieve parameter sparsity, and a sparsified model is obtained through joint optimization of mean square error and sparsity constraints. Partial weight updates are performed based on power levels, and the output results are weighted and fused when switching power ranges to achieve power adaptive optimization. The trained sparse GRU is pruned and compressed and deployed in the predistortion module to compensate for power amplifier nonlinearity and improve transmit link linearization performance.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0008] A power-adaptive digital predistortion method based on sparse GRU includes the following steps:

[0009] Step 1: Acquire broadband input signals and power amplifier output signals under different input power conditions. Perform amplitude normalization and phase alignment processing on the input and output signals respectively. Construct a composite feature vector containing the real part, imaginary part, amplitude, and phase information of the input signal. Set at least two power thresholds to divide the input power into low, medium, and high power intervals. Assign weighting coefficients to the output signals of each interval, ensuring that the weight sums to 1, and then weight and fuse them to form an intermediate output signal that characterizes the steady-state nonlinearity of the power amplifier. Using the acquired input signals and corresponding output signals, train an approximate model of the power amplifier. Derive the predistorted target output signal in reverse. Combine the predistorted target output signal with the composite feature vector to form training samples.

[0010] Step 2: Construct a sparse GRU neural network model based on the training samples constructed in Step 1. Use an indirect learning structure (ILA) to generate a predistorted target signal and use mean square error as the optimization index. Use the input signal and the predistorted target signal as supervision samples. Update the network parameters through backpropagation. Introduce regularization constraints based on the L1 norm to construct a joint optimization objective function of "mean square error loss + L1 sparsity constraint". Through iterative training, make some weight parameters approach zero to obtain a sparse predistortion model.

[0011] Step 3: Configure a subset of modeling parameters for each power range set in Step 1. During online updates, only the weight parameters corresponding to the current power range are updated first, while the other parameters remain unchanged. When the input signal switches between different power ranges, the model output results of adjacent ranges are weighted and fused, and the predicted signal is output by combining ReLU or LeakyReLU activation functions to achieve power adaptive optimization.

[0012] Step four: Perform weight pruning on the sparse GRU neural network model trained in step two, remove redundant weights whose absolute value is less than the pruning threshold, deploy the pruned sparse GRU neural network model on the digital predistortion model, update the model weights in conjunction with the power adaptive mechanism in step three, and output the predistortion signal to the power amplifier to compensate for the nonlinear distortion of the power amplifier.

[0013] The specific process of step one is as follows:

[0014] Acquire wideband input signal x(n) and power amplifier output signal y under different input power conditions. i (n), and set two power thresholds T. low T high The input power is divided into at least two power ranges to cover low power, medium power and high power ranges;

[0015] For the input signal x(n) and the output signal y i (n) Amplitude normalization and phase alignment are performed respectively to eliminate the influence of amplitude and phase; the real part, imaginary part, amplitude and phase information of the input signal x(n) are used to construct a feature vector, X n =[I[n],Q[n],A(t),cos(φ(t)),sin(φ(t))], as the input to the power amplifier model; the output signal y in each power range low (n), y mid (n), y high (n) Assign weighting coefficients α1, α2, α3 such that α1 + α2 + α3 = 1, and perform weighted fusion to form the target output signal characterizing the steady-state nonlinear characteristics of the power amplifier model:

[0016] Using the acquired input signal x(n) and the corresponding output signal y i (n), and an approximate model of the power amplifier is obtained through training. By performing reverse derivation on the input signal, the corresponding predistortion target output signal x is obtained. DPD (n), using the predistorted target output signal as the monitoring signal x DPD x(n), together with the input signal x(n), constitute training samples for constructing a digital predistortion model.

[0017] 3. The power adaptive digital predistortion method based on sparse GRU as described in claim 1, characterized in that the amplitude normalization process maps the amplitude values ​​of the input signal and the output signal to the [-1,1] interval, and the phase alignment process uses the cross-correlation method to calculate and compensate for the phase difference between the input and output signals, so that the two are phase synchronized.

[0018] The specific process of step two is as follows:

[0019] Based on the training sample data obtained in step one, the amplitude, phase, and higher-order combination features of the signal are used as input. Temporal dependency features are extracted through multi-layer GRU hidden units with residual connections. Finally, the signal is regressed to the ideal pre-distortion target—a target signal generated by an indirect learning structure (ILA) that can cancel the nonlinear distortion of the power amplifier—through a fully connected output layer, thus constructing a sparse GRU neural network model. Based on this model, the indirect learning structure (ILA) is used to generate the pre-distortion target signal, with mean square error as the optimization metric. The input and target signals are used as supervision samples, and the network parameters are updated through backpropagation. Before each round of backpropagation, L1-norm-based regularization constraints are embedded in the loss function calculation, compressing unimportant weights to zero, thereby achieving sparsity of the network parameters. L MSE The mean squared error loss is used for predistortion modeling, where λ is a sparsity hyperparameter and W is the mean squared error loss. * Indicates the weights of each layer in the network. This is a sparsity constraint term;

[0020] Constructing a "mean squared error loss + L1 sparsity constraint" means combining the mean squared error loss and the sparsity constraint term to form the optimization objective function. A sparse GRU neural network model f is obtained through iterative training. SRGRU (n); the sparse GRU neural network model f SRGRU (n) is used as a digital predistortion model, i.e., a sparse predistortion model, to compensate for the nonlinear distortion of the power amplifier.

[0021] The specific process of step three is as follows:

[0022] The predistortion model constructed in step two is updated online. During the update process, only the hidden layer weights bound to the current input power interval are updated first, including the weight parameter W related to the current power interval. *The remaining weights remain unchanged to reduce the computational overhead of the overall model update. When the input signal switches between different power ranges, the output results corresponding to adjacent power ranges are weighted and fused to achieve a smooth transition. The output is updated through the above-mentioned partial weight update and weighted fusion mechanism.

[0023]

[0024] Where φ(·) is an activation function such as ReLU; the final network output This is the prediction signal of SRGRU in DPD modeling, which enables the digital predistortion model to maintain stable nonlinear compensation performance under different input power conditions and achieves adaptive optimization.

[0025] The partial weight update in step three is as follows: each power interval corresponds to a weight subset of the sparse GRU hidden layer. During online update, the parameters of the current interval weight subset are optimized only through gradient descent. The weight coefficients of the weighted fusion processing are determined based on the Euclidean distance between the current input power and the interval boundary. The closer the distance, the larger the coefficient.

[0026] The specific process of step four is as follows:

[0027] The weight parameters of the sparse GRU neural network model trained in step two are pruned. The pruned sparse GRU neural network model is then deployed in the digital predistortion model, i.e., the sparsified predistortion model, to construct the final DPD module.

[0028] x DPD (t)=f SRGRU (x(t))

[0029] By combining the power adaptive mechanism to update the weights of the DPD module and optimize the model output, the nonlinear distortion of the power amplifier is compensated and the linearization performance of the transmit link is improved.

[0030] A power-adaptive digital predistortion system based on sparse GRU, the system comprising:

[0031] The data acquisition and feature modeling module is used to acquire input and output signals under different power conditions in step one, perform amplitude normalization and phase alignment processing, construct composite feature vectors, and generate target output signals that characterize the nonlinear characteristics of the power amplifier through an indirect learning structure, which serve as supervision data for neural network training.

[0032] The sparse GRU modeling module is used to construct a sparse gated recurrent unit neural network based on the input feature vector and the target output signal in step two. During the training process, L1 norm regularization is introduced to achieve parameter sparsity. A sparse predistortion model is obtained through joint optimization of mean square error and sparsity constraints for nonlinear compensation.

[0033] The power adaptive optimization module is used in step three to classify the power level of the input signal, update the corresponding weight parameters according to the interval, and perform weighted fusion of the output results when switching power intervals, so as to reduce computational overhead and improve the adaptive compensation performance of the predistortion model under different power conditions.

[0034] The predistortion model deployment and compensation execution module is used in step four to prune and compress the trained sparse GRU neural network model, i.e., the sparse predistortion model, to remove redundant weights to reduce computational complexity and storage overhead. The pruned model is then deployed to the digital predistortion model, i.e., the sparse predistortion model, and the input signal is processed in conjunction with the power adaptive mechanism to achieve compensation for the nonlinear distortion of the power amplifier and improvement of the linearization performance of the transmit link.

[0035] A power-adaptive digital predistortion device based on sparse GRU, comprising:

[0036] Memory, used to store computer programs;

[0037] A processor is used to implement the power adaptive digital predistortion method based on sparse GRU described in steps one to four when executing the computer program.

[0038] A computer-readable storage medium storing a computer program, which, when executed by a processor, enables the implementation of the power adaptive digital predistortion method based on sparse GRU described in steps one to four, reducing resource consumption and achieving online adaptive convergence of the predistortion model under different input powers.

[0039] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0040] 1. This invention proposes a power-adaptive digital predistortion method based on sparse GRU. By introducing L1 sparsity constraints, it effectively reduces the network parameter size and computational complexity, providing a lightweight modeling foundation for wireless transmission systems. Low complexity and low power consumption: Through L1 regularization sparsity (parameter sparsity rate reaches 0.5) and weight pruning (parameter compression rate exceeds 50%), the model computation is reduced by more than 60%, storage overhead is reduced by more than 55%, and it is compatible with low-power hardware (such as FPGA deployment, where power consumption is reduced from 1.2W to 0.5W).

[0041] 2. This invention establishes an input power grading and partial weight adaptive update mechanism. This mechanism can flexibly adjust model parameters under different power ranges and adopts a weighted fusion strategy during power switching to achieve power adaptive optimization of the digital predistortion model, improving linearization performance under dynamic operating conditions. Based on "partial weight update + range switching fusion", the model's ACPR fluctuation is less than 2dB and NMSE fluctuation is less than 1dB in low, medium, and high power ranges, solving the problem of unstable performance of traditional models under dynamic power.

[0042] 3. This invention develops sparsification and pruning compression techniques to compress the structure of sparse GRU neural network models, reducing redundant weights and storage bandwidth requirements. While ensuring modeling accuracy, it significantly improves hardware deployment efficiency and low-power adaptability. The pruned model can be directly deployed on the transmission link of existing wireless communication equipment without large-scale hardware modifications.

[0043] 4. This invention constructs a complete acquisition-modeling-training-optimization-deployment process. This process ensures modeling accuracy while balancing complexity and power consumption, enabling effective compensation for nonlinear distortion of power amplifiers in broadband wireless transmission systems. Under 200MHz OFDM signals, the ACPR suppression effect of the sparse GRU neural network model is superior to that of the traditional GRU (the ACPR on the left side is optimized from -45dBc to -52dBc, and the ACPR on the right side is optimized from -44dBc to -51dBc), and the NMSE convergence speed is improved by 30% (NMSE reaches -42dB after 100 iterations, while the traditional GRU requires 140 iterations); it has the advantages of high precision, low power consumption, and strong engineering practicality.

[0044] The proposed power-adaptive digital predistortion method based on sparse GRU can achieve nonlinear compensation accuracy comparable to that of traditional neural network DPD in a 200MHz broadband scenario. Furthermore, it can significantly reduce model complexity and hardware power consumption through sparsification and power adaptive optimization, thus meeting the requirements of low-power, high-performance wireless communication devices. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a power adaptive digital predistortion method based on sparse GRU according to an embodiment of the present invention.

[0047] Figure 2This is a schematic diagram of a power adaptive digital predistortion method based on sparse GRU according to an embodiment of the present invention.

[0048] Figure 2 The module consists of: 1. Data acquisition and feature modeling module; 2. Sparse GRU modeling module; 3. Power adaptive optimization module; 4. Predistortion model deployment and compensation execution module.

[0049] Figure 3 This is a schematic diagram comparing the performance and simulation results of a power adaptive digital predistortion method based on sparse GRU under different sparsities according to an embodiment of the present invention; wherein, Figure 3 (a) is ACPR L Performance comparison; Figure 3 (b) is ACPR R Performance comparison; Figure 3 (c) shows the performance comparison of NMSE.

[0050] Figure 4 This is a schematic diagram comparing the power spectrum of the output signal of a power adaptive digital predistortion method based on sparse GRU according to an embodiment of the present invention with that of other methods. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] To address the problems existing in the prior art, this invention provides a power adaptive digital predistortion method based on sparse GRU. The invention will be described in detail below with reference to the accompanying drawings.

[0053] like Figure 1 As shown in the figure, a power adaptive digital predistortion method based on sparse GRU according to an embodiment of the present invention specifically includes the following steps.

[0054] Step 1: Acquire broadband input signals and power amplifier output signals under different input power conditions. Perform amplitude normalization and phase alignment processing on the input and output signals respectively. Construct a composite feature vector containing the real, imaginary, amplitude, and phase information of the input signal. Set at least two power thresholds to divide the input power into low, medium, and high power intervals. Assign weighting coefficients to the output signals of each interval, ensuring the weight sums to 1, and then weight and fuse them to form an intermediate output signal characterizing the steady-state nonlinearity of the power amplifier. Using the acquired input signals and corresponding output signals, train an approximate model of the power amplifier. Derive the predistorted target output signal in reverse. Combine the predistorted target output signal and the composite feature vector to form training samples. Details are as follows:

[0055] Acquire wideband input signal x(n) and power amplifier output signal y under different input power conditions. i (n), and set two power thresholds T. low T high The input power is divided into at least two power ranges to cover low power, medium power and high power ranges;

[0056] For the input signal x(n) and the output signal y i (n) Amplitude normalization and phase alignment are performed respectively to eliminate the influence of amplitude and phase; the real part, imaginary part, amplitude and phase information of the input signal x(n) are used to construct a feature vector, X n =[I[n],Q[n],A(t),cos(φ(t)),sin(φ(t))], as the input to the power amplifier model; the output signal y in each power range low (n), y mid (n), y high (n) Assign weighting coefficients α1, α2, α3 such that α1 + α2 + α3 = 1, and perform weighted fusion to form the target output signal characterizing the steady-state nonlinear characteristics of the power amplifier model:

[0057] Using the acquired input signal x(n) and the corresponding output signal y i (n), and an approximate model of the power amplifier is obtained through training. By performing reverse derivation on the input signal, the corresponding predistortion target output signal x is obtained. DPD (n), using the predistorted target output signal as the monitoring signal x DPD x(n), together with the input signal x(n), constitute training samples for constructing a digital predistortion model.

[0058] The amplitude normalization process maps the amplitude values ​​of the input and output signals to the [-1,1] interval. The phase alignment process uses the cross-correlation method to calculate and compensate for the phase difference between the input and output signals, so that the two are in phase synchronization.

[0059] Step two: Based on the training sample data obtained in step one, the amplitude, phase, and higher-order combination features of the signal are used as input. Temporal dependency features are extracted through multi-layer GRU hidden units with residual connections. Finally, the signal is regressed to the ideal pre-distortion target through a fully connected output layer, i.e., the target signal generated by the indirect learning structure (ILA) that can cancel the nonlinear distortion of the power amplifier. The mean square error is used as the optimization index. The input signal and the target signal are used as supervision samples. The network parameters are updated through backpropagation. Before each round of backpropagation to update the network parameters, the L1 norm-based regularization constraint is embedded in the loss function calculation, and unimportant weight parameters are compressed to zero, thereby achieving sparsity of network parameters. L MSE The mean squared error loss is used for predistortion modeling, where λ is a sparsity hyperparameter (ranging from 0.001 to 0.01, with the optimal value of λ = 0.005 determined through cross-validation), and W... * Indicates the weights of each layer in the network. This is a sparsity constraint term;

[0060] Constructing a "mean squared error loss + L1 sparsity constraint" means combining the mean squared error loss and the sparsity constraint term to form the optimization objective function. A sparse GRU neural network model f is obtained through iterative training. SRGRU (n); the sparse GRU neural network model f SRGRU (n) serves as a digital predistortion model used to compensate for the nonlinear distortion of the power amplifier.

[0061] Step 3: Configure a subset of modeling parameters for each power range set in Step 1. During online updates, only the weight parameters corresponding to the current power range are updated first, while the other parameters remain unchanged. When the input signal switches between different power ranges, the model output results of adjacent ranges are weighted and fused, and the predicted signal is output by combining ReLU or Leaky ReLU activation functions to achieve power adaptive optimization.

[0062] The predistortion model constructed in step two is updated online. During the update process, only the hidden layer weights bound to the current input power interval are updated first, including the weight parameter W related to the current power interval. * The remaining weights remain unchanged to reduce the computational overhead of the overall model update. When the input signal switches between different power ranges, the output results corresponding to adjacent power ranges are weighted and fused to achieve a smooth transition. The output is updated through the above-mentioned partial weight update and weighted fusion mechanism.

[0063]

[0064] Where φ(·) is an activation function such as ReLU; the final network output This is the prediction signal of SRGRU in DPD modeling, which enables the digital predistortion model to maintain stable nonlinear compensation performance under different input power conditions and achieves adaptive optimization.

[0065] The partial weight update in step three is as follows: each power interval corresponds to a weight subset of the sparse GRU hidden layer. During online update, the parameters of the current interval weight subset are optimized only through gradient descent. The weight coefficients of the weighted fusion processing are determined based on the Euclidean distance between the current input power and the interval boundary. The closer the distance, the larger the coefficient.

[0066] Step four involves pruning the weights of the trained sparse gated recurrent unit (GRU) model, removing redundant weights whose absolute values ​​are less than the pruning threshold, and deploying the pruned sparse GRU neural network model in the digital predistortion module. The module weights are then updated using the power adaptive mechanism from step three to compensate for the nonlinear distortion of the power amplifier and improve the linearization performance of the transmit link.

[0067] The weight parameters of the sparse GRU neural network model trained in step two are pruned, and the pruned sparse GRU neural network model is deployed in the digital predistortion model to construct the final DPD module:

[0068] x DPD (t)=f SRGRU (x(t))

[0069] By combining the power adaptive mechanism to update the weights of the DPD module and optimize the output of the model, the nonlinear distortion of the power amplifier is compensated and the linearization performance of the transmit link is improved.

[0070] like Figure 2 As shown, an embodiment of the present invention provides a power adaptive digital predistortion system based on sparse GRU, comprising:

[0071] The data acquisition and feature modeling module 1 is used to acquire input and output signals under different power conditions in step one, perform amplitude normalization and phase alignment processing, construct composite feature vectors, and generate target output signals that characterize the nonlinear characteristics of the power amplifier through an indirect learning structure, which serve as supervision data for neural network training.

[0072] Sparse GRU modeling module 2 is used to construct a sparse gated recurrent unit neural network based on the input feature vector and the target output signal in step two. During the training process, L1 norm regularization is introduced to achieve parameter sparsity. A sparsity predistortion model is obtained through joint optimization of mean square error and sparsity constraints for nonlinear compensation.

[0073] The power adaptive optimization module 3 is used in step three to classify the power level of the input signal, update the corresponding weight parameters according to the interval, and perform weighted fusion of the output results when switching power intervals, so as to reduce computational overhead and improve the adaptive compensation performance of the predistortion model under different power conditions.

[0074] The predistortion model deployment and compensation execution module 4 is used to prune and compress the trained sparse GRU neural network model in step four, remove redundant weights to reduce computational complexity and storage overhead, and deploy the pruned model to the digital predistortion module. Combined with the power adaptive mechanism, the input signal is processed to achieve compensation for the nonlinear distortion of the power amplifier and improvement of the linearization performance of the transmit link.

[0075] The technical effects of the present invention will be described in detail below with reference to simulation experiments.

[0076] I. Experimental Environment and Parameter Settings

[0077] Input signal: 200MHz bandwidth OFDM signal, 1024 subcarriers, QPSK modulation mode, 120Msps symbol rate;

[0078] PA Model: A memory polynomial model (nonlinear order 5, memory depth 3) is used to simulate the nonlinearity and memory effect of a real PA, with an input power range of -20dBm to 10dBm.

[0079] The impact of sparsity on performance (see Figure 3 )

[0080] Figure 3 (a)~ Figure 3 (c) shows the ACPR and NMSE convergence curves under different sparsity ratios (0.1 / 0.3 / 0.5 / 0.7):

[0081] The best performance is achieved when the sparsity is 0.5: ACPR is -52dBc on the left and -51dBc on the right, NMSE is -46dB, and the convergence speed is the fastest (Loss stabilizes after 50 rounds).

[0082] When the sparsity ratio is >0.7, excessive pruning leads to a decrease in modeling capability, and the ACPR deteriorates to -47dBc; when the sparsity ratio is <0.3, the complexity is close to that of a normal GRU, and the power consumption is reduced by less than 10%. Verification shows that 0.5 is the optimal sparsity ratio, which matches the pruning design of this invention.

[0083] Output power spectrum comparison (see) Figure 4 )

[0084] Figure 4 Comparison of the output spectra of no DPD, ordinary GRU-DPD, and the present invention (SRGRU-sparsity=0.5):

[0085] In scenarios without DPD, out-of-band leakage is severe (-28dBc), failing to meet communication standards;

[0086] The out-of-band leakage of a standard GRU-DPD is reduced to -48dBc, but there is still significant leakage in the far-end frequency band;

[0087] This invention reduces out-of-band leakage to -52dBc, achieving optimal spectral purity and fully meeting the ACPR requirements of 5G NR (≤-45dBc).

[0088] II. Experimental Results and Analysis

[0089] Relationship between model sparsity and performance:

[0090] When the sparsity ratio (zero weight percentage) is 0.5, the model performance is optimal: ACPR -52dBc on the left and -51dBc on the right, NMSE -46dB.

[0091] When the sparsity is too high (e.g., 0.7), excessive weight pruning leads to a decrease in modeling capability, and the ACPR deteriorates to -47dBc; when the sparsity is too low (e.g., 0.1), the complexity is close to that of a normal GRU, and the power consumption is reduced by less than 10%; it is verified that 0.5 is the optimal sparsity.

[0092] The following is a comparison of the core performance indicators of the present invention (SRGRU-DPD) with existing digital predistortion (DPD) technology:

[0093]

[0094] Compared with ordinary GRU-DPD, this invention reduces parameters by 50%, power consumption by 35%, and has better ACPR and NMSE.

[0095] Compared to 1D-CNN DPD, it reduces the parameter size by 75% and power consumption by 54%, making it more practical for engineering applications.

[0096] Power adaptive performance verification:

[0097] When the input power switches from -16dBm (low power) to -14dBm (medium power), this invention achieves an ACPR fluctuation of only 1.5dB through weighted fusion.

[0098] Traditional GRU-DPD lacks an adaptive mechanism, resulting in ACPR fluctuations of up to 6 dB; this verifies the power adaptive advantage of the present invention.

[0099] III. Deployment Process

[0100] Offline phase: Acquire PA input and output signals, train sparse GRU neural network model, prune and compress the data and store the optimal parameters (including weight subsets of 3 power ranges);

[0101] Online phase:

[0102] Real-time acquisition of input signal power to determine the corresponding interval;

[0103] Only update the weight subset of the current interval to generate the predistortion signal;

[0104] Weighted fusion is performed during power switching to output a stable predistortion signal;

[0105] The predistorted signal is input to PA, which compensates for nonlinear distortion and outputs a linearized signal.

[0106] To evaluate the performance of the proposed power adaptive digital predistortion method based on sparse GRU, a series of simulation experiments were conducted. The experimental platform was implemented in Python, using a 200MHz OFDM signal as input data, and the Adjacent Channel Power Ratio (ACPR) and Normalized Mean Square Error (NMSE) were used as the main performance indicators. By training and comparing models under different sparsity rates, the comprehensive impact of the sparsity strategy on the model's accuracy and complexity was verified.

[0107] Figure 3 Figures (a)–(c) show the convergence of ACPR (left side), ACPR (right side), and NMSE (Number of Minimum Sequences) with different sparsity rates. As can be seen from the figures, all models exhibit good convergence trends. Specifically, when the sparsity rate is 0.5, the model converges faster in both ACPR and NMSE, and its final stable value is better than other sparsity rates, indicating that moderate sparsity can improve the model's generalization ability and convergence efficiency. When the sparsity rate is too high (e.g., 0.7), performance fluctuates, indicating that excessive sparsity weakens the nonlinear modeling ability.

[0108] Furthermore, Figure 4 The output spectra of No-DPD, conventional GRU, and the sparse GRU (sparse ratio of 0.5) of this invention were compared under a 200MHz OFDM signal. The results show that No-DPD results in severe out-of-band leakage. Although conventional GRU can effectively suppress distortion, strong leakage still exists in the far-end frequency band. The sparse GRU model, on the other hand, performs best in out-of-band suppression and significantly improves spectral purity, verifying its linearization advantage under low complexity conditions.

[0109] In summary, the method of this invention effectively reduces computational complexity while ensuring modeling accuracy, and shows significantly better performance than traditional methods in out-of-band distortion suppression and convergence, making it suitable for widespread application in broadband wireless transmission systems.

[0110] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0111] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A power adaptive digital predistortion method based on sparse GRU, characterized in that, Includes the following steps: Step 1: Acquire broadband input signals and power amplifier output signals under different input power conditions. Perform amplitude normalization and phase alignment processing on the input and output signals respectively. Construct a composite feature vector containing the real part, imaginary part, amplitude, and phase information of the input signal. Set at least two power thresholds to divide the input power into low, medium, and high power intervals. Assign weighting coefficients to the output signals of each interval, ensuring that the weight sums to 1, and then weight and fuse them to form an intermediate output signal that characterizes the steady-state nonlinearity of the power amplifier. Using the acquired input signals and corresponding output signals, train an approximate model of the power amplifier. Derive the predistorted target output signal in reverse. Combine the predistorted target output signal with the composite feature vector to form training samples. Step 2: Construct a sparse GRU neural network model based on the training samples constructed in Step 1. Use an indirect learning structure (ILA) to generate a predistorted target signal and use mean squared error as the optimization index. Use the input signal and the predistorted target signal as supervision samples, update the network parameters through backpropagation, introduce regularization constraints based on the L1 norm, and construct a joint optimization objective function of "mean squared error loss + L1 sparsity constraint". Through iterative training, make some weight parameters approach zero to obtain a sparse predistorted model. Step 3: Configure a subset of modeling parameters for each power range set in Step 1. During online updates, only the weight parameters corresponding to the current power range are updated first, while the other parameters remain unchanged. When the input signal switches between different power ranges, the model output results of adjacent ranges are weighted and fused, and the predicted signal is output by combining ReLU or Leaky ReLU activation functions to achieve power adaptive optimization. Step four: Perform weight pruning on the sparse GRU neural network model trained in step two, remove redundant weights whose absolute value is less than the pruning threshold, deploy the pruned sparse GRU neural network model on the digital predistortion model, update the model weights in conjunction with the power adaptive mechanism in step three, and output the predistortion signal to the power amplifier to compensate for the nonlinear distortion of the power amplifier.

2. The power adaptive digital predistortion method based on sparse GRU as described in claim 1, characterized in that, The specific process of step one is as follows: Acquire wideband input signal x(n) and power amplifier output signal y under different input power conditions. i (n), and set two power thresholds T. low T high The input power is divided into at least two power ranges to cover low power, medium power and high power ranges; For the input signal x(n) and the output signal y i (n) Amplitude normalization and phase alignment are performed respectively to eliminate the influence of amplitude and phase; the real part, imaginary part, amplitude and phase information of the input signal x(n) are used to construct a feature vector, X n =[I[n],Q[n],A(t),cos(φ(t)),sin(φ(t))], as the input to the power amplifier model; the output signal y in each power range low (n), y mid (n), y high (n) Assign weighting coefficients α1, α2, α3 such that α1 + α2 + α3 = 1, and perform weighted fusion to form the target output signal characterizing the steady-state nonlinear characteristics of the power amplifier model: Using the acquired input signal x(n) and the corresponding output signal y i (n), and an approximate model of the power amplifier is obtained through training. By performing reverse derivation on the input signal, the corresponding predistortion target output signal x is obtained. DPD (n), using the predistorted target output signal as the monitoring signal x DPD x(n), together with the input signal x(n), constitute training samples for constructing a digital predistortion model.

3. The power adaptive digital predistortion method based on sparse GRU as described in claim 1, characterized in that, The amplitude normalization process maps the amplitude values ​​of the input and output signals to the [-1,1] interval. The phase alignment process uses the cross-correlation method to calculate and compensate for the phase difference between the input and output signals, so that the two are in phase synchronization.

4. The power adaptive digital predistortion method based on sparse GRU as described in claim 1, characterized in that, The specific process of step two is as follows: Based on the training sample data obtained in step one, the amplitude, phase, and higher-order combination features of the signal are used as input. Temporal dependency features are extracted through multi-layer GRU hidden units with residual connections. Finally, the signal is regressed to the ideal pre-distortion target—a target signal generated by an indirect learning structure (ILA) that can cancel the nonlinear distortion of the power amplifier—through a fully connected output layer, thus constructing a sparse GRU neural network model. Based on this model, the indirect learning structure (ILA) is used to generate the pre-distortion target signal, with mean square error as the optimization metric. The input and target signals are used as supervision samples, and the network parameters are updated through backpropagation. Before each round of backpropagation, L1-norm-based regularization constraints are embedded in the loss function calculation, compressing unimportant weights to zero, thereby achieving sparsity of the network parameters. L MSE The mean squared error loss is used for predistortion modeling, where λ is a sparsity hyperparameter and W is the mean squared error loss. * Represents the weights of each layer in the network. This is a sparsity constraint term; Constructing a "mean squared error loss + L1 sparsity constraint" means combining the mean squared error loss with the sparsity constraint term to form the optimization objective function. A sparse GRU neural network model f is obtained through iterative training. SRGRU (n); the sparse GRU neural network model f SRGRU (n) is used as a digital predistortion model, i.e., a sparse predistortion model, to compensate for the nonlinear distortion of the power amplifier.

5. The power adaptive digital predistortion method based on sparse GRU as described in claim 1, characterized in that, The specific process of step three is as follows: The predistortion model constructed in step two is updated online. During the update process, only the hidden layer weights bound to the current input power interval are updated first, including the weight parameter W related to the current power interval. * The remaining weights remain unchanged to reduce the computational overhead of the overall model update. When the input signal switches between different power ranges, the output results corresponding to adjacent power ranges are weighted and fused to achieve a smooth transition. The output is updated through the above-mentioned partial weight update and weighted fusion mechanism. Where φ(·) is an activation function such as ReLU; the final network output This is the prediction signal of SRGRU in DPD modeling, which enables the digital predistortion model to maintain stable nonlinear compensation performance under different input power conditions and achieves adaptive optimization.

6. The power adaptive digital predistortion method based on sparse GRU as described in claim 1, characterized in that, The partial weight update in step three is as follows: each power interval corresponds to a weight subset of the hidden layer of the sparse GRU neural network model. During online update, the parameters of the current interval weight subset are optimized only through gradient descent. The weight coefficients of the weighted fusion processing are determined based on the Euclidean distance between the current input power and the interval boundary. The closer the distance, the larger the coefficient.

7. The power adaptive digital predistortion method based on sparse GRU as described in claim 1, characterized in that, The specific process of step four is as follows: The weight parameters of the sparse GRU neural network model trained in step two are pruned. The pruned sparse GRU neural network model is then deployed in the digital predistortion model, i.e., the sparsified predistortion model, to construct the final DPD module. x DPD (t)=f SRGRU (x(t)) By combining the power adaptive mechanism to update the weights of the DPD module and optimize the model output, the nonlinear distortion of the power amplifier is compensated and the linearization performance of the transmit link is improved.

8. A system for implementing the power adaptive digital predistortion method based on sparse GRU as described in any one of claims 1 to 7, characterized in that, The system includes: The data acquisition and feature modeling module is used to acquire input and output signals under different power conditions in step one, perform amplitude normalization and phase alignment processing, construct composite feature vectors, and generate target output signals that characterize the nonlinear characteristics of the power amplifier through an indirect learning structure, which serve as supervision data for neural network training. The sparse GRU modeling module is used to construct a sparse gated recurrent unit neural network based on the input feature vector and the target output signal in step two. During the training process, L1 norm regularization is introduced to achieve parameter sparsity. A sparse predistortion model is obtained through joint optimization of mean square error and sparsity constraints for nonlinear compensation. The power adaptive optimization module is used in step three to classify the power level of the input signal, update the corresponding weight parameters according to the interval, and perform weighted fusion of the output results when switching power intervals, so as to reduce computational overhead and improve the adaptive compensation performance of the predistortion model under different power conditions. The predistortion model deployment and compensation execution module is used in step four to prune and compress the trained sparse GRU neural network model, i.e., the sparse predistortion model, to remove redundant weights to reduce computational complexity and storage overhead. The pruned model is then deployed to the digital predistortion model, i.e., the sparse predistortion model, and the input signal is processed in conjunction with the power adaptive mechanism to achieve compensation for the nonlinear distortion of the power amplifier and improvement of the linearization performance of the transmit link.

9. A power adaptive digital predistortion device based on sparse GRU, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the power adaptive digital predistortion method based on sparse GRU as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement the power adaptive digital predistortion method based on sparse GRU as described in any one of claims 1 to 7, reduce resource consumption, and achieve online adaptive convergence of the predistortion model under different input powers.

Images