Frequency hopping broadband unified digital pre-distortion method and system based on FiLM-GRU architecture

Abstract

The invention relates to the technical field of digital information transmission, in particular to a frequency hopping broadband unified digital pre-distortion method and system based on a FiLM-GRU architecture, and the method comprises the steps: constructing a broadband frequency hopping nonlinear data set; constructing a time sequence baseband feature vector, and extracting a center frequency value at the same time; establishing a unified pre-distortion neural network model architecture; inputting the center frequency value as a frequency condition feature into a condition modulation network, generating a modulation parameter vector, inputting the time sequence baseband feature vector into a backbone network feature, and performing dynamic modulation by using the modulation parameter vector to obtain a unified pre-distortion neural network model; and carrying out end-to-end training on the unified predistortion neural network model to obtain a frequency hopping scene-oriented digital predistorter in the target frequency band. According to the method, the problems of high resource consumption, high model complexity, weak generalization ability and poor adaptability to multi-frequency-point nonlinear dynamic characteristics in an existing frequency hopping scene are solved.

Description

Technical Field

[0001] This invention relates to the field of digital information transmission technology, and in particular to a frequency-hopping broadband unified digital predistortion method and system based on the FiLM-GRU architecture. Background Technology

[0002] In modern broadband wireless communication systems, the power amplifier is a core component of the radio frequency (RF) front-end, responsible for amplifying the RF signal to the power level required for long-distance transmission according to the expected gain. However, its inherent nonlinear characteristics can lead to signal distortion and spectral regeneration, affecting the stability, accuracy, and efficiency of the communication system. Digital predistortion technology corrects these nonlinearities by cascading an inverse model before the power amplifier. Classical digital predistortion techniques are typically optimized for single, fixed operating conditions. As modern wireless communication evolves towards higher frequencies, wider bandwidth, and greater intelligence, multi-condition dynamic scenarios have become a typical characteristic of practical communication systems.

[0003] Existing digital predistortion techniques for multiple operating conditions include lookup table methods and deep learning models. Traditional digital predistortion solutions typically employ multi-coefficient lookup table methods, such as training and storing a separate set of coefficients for each operating frequency. Theoretically, lookup table methods can pre-store predistortion coefficients for all possible operating conditions. However, as communication systems demand higher precision and coverage of higher-order memory effects, the required storage capacity grows exponentially, significantly consuming hardware resources. In recent years, although deep learning technology has made some progress in the field of multi-operating condition digital predistortion, existing models are mostly designed for single operating conditions. Currently, many offline learning methods for power amplifier predistortion models under multiple operating conditions often have complex architectures and training processes. If operating conditions such as frequency are directly used as ordinary feature inputs, simple network architectures often struggle to accurately capture the fine modulation effect of operating condition features such as frequency on the nonlinear characteristics of the power amplifier, resulting in insufficient adaptability across different operating conditions. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a frequency-hopping broadband unified digital predistortion method and system based on the FiLM-GRU architecture. It solves the problems of high resource consumption, high model complexity, weak generalization ability and poor adaptability to nonlinear dynamic characteristics of multiple frequency points in existing frequency-hopping scenarios. It achieves unified and efficient learning of nonlinear characteristics of different frequency points in the target frequency band by using a single model.

[0005] A frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture includes the following steps:

[0006] S1: Within the target operating frequency band, collect the baseband input signal sequence and output signal sequence of the power amplifier under various frequency configurations to construct a wideband frequency hopping nonlinear dataset;

[0007] S2: The input and output signal data in the wideband frequency hopping nonlinear dataset are grouped and normalized according to the center frequency. Then, a time-series baseband feature vector containing the normalized in-phase component, quadrature component, and higher power of amplitude is constructed. At the same time, the center frequency values ​​corresponding to the normalized in-phase component and quadrature component are extracted.

[0008] S3: Build a unified predistortion neural network model architecture based on FiLM-GRU, including a backbone network and a conditional modulation network;

[0009] S4: The center frequency values ​​corresponding to the normalized in-phase and quadrature components of the signal are used as frequency conditional features and input into the conditional modulation network to generate a modulation parameter vector. The timing baseband feature vector is input into the backbone network to extract the timing nonlinear features of the power amplifier signal. The modulation parameter vector is used to perform element-wise linear radiometric transformation on the state vector of the last GRU hidden layer for dynamic modulation. The hidden layer state vector after element-wise linear radiometric transformation is then fed into the output layer of the last layer of the backbone network to obtain a unified predistortion neural network model.

[0010] S5: Based on the indirect learning architecture, the input signal sequence and output signal sequence in the wideband frequency hopping nonlinear dataset are input into the unified predistortion neural network model for end-to-end training, and a digital predistorter for frequency hopping scenarios in the target frequency band is obtained.

[0011] Furthermore, step S1 constructs a wideband frequency-hopping nonlinear dataset using the following method:

[0012] S111: Within the target operating frequency band, using the collected baseband input signal sequence and output signal sequence of the power amplifier under various frequency configurations, a series of single-frequency predistortion models are first trained to obtain the data parameters of each frequency, and the data parameters of adjacent frequencies are statistically analyzed.

[0013] S112: Force the adjacent frequency data parameters to be aligned to the center frequency parameters for cross-testing, obtain the predistortion effect of the adjacent frequency data on the current center frequency predistortion model, clarify the effective frequency coverage boundary of the current center frequency predistortion model, and determine the effective coverage range of each single frequency predistortion model for the surrounding frequency data.

[0014] S113: Within the effective coverage range of surrounding frequency data for each single-frequency predistortion model, a minimum coverage set strategy is adopted to select a minimum number of center frequency combinations, extract the input and output signal data corresponding to each center frequency in the center frequency combination, and construct a power amplifier wideband frequency hopping nonlinear dataset.

[0015] In the optimized version, the higher powers of the amplitude mentioned in step S2 are the first, third, fifth, and seventh powers of the signal amplitude.

[0016] The optimized time-series baseband feature vector in step S2 is: ,in: Represents the in-phase component of the signal, Indicates the quadrature components of the signal. This indicates the signal amplitude.

[0017] Furthermore, in step S4, the center frequency values ​​corresponding to the extracted normalized in-phase and quadrature components of the signal are used as frequency conditional features input to the conditional modulation network to generate the modulation parameter vector, as follows:

[0018] S411: The frequency condition features are encoded using a sine-cosine hybrid encoding method, expanding the one-dimensional frequency value into a high-dimensional encoded vector with the mathematical expression (1):

[0019] (1);

[0020] in: Represents a high-dimensional encoded vector. This represents the center frequency values ​​corresponding to the in-phase and quadrature components of the normalized signal.

[0021] S412: Map high-dimensional encoded vectors to higher-dimensional embedding vectors through linear neural network layers and activation functions;

[0022] S413: The embedded vector is processed using a multilayer perceptron, and the output is a modulation parameter vector with the same dimension as the hidden unit of the last hidden layer of the gated recurrent unit in the backbone network.

[0023] The optimized modulation parameter vector in step S4 includes a scaling factor and a translation factor.

[0024] In the optimized step S4, the temporal nonlinear features of the power amplifier signal are extracted by inputting the time-series baseband feature vector into the backbone network and using a gated recurrent unit multi-layer hidden layer forward propagation method.

[0025] Furthermore, in step S4, the state vector of the last GRU hidden layer is dynamically modulated using an element-wise linear radiometric transformation according to equation (2) using the modulation parameter vector:

[0026] (2);

[0027] in: This represents the hidden layer state vector after element-wise linear radiometric transformation. Indicates the scaling factor. This represents the center frequency values ​​corresponding to the in-phase and quadrature components of the normalized signal. This represents the original state vector of the last hidden layer. ☉ represents the translation coefficient, and ☉ represents element-wise multiplication.

[0028] Furthermore, in step S5, end-to-end training is performed using the following method:

[0029] S511: Directly model the inverse model of the power amplifier through an indirect learning architecture;

[0030] S512: The input and output data of the switching power amplifier are used as the input and output data for training. The power amplifier inverse model is trained to minimize the error between the output signal of the power amplifier inverse model and the original input signal of the switching power amplifier.

[0031] A unified power amplifier digital predistortion system for multiple power conditions is used to execute a frequency-hopping broadband unified digital predistortion method based on the FiLM-GRU architecture as described in any of the above. It includes a power amplifier, a data acquisition module, a preprocessing module, a feature extraction module, a unified predistortion neural network model construction module, a unified predistortion neural network model dynamic modulation module, and a unified predistortion neural network model training module.

[0032] The data acquisition module is used to acquire the baseband input signal sequence and output signal sequence of the power amplifier under various frequency configurations within the target operating frequency band, and to construct a wideband frequency hopping nonlinear dataset.

[0033] The preprocessing module is used to group and normalize the input and output signal data in the wideband frequency hopping nonlinear dataset according to the center frequency.

[0034] The feature extraction module is used to construct a time-series baseband feature vector containing the normalized in-phase component, quadrature component, and higher powers of amplitude of the signal, and to extract the center frequency values ​​corresponding to the normalized in-phase component and quadrature component of the signal.

[0035] The unified predistortion neural network model building module is used to build a unified predistortion neural network model architecture based on FiLM-GRU, which includes a backbone network and a conditional modulation network.

[0036] The unified predistortion neural network model dynamic modulation module is used to input the center frequency values ​​corresponding to the extracted normalized in-phase and quadrature components of the signal as frequency condition features into the conditional modulation network to generate a modulation parameter vector. The timing baseband feature vector is input into the backbone network to extract the timing nonlinear features of the power amplifier signal. The modulation parameter vector is used to perform element-wise linear radiometric transformation on the state vector of the last GRU hidden layer for dynamic modulation. The hidden layer state vector after element-wise linear radiometric transformation is then fed into the output layer of the last layer of the backbone network to obtain the unified predistortion neural network model.

[0037] The unified predistortion neural network model training module, based on an indirect learning architecture, inputs the input signal sequence and output signal sequence from the broadband frequency hopping nonlinear dataset into the unified predistortion neural network model for end-to-end training, thereby obtaining a digital predistorter for frequency hopping scenarios within the target frequency band.

[0038] Beneficial effects of the invention:

[0039] The present invention provides a frequency-hopping wideband unified digital predistortion method and system based on the FiLM-GRU architecture, which has the following advantages:

[0040] 1. This invention breaks away from the traditional approach of modeling all frequencies one by one and searching for each frequency. First, it accurately determines the effective generalization boundary of a series of single-frequency predistortion models within the frequency band. Then, based on this boundary, a minimum coverage strategy is adopted to select a minimal combination of center frequencies that can cover the nonlinear characteristics of the power amplifier across the entire frequency band, avoiding uniform training for all frequencies. This method drastically reduces the originally dense scanning frequencies to a minimum number of backbone frequencies. Data within the same backbone frequency cluster shares statistical parameters, significantly reducing complexity and computational cost while maintaining performance.

[0041] 2. A unified digital predistortion model architecture of FiLM-GRU with "backbone-conditional" dual-stream decoupling was constructed. The extraction of the power amplifier's time-series memory nonlinear features is completely handled independently by the backbone network, and the nonlinear modulation mapping of frequency features is completely handled independently by the conditional modulation network. The two do not interfere with each other until the deep feature space, where the generated modulation vectors are then fused through explicit, channel-level affine transformations. This architecture not only effectively avoids the catastrophic forgetting phenomenon of traditional models forgetting old frequencies when learning new frequency nonlinear features, but also effectively prevents the neutralization and fuzzing of nonlinear features of different frequencies in the backbone network, ensuring high-precision convergence of the model. This allows the single model to accurately approximate the nonlinear characteristics of the power amplifier at each specific frequency across the entire frequency band, exhibiting excellent robustness and generalization ability in frequency hopping scenarios.

[0042] 3. By using a unified predistortion neural network model, the nonlinear characteristics of all frequency points within a wide bandwidth can be learned, achieving fine-grained adaptation to different frequency configurations. While effectively ensuring the accurate adjustment and capture of features, compared with the traditional lookup table method, this invention only needs to store one set of neural network parameter weights, which greatly reduces the consumption of hardware storage resources and avoids the impact of latency in model switching and reloading. It has a fast response and better matches the stringent requirements of high frequency and ultra-fast frequency hopping in anti-interference communication and modern broadband communication. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the process of this invention. Detailed Implementation

[0044] A frequency-hopping broadband unified digital predistortion method based on the FiLM-GRU architecture, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps:

[0045] S1: Within the target operating frequency band, collect the baseband input signal sequence and output signal sequence of the power amplifier under various frequency configurations to construct a wideband frequency hopping nonlinear dataset;

[0046] Since the nonlinear characteristics of power amplifiers often change gradually with frequency configuration, a predistortion model corresponding to a certain frequency often still has good predistortion performance for data within a certain frequency range around it. To ensure coverage while reducing training complexity, this invention explores the generality boundary of a single-frequency predistortion model, based on the idea of ​​data clustering-assisted predistortion.

[0047] Specifically, the following method can be used to construct a wideband frequency-hopping nonlinear dataset:

[0048] S111: Within the target operating frequency band, using the baseband input signal sequence and output signal sequence of the power amplifier under various frequency configurations, a series of single-frequency predistortion models are first trained to obtain the data parameters of each frequency, and the data parameters of adjacent frequencies are statistically analyzed to explore the universality of each frequency data to its surrounding frequency data.

[0049] S112: Force the adjacent frequency data parameters to be aligned to the center frequency parameters for cross-testing, obtain the predistortion effect of the adjacent frequency data on the current center frequency predistortion model, clarify the effective frequency coverage boundary of the current center frequency predistortion model, and determine the effective coverage range of each single frequency predistortion model for the surrounding frequency data.

[0050] The predistortion effect of adjacent frequency data on the current center frequency predistortion model can be referenced by two indicators: error vector amplitude and adjacent channel leakage ratio.

[0051] S113: Within the effective coverage range of surrounding frequency data for each single-frequency predistortion model, a minimum coverage set strategy is adopted to select a minimum number of center frequency combinations, extract the input and output signal data corresponding to each center frequency in the center frequency combination, and construct a power amplifier wideband frequency hopping nonlinear dataset.

[0052] The minimum number of center frequency combinations can be optimized to select 8 center frequencies. By splicing the power amplifier data using only these 8 selected center frequencies and adding a center frequency label column, a wideband frequency hopping nonlinear dataset of power amplifiers can be constructed, which can cover the range of (1300M-1800M). This breaks the inertia of traditional methods that model all frequencies and search them one by one, and avoids unified training for all frequencies. It sharply reduces the originally dense scanning frequency points (such as 501 in the 1300MHz-1800MHz frequency range) to at least 8 backbone frequency points. Data within the same backbone frequency cluster share statistical parameters, which greatly reduces complexity and computational cost while ensuring the effect.

[0053] S2: The input and output signal data in the wideband frequency hopping nonlinear dataset are grouped and normalized according to the center frequency. Then, a time-series baseband feature vector containing the normalized in-phase component, quadrature component, and higher power of amplitude is constructed. At the same time, the center frequency values ​​corresponding to the normalized in-phase component and quadrature component are extracted.

[0054] When grouping and normalizing the input and output signals in a wideband frequency-hopping nonlinear dataset according to the center frequency, the in-phase and quadrature components of the signal are first grouped according to the center frequency, and Z-score normalization is performed on each group to eliminate the influence of scale and accelerate the convergence of subsequent network training. However, the frequency label column is not grouped and global Z-score normalization is performed directly.

[0055] Considering the nonlinear memory effect of power amplifiers, a timing feature vector is constructed based on the normalized signal. This vector not only includes the in-phase and quadrature components of the current signal, but also the higher powers of the signal amplitude, thereby enhancing the ability to characterize the nonlinear memory effect of power amplifiers.

[0056] The optimized amplitude is represented by higher powers of the signal amplitude, which are the first, third, fifth, and seventh powers of the signal amplitude.

[0057] Specifically, the time-series baseband feature vector is ,in: Represents the in-phase component of the signal, Indicates the quadrature components of the signal. Indicates the signal amplitude.

[0058] S3: Build a unified predistortion neural network model architecture based on FiLM-GRU, including a backbone network and a conditional modulation network;

[0059] The backbone network adopts a gated recurrent unit (GRU) to receive the constructed baseband feature vector and extract the temporal nonlinear features of the signal. The conditional modulation network architecture adopts a FiLM structure to receive frequency condition features and convert the frequency, an external condition feature input, into a control signal for the backbone network. The two are fused through subsequent mechanisms.

[0060] S4: The center frequency values ​​corresponding to the normalized in-phase and quadrature components of the signal are used as frequency conditional features and input into the conditional modulation network to generate a modulation parameter vector. The timing baseband feature vector is input into the backbone network to extract the timing nonlinear features of the power amplifier signal. The modulation parameter vector is used to perform element-wise linear radiometric transformation on the state vector of the last GRU hidden layer for dynamic modulation. The hidden layer state vector after element-wise linear radiometric transformation is then fed into the output layer of the last layer of the backbone network to obtain a unified predistortion neural network model.

[0061] Furthermore, in step S4, the center frequency values ​​corresponding to the extracted normalized in-phase and quadrature components of the signal are used as frequency conditional features input to the conditional modulation network to generate the modulation parameter vector, as follows:

[0062] S411: The frequency condition features are encoded using a sine-cosine hybrid encoding method, expanding the one-dimensional frequency value into a high-dimensional encoded vector with the mathematical expression (1):

[0063] (1);

[0064] in: Represents a high-dimensional encoded vector. This represents the center frequency values ​​corresponding to the in-phase and quadrature components of the normalized signal; the specific normalization formula is: ,in This represents the original value of the center frequency corresponding to the current signal. This represents the mean of the center frequency combination. The standard deviation of the center frequency combination is represented by the normalized value. It is mapped to a numerical distribution with a mean of 0 and a variance of 1, typically in the range of [-3, 3].

[0065] By using sine and cosine hybrid coding, a one-dimensional frequency value is extended into a high-dimensional coding vector, enabling conditional modulation networks to sense frequency changes.

[0066] S412: Map high-dimensional encoded vectors to higher-dimensional embedding vectors through linear neural network layers and activation functions;

[0067] S413: The embedded vector is processed using a multilayer perceptron, and the output is a modulation parameter vector with the same dimension as the hidden unit of the last hidden layer of the gated recurrent unit in the backbone network.

[0068] The optimized modulation parameter vector in step S4 includes a scaling factor and a translation factor.

[0069] In the optimized step S4, the temporal nonlinear features of the power amplifier signal are extracted by inputting the time-series baseband feature vector into the backbone network and using a gated recurrent unit multi-layer hidden layer forward propagation method.

[0070] Furthermore, in step S4, the state vector of the last GRU hidden layer is dynamically modulated using an element-wise linear radiometric transformation according to equation (2) using the modulation parameter vector:

[0071] (2);

[0072] in: This represents the hidden layer state vector after element-wise linear radiometric transformation. Indicates the scaling factor. This represents the center frequency values ​​corresponding to the in-phase and quadrature components of the normalized signal. This represents the original state vector of the last hidden layer. ☉ represents the translation coefficient, and ☉ represents element-wise multiplication.

[0073] This transformation alters the activation distribution of the backbone network, enabling it to adapt to the specific nonlinear characteristics at the current frequency. After dynamic modulation, the backbone network continues to propagate forward through the output layer and output.

[0074] By constructing a unified digital predistortion model architecture of FiLM-GRU with "backbone-conditional" dual-stream decoupling, the extraction of the power amplifier's temporal memory nonlinear features is entirely handled independently by the backbone network, and the nonlinear modulation mapping of frequency features is entirely handled independently by the conditional modulation network. These two networks operate independently without interference until the deep feature space, where the generated modulation vectors are then fused through explicit, channel-level affine transformations. This effectively avoids the catastrophic forgetting phenomenon of traditional models forgetting old frequencies when learning new frequency nonlinear features, and prevents the neutralization and fuzzing of nonlinear features from different frequencies within the backbone network. This ensures high-precision convergence of the model, enabling this single model to accurately approximate the nonlinear characteristics of the power amplifier at specific frequencies across the entire frequency band, exhibiting excellent robustness and generalization ability in frequency-hopping scenarios.

[0075] S5: Based on the indirect learning architecture, the input signal sequence and output signal sequence in the wideband frequency hopping nonlinear dataset are input into the unified predistortion neural network model for end-to-end training, and a digital predistorter for frequency hopping scenarios in the target frequency band is obtained.

[0076] Specifically, end-to-end training can be performed using the following methods:

[0077] S511: Directly model the inverse model of the power amplifier through an indirect learning architecture;

[0078] S512: The input and output data of the switching power amplifier are used as the input and output data for training. The power amplifier inverse model is trained to minimize the error between the output signal of the power amplifier inverse model and the original input signal of the switching power amplifier.

[0079] The indirect learning architecture directly models the inverse model of the power amplifier. It trains the inverse model by exchanging the power amplifier's input and output data as the input and output data during model training. The optimization objective is to minimize the error between the model's output signal and the original power amplifier's input signal. Through training, the model learns not only the power amplifier's temporal nonlinear characteristics (represented by GRU weights) but also the intrinsic relationship between frequency conditional characteristics and temporal nonlinear characteristics (represented by FiLM network weights), enabling explicit control of the backbone network by frequency conditional characteristics. The resulting set of globally shared model weights serves as a predistorter, suitable for frequency hopping scenarios and applicable to full-band digital predistortion within the target frequency band.

[0080] This invention, through a unified predistortion neural network model, can learn the nonlinear characteristics of all frequency points within a wide bandwidth, achieving fine-grained adaptation to different frequency configurations. While effectively ensuring the accurate adjustment and capture of features, compared to the traditional lookup table method, this invention only needs to store one set of neural network parameter weights, greatly reducing hardware storage resource consumption and avoiding the time delay impact of model switching and reloading. It has a fast response and better matches the stringent requirements of high-frequency, ultra-fast frequency hopping in anti-interference communication and modern broadband communication.

[0081] A unified power amplifier digital predistortion system for multiple power conditions is used to execute a frequency-hopping broadband unified digital predistortion method based on the FiLM-GRU architecture as described in any of the above. It includes a power amplifier, a data acquisition module, a preprocessing module, a feature extraction module, a unified predistortion neural network model construction module, a unified predistortion neural network model dynamic modulation module, and a unified predistortion neural network model training module.

[0082] The data acquisition module is used to acquire the baseband input signal sequence and output signal sequence of the power amplifier under various frequency configurations within the target operating frequency band, and to construct a wideband frequency hopping nonlinear dataset.

[0083] The preprocessing module is used to group and normalize the input and output signal data in the wideband frequency hopping nonlinear dataset according to the center frequency.

[0084] The feature extraction module is used to construct a time-series baseband feature vector containing the normalized in-phase component, quadrature component, and higher powers of amplitude of the signal, and to extract the center frequency values ​​corresponding to the normalized in-phase component and quadrature component of the signal.

[0085] The unified predistortion neural network model building module is used to build a unified predistortion neural network model architecture based on FiLM-GRU, which includes a backbone network and a conditional modulation network.

[0086] The unified predistortion neural network model dynamic modulation module is used to input the center frequency values ​​corresponding to the extracted normalized in-phase and quadrature components of the signal as frequency condition features into the conditional modulation network to generate a modulation parameter vector. The timing baseband feature vector is input into the backbone network to extract the timing nonlinear features of the power amplifier signal. The modulation parameter vector is used to perform element-wise linear radiometric transformation on the state vector of the last GRU hidden layer for dynamic modulation. The hidden layer state vector after element-wise linear radiometric transformation is then fed into the output layer of the last layer of the backbone network to obtain the unified predistortion neural network model.

[0087] The unified predistortion neural network model training module, based on an indirect learning architecture, inputs the input signal sequence and output signal sequence from the broadband frequency hopping nonlinear dataset into the unified predistortion neural network model for end-to-end training, thereby obtaining a digital predistorter for frequency hopping scenarios within the target frequency band.

[0088] In a practical power amplifier system, the predistortion performance of the trained unified predistortion neural network model can be measured. Specifically, the power amplifier is configured for different frequency conditions. The original baseband input signal of the power amplifier is first processed by the forward propagation of the obtained unified predistortion neural network model to obtain the predistortion signal. Then, the obtained predistortion signal is fed into the actual power amplifier, and the quality of the final power amplifier output signal is observed and recorded (reference error vector amplitude EVM, adjacent channel leakage ratio ACLR, etc.). The application evaluation of the predistortion capability of the obtained unified predistortion neural network model for frequency hopping scenarios is then carried out.

[0089] The specific verification is as follows:

[0090] Table 1 shows the test data of the unified predistortion neural network model obtained by the method described in this invention within the 1300MHz-1800MHz frequency band under a fixed output power. In the frequency hopping scenario, the fixed power is 29.70dBm. In Table 1, "1300 / 1400" indicates that the test data is the power amplifier data corresponding to the 1300MHz frequency configuration, which belongs to the 1400MHz center frequency. "P" represents the original data index before processing by the unified predistortion neural network model; while "D" represents the operating condition linear correction data index after processing by the unified predistortion neural network model. For ease of recording, Table 1 shows the effect of one frequency every 10MHz, recording the predistortion effect of the power amplifier output signal after processing by the unified predistortion neural network model within the 1300-1800MHz frequency band.

[0091] Table 1

[0092]

[0093] Experimental results within the 1300MHz-1500MHz frequency band demonstrate that the unified predistortion neural network model obtained in this invention exhibits stable and excellent predistortion performance across all frequency configurations within the target frequency band. Throughout the entire 501MHz band, the first adjacent channel ACLR is strictly less than -45dB, and the error vector amplitude EVM is less than 1%, fully verifying the effectiveness and reliability of the unified predistortion neural network model in frequency-hopping communication applications. The construction of the unified predistortion neural network model allows power amplifier communication to move beyond the traditional fixed-frequency single-operation scenario. In frequency-hopping communication scenarios, the unified predistortion neural network model can also quickly achieve efficient predistortion of data at multiple frequency points, providing strong support for improving communication reliability.

[0094] In summary, this invention provides a frequency-hopping broadband unified digital predistortion method and system based on the FiLM-GRU architecture, targeting broadband frequency-hopping communication scenarios. It improves linearization efficiency by adaptively covering the nonlinear characteristics of the entire frequency band through a single shared unified predistortion neural network model, meeting the requirements of lightweight design and linearization. This invention constructs a minimum center frequency dataset based on general boundary exploration and designs a dynamic conditional modulation architecture based on the FiLM mechanism. This allows the backbone network to perform dynamic affine transformations on the hidden layer feature distribution according to real-time frequency changes, effectively improving the learning ability and accuracy of the unified predistortion neural network model. While ensuring excellent predistortion performance, it endows the single shared unified predistortion neural network model with outstanding adaptive, fast, and seamless switching capabilities across the entire frequency band dynamic range. This provides a breakthrough linearization solution with low resource consumption and high deployment efficiency for the RF front-end of next-generation agile frequency-hopping systems.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture, characterized in that: Includes the following steps: S1: Within the target operating frequency band, collect the baseband input signal sequence and output signal sequence of the power amplifier under various frequency configurations to construct a wideband frequency hopping nonlinear dataset; S2: The input and output signal data in the wideband frequency hopping nonlinear dataset are grouped and normalized according to the center frequency. Then, a time-series baseband feature vector containing the normalized in-phase component, quadrature component, and higher power of amplitude is constructed. At the same time, the center frequency values ​​corresponding to the normalized in-phase component and quadrature component are extracted. S3: Build a unified predistortion neural network model architecture based on FiLM-GRU, including a backbone network and a conditional modulation network; S4: The center frequency values ​​corresponding to the normalized in-phase and quadrature components of the signal are used as frequency conditional features and input into the conditional modulation network to generate a modulation parameter vector. The timing baseband feature vector is input into the backbone network to extract the timing nonlinear features of the power amplifier signal. The modulation parameter vector is used to perform element-wise linear radiometric transformation on the state vector of the last GRU hidden layer for dynamic modulation. The hidden layer state vector after element-wise linear radiometric transformation is then fed into the output layer of the last layer of the backbone network to obtain a unified predistortion neural network model. S5: Based on the indirect learning architecture, the input signal sequence and output signal sequence in the wideband frequency hopping nonlinear dataset are input into the unified predistortion neural network model for end-to-end training, and a digital predistorter for frequency hopping scenarios in the target frequency band is obtained.

2. The frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture according to claim 1, characterized in that: Step S1 uses the following method to construct a wideband frequency hopping nonlinear dataset: S111: Within the target operating frequency band, using the collected baseband input signal sequence and output signal sequence of the power amplifier under various frequency configurations, a series of single-frequency predistortion models are first trained to obtain the data parameters of each frequency, and the data parameters of adjacent frequencies are statistically analyzed. S112: Force the adjacent frequency data parameters to be aligned to the center frequency parameters for cross-testing, obtain the predistortion effect of the adjacent frequency data on the current center frequency predistortion model, clarify the effective frequency coverage boundary of the current center frequency predistortion model, and determine the effective coverage range of each single frequency predistortion model for the surrounding frequency data. S113: Within the effective coverage range of surrounding frequency data for each single-frequency predistortion model, a minimum coverage set strategy is adopted to select a minimum number of center frequency combinations, extract the input and output signal data corresponding to each center frequency in the center frequency combination, and construct a power amplifier wideband frequency hopping nonlinear dataset.

3. The frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture according to claim 1, characterized in that: The higher powers of the amplitude mentioned in step S2 are the first, third, fifth, and seventh powers of the signal amplitude.

4. The frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture according to claim 1, characterized in that: The time-series baseband feature vector mentioned in step S2 is ,in: Represents the in-phase component of the signal, Indicates the quadrature components of the signal. This indicates the signal amplitude.

5. The frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture according to claim 1, characterized in that: In step S4, the center frequency values ​​corresponding to the extracted normalized in-phase and quadrature components of the signal are used as frequency conditional features input to the conditional modulation network to generate the modulation parameter vector. The method is as follows: S411: The frequency condition features are encoded using a sine-cosine hybrid encoding method, expanding the one-dimensional frequency value into a high-dimensional encoded vector with the mathematical expression (1): （1）； in: Represents a high-dimensional encoded vector. This represents the center frequency values ​​corresponding to the in-phase and quadrature components of the normalized signal. S412: Map high-dimensional encoded vectors to higher-dimensional embedding vectors through linear neural network layers and activation functions; S413: The embedded vector is processed using a multilayer perceptron, and the output is a modulation parameter vector with the same dimension as the hidden unit of the last hidden layer of the gated recurrent unit in the backbone network.

6. The frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture according to claim 1, characterized in that: The modulation parameter vector in step S4 includes a scaling factor and a translation factor.

7. The frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture according to claim 1, characterized in that: In step S4, the timing nonlinear features of the power amplifier signal are extracted by inputting the timing baseband feature vector into the backbone network and using a gated recurrent unit multi-layer hidden layer forward propagation method.

8. The frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture according to claim 1, characterized in that: In step S4, the state vector of the last GRU hidden layer is dynamically modulated using an element-wise linear radiometric transformation according to equation (2) using the modulation parameter vector: （2）； in: This represents the hidden layer state vector after element-wise linear radiometric transformation. Indicates the scaling factor. This represents the center frequency values ​​corresponding to the in-phase and quadrature components of the normalized signal. This represents the original state vector of the last hidden layer. ☉ represents the translation coefficient, and ☉ represents element-wise multiplication.

9. The frequency-hopping wideband unified digital predistortion method based on the FiLM-GRU architecture according to claim 1, characterized in that: In step S5, end-to-end training is performed using the following method: S511: Directly model the inverse model of the power amplifier through an indirect learning architecture; S512: The input and output data of the switching power amplifier are used as the input and output data for training. The power amplifier inverse model is trained to minimize the error between the output signal of the power amplifier inverse model and the original input signal of the switching power amplifier.

10. A multi-power condition unified power amplifier digital predistortion system, used to execute a frequency-hopping broadband unified digital predistortion method based on the FiLM-GRU architecture as described in any one of claims 1 to 9, characterized in that: It includes a power amplifier, a data acquisition module, a preprocessing module, a feature extraction module, a unified predistortion neural network model construction module, a unified predistortion neural network model dynamic modulation module, and a unified predistortion neural network model training module; The data acquisition module is used to acquire the baseband input signal sequence and output signal sequence of the power amplifier under various frequency configurations within the target operating frequency band, and to construct a wideband frequency hopping nonlinear dataset. The preprocessing module is used to group and normalize the input and output signal data in the wideband frequency hopping nonlinear dataset according to the center frequency. The feature extraction module is used to construct a time-series baseband feature vector containing the normalized in-phase component, quadrature component, and higher powers of amplitude of the signal, and to extract the center frequency values ​​corresponding to the normalized in-phase component and quadrature component of the signal. The unified predistortion neural network model building module is used to build a unified predistortion neural network model architecture based on FiLM-GRU, which includes a backbone network and a conditional modulation network. The unified predistortion neural network model dynamic modulation module is used to input the center frequency values ​​corresponding to the extracted normalized in-phase and quadrature components of the signal as frequency condition features into the conditional modulation network to generate a modulation parameter vector. The timing baseband feature vector is input into the backbone network to extract the timing nonlinear features of the power amplifier signal. The modulation parameter vector is used to perform element-wise linear radiometric transformation on the state vector of the last GRU hidden layer for dynamic modulation. The hidden layer state vector after element-wise linear radiometric transformation is then fed into the output layer of the last layer of the backbone network to obtain the unified predistortion neural network model. The unified predistortion neural network model training module, based on an indirect learning architecture, inputs the input signal sequence and output signal sequence from the broadband frequency hopping nonlinear dataset into the unified predistortion neural network model for end-to-end training, thereby obtaining a digital predistorter for frequency hopping scenarios within the target frequency band.

Images