CN-122027408-A - FiLM-GRU architecture-based frequency hopping broadband unified digital predistortion method and system

CN122027408ACN 122027408 ACN122027408 ACN 122027408ACN-122027408-A

Abstract

The invention relates to the technical field of digital information transmission, in particular to a method and a system for unified digital predistortion of frequency hopping broadband based on FiLM-GRU architecture, wherein the method comprises the steps of constructing a broadband frequency hopping nonlinear data set; the method comprises the steps of constructing a time sequence baseband feature vector, extracting a central frequency value, constructing a unified predistortion neural network model framework, taking the central frequency value as a frequency condition feature input condition modulation network, generating a modulation parameter vector, inputting the time sequence baseband feature vector into a main network feature, dynamically modulating the time sequence baseband feature vector by using the modulation parameter vector to obtain a unified predistortion neural network model, and carrying out end-to-end training on the unified predistortion neural network model to obtain a digital predistorter facing a frequency hopping scene in a target frequency band. The method solves the problems of large resource consumption, high model complexity, weak generalization capability and poor adaptability to the nonlinear dynamic characteristics of multiple frequency points in the existing frequency hopping scene.

Inventors

ZHANG XUEBO
WANG MENGMENG
ZHOU XUDONG
WANG RUNHUA

Assignees

南开大学

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. A frequency hopping broadband unified digital predistortion method based on FiLM-GRU architecture is characterized by comprising the following steps: S1, in a target working frequency band, acquiring a baseband input signal sequence and an output signal sequence of a power amplifier under each frequency configuration, and constructing a broadband frequency hopping nonlinear data set; S2, carrying out grouping normalization processing on input signal and output signal data in the broadband frequency hopping nonlinear data set according to the center frequency, then constructing a time sequence baseband feature vector containing normalized signal inphase components, orthogonal components and amplitude high-order power, and simultaneously extracting center frequency values corresponding to the normalized signal inphase components and orthogonal components; S3, constructing a uniform predistortion neural network model framework based on FiLM-GRU and comprising a backbone network and a conditional modulation network; S4, inputting the central frequency values corresponding to the extracted in-phase components and the quadrature components of the normalized signals into a conditional modulation network as frequency condition characteristics to generate modulation parameter vectors, inputting the time sequence baseband characteristic vectors into a main network to extract time sequence nonlinear characteristics of the power amplifier signals, carrying out dynamic modulation of element-by-element linear radiation conversion on the state vector of the hidden layer of the last GRU layer by using the modulation parameter vectors, and feeding the hidden layer state vector after the element-by-element linear radiation conversion into the output layer of the last network of the main network to obtain a unified predistortion neural network model; S5, inputting the input signal sequence and the output signal sequence in the broadband frequency hopping nonlinear data set into a unified predistortion neural network model based on an indirect learning architecture, and performing end-to-end training to obtain the digital predistorter facing the frequency hopping scene in the target frequency band.
2. The method for unified digital predistortion of frequency hopping bandwidth based on FiLM-GRU architecture according to claim 1, wherein step S1 is to construct a nonlinear data set of frequency hopping of broadband by adopting the following method: S111, in a target working frequency band, training a series of single-frequency predistortion models to obtain each frequency data parameter and counting adjacent frequency data parameters by utilizing a baseband input signal sequence and an output signal sequence of an acquired power amplifier under each frequency configuration; s112, forcibly aligning adjacent frequency data parameters to central frequency parameters for cross test, obtaining the predistortion effect of the adjacent frequency data on the current central frequency predistortion model, determining the effective frequency coverage boundary of the current central frequency predistortion model, and determining the effective coverage of each single frequency predistortion model on surrounding frequency data; S113, selecting a minimum number of center frequency combinations by adopting a minimum coverage set strategy in the effective coverage area of each single-frequency predistortion model on surrounding frequency data, extracting input signal and output signal data corresponding to each center frequency in the center frequency combinations, and constructing a power amplifier broadband frequency hopping nonlinear data set.
3. The method for unified digital predistortion of a frequency hopping bandwidth based on FiLM-GRU architecture as set forth in claim 1, wherein the higher order power of magnitude in step S2 is a first order, a third order, a fifth order and a seventh order of magnitude of the signal.
4. The method for unified digital predistortion of a frequency hopping bandwidth based on FiLM-GRU architecture as set forth in claim 1, wherein the timing baseband feature vector in step S2 is Wherein: representing the in-phase component of the signal, Representing the orthogonal component of the signal, Representing the signal amplitude.
5. The method for uniformly pre-distorting the frequency hopping bandwidth based on FiLM-GRU architecture is characterized in that in step S4, the center frequency values corresponding to the extracted in-phase component and the quadrature component of the normalized signal are used as frequency condition characteristics to be input into a condition modulation network, and the method for generating the modulation parameter vector is as follows: S411, performing sine and cosine mixed coding on the frequency condition characteristics, and expanding a one-dimensional frequency value into a high-dimensional coding vector with a mathematical expression of formula (1): (1); Wherein: representing the high-dimensional encoded vector, Representing the center frequency value corresponding to the inphase component and the orthogonal component of the normalized signal; S412, mapping the high-dimensional coding vector into an embedded vector with higher dimension through a linear neural network layer and an activation function; s413, processing the embedded vector by using a multi-layer perceptron, and outputting a modulation parameter vector consistent with the dimension of the hidden unit of the last layer of hidden layer of the main network gating circulation unit.
6. The method for unified digital predistortion of a frequency hopping bandwidth based on FiLM-GRU architecture of claim 1, wherein said modulation parameter vector comprises a scaling factor and a translation factor in step S4.
7. The method of claim 1, wherein the step S4 is characterized in that the time sequence nonlinear characteristic of the power amplifier signal is extracted by inputting the time sequence baseband characteristic vector into a backbone network and adopting a mode of forward propagation of a plurality of hidden layers of a gating circulation unit layer by layer.
8. The method for unified digital predistortion of a frequency hopping broadband based on FiLM-GRU architecture according to claim 1, wherein in step S4, dynamic modulation of element-by-element linear radiation conversion is performed on the state vector of the last GRU hidden layer according to equation (2) by using a modulation parameter vector: (2); Wherein: represents the hidden layer state vector after element-by-element linear radiation conversion, The scaling factor is represented as such, Represents the center frequency value corresponding to the inphase component and the quadrature component of the normalized signal, Representing the original state vector of the last hidden layer, Representing the translation coefficient, ☉ represents the element-wise multiplication.
9. The method for unified digital predistortion of a frequency hopping broadband based on FiLM-GRU architecture according to claim 1, wherein the end-to-end training is performed in step S5 by adopting the following method: S511, directly modeling a power amplifier inverse model through an indirect learning architecture; And S512, exchanging input and output data of the power amplifier as trained output and input data, training an inverse power amplifier model, and minimizing errors between signals output by the inverse power amplifier model and original input signals of the power amplifier.
10. The digital predistortion system of the unified power amplifier of the multiple power working conditions, is used for carrying out a frequency hopping broadband unified digital predistortion method based on FiLM-GRU framework as set forth in any one of claims 1 to 9, and is characterized by comprising 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 for acquiring a baseband input signal sequence and an output signal sequence of the power amplifier under each frequency configuration in a target working frequency band to construct a broadband frequency hopping nonlinear data set; The preprocessing module is used for carrying out grouping normalization processing on the input signal and the output signal data in the broadband frequency hopping nonlinear data set according to the center frequency; The characteristic extraction module is used for constructing a time sequence baseband characteristic vector containing normalized signal inphase components, orthogonal components and amplitude high-order power, and extracting center frequency values corresponding to the normalized signal inphase components and the orthogonal components; The unified predistortion neural network model construction module is used for constructing a unified predistortion neural network model framework which comprises a backbone network and a conditional modulation network and is based on FiLM-GRU; The unified predistortion neural network model dynamic modulation module is used for taking the central frequency values corresponding to the extracted in-phase components and the orthogonal components of the normalized signals as frequency condition characteristics to be input into a condition modulation network to generate modulation parameter vectors, inputting the time sequence baseband characteristic vectors into a main network to extract time sequence nonlinear characteristics of the power amplifier signals, carrying out dynamic modulation of element-by-element linear radiation conversion on the state vector of the hidden layer of the last GRU layer by using the modulation parameter vectors, and feeding the hidden layer state vector after the element-by-element linear radiation conversion into the last network output layer of the main network to obtain a unified predistortion neural network model; the unified predistortion neural network model training module inputs an input signal sequence and an output signal sequence in the broadband frequency hopping nonlinear data set into the unified predistortion neural network model based on an indirect learning architecture, performs end-to-end training, and obtains a digital predistorter facing a frequency hopping scene in a target frequency band.

Description

FiLM-GRU architecture-based frequency hopping broadband unified digital predistortion method and system Technical Field The invention relates to the technical field of digital information transmission, in particular to a FiLM-GRU architecture-based frequency hopping broadband unified digital predistortion method and system. Background In modern broadband wireless communication systems, a power amplifier is the core of the radio frequency front-end, responsible for amplifying the radio frequency signal to the power level required for proper long-range transmission with the desired gain. However, the inherent nonlinear characteristics of the communication system can bring about problems such as signal distortion and spectrum regeneration, so that the stability and accuracy of the communication system and the communication efficiency are affected. Digital predistortion techniques correct for these non-linearities by concatenating inverse models before the power amplifier. Classical digital predistortion techniques are typically optimized for single fixed operating conditions scenarios. As modern wireless communications evolve toward high-band, wideband, and intelligent directions, multi-operating dynamic scenarios have become typical features of practical communication systems. The existing digital predistortion technology under multiple working conditions comprises a lookup table method, a deep learning model and the like. Conventional digital predistortion solutions typically employ a multi-coefficient table look-up approach, e.g., a set of coefficients are trained and stored separately for each operating frequency bin. In theory, the look-up table method can pre-store predistortion coefficients under all possible working conditions, however, as the communication system needs to cover the precision and higher-order memory effect, the required storage capacity will increase exponentially, and hardware resources are greatly consumed. In recent years, although the deep learning technology has made a certain progress in the technical field of multi-working condition digital predistortion, the existing model is designed for a single working condition. The current off-line learning methods for the predistortion models of the power amplifiers under multiple working conditions often have complex architecture and training flow, and if the working conditions such as frequency are directly input as common characteristics, the simple network architecture often has difficulty in accurately capturing the fine modulation effect of the working condition characteristics such as frequency on the nonlinear characteristics of the power amplifiers, so that the adaptability in different working conditions is insufficient. Disclosure of Invention The invention aims to solve the technical problems of large resource consumption, high model complexity, weak generalization capability and poor adaptability to nonlinear dynamic characteristics of multiple frequency points in the existing frequency hopping scene, and the unified and efficient learning of nonlinear characteristics of different frequency points in a target frequency band is realized by adopting a single model. A frequency hopping broadband unified digital predistortion method based on FiLM-GRU architecture comprises the following steps: S1, in a target working frequency band, acquiring a baseband input signal sequence and an output signal sequence of a power amplifier under each frequency configuration, and constructing a broadband frequency hopping nonlinear data set; S2, carrying out grouping normalization processing on input signal and output signal data in the broadband frequency hopping nonlinear data set according to the center frequency, then constructing a time sequence baseband feature vector containing normalized signal inphase components, orthogonal components and amplitude high-order power, and simultaneously extracting center frequency values corresponding to the normalized signal inphase components and orthogonal components; S3, constructing a uniform predistortion neural network model framework based on FiLM-GRU and comprising a backbone network and a conditional modulation network; S4, inputting the central frequency values corresponding to the extracted in-phase components and the quadrature components of the normalized signals into a conditional modulation network as frequency condition characteristics to generate modulation parameter vectors, inputting the time sequence baseband characteristic vectors into a main network to extract time sequence nonlinear characteristics of the power amplifier signals, carrying out dynamic modulation of element-by-element linear radiation conversion on the state vector of the hidden layer of the last GRU layer by using the modulation parameter vectors, and feeding the hidden layer state vector after the element-by-element linear radiation conversion into the output layer of the last network of the main network to obtain