Deep learning driven power amplifier predistortion method

Abstract

The invention belongs to the technical field of signal processing, and particularly discloses a deep learning driven power amplifier predistortion method, which comprises the following steps: constructing a signal matrix of a target signal sequence, the target signal sequence being obtained by preprocessing an original time signal sequence collected from a signal source; sequentially inputting the signal matrix into a PDTRANS predistortion model and a power amplifier, and obtaining an output signal sequence of the power amplifier; determining a loss function value based on the output signal sequence and the target signal sequence, and adjusting hyper-parameters of a PDTRANS pre-distortion model based on the loss function value; the steps of obtaining the output signal sequence of the power amplifier and adjusting the hyper-parameters of the PDTRANS predistortion model are repeated until the loss function is minimized, and the trained PDTRANS predistortion model matched with the power amplifier is obtained; and performing pre-distortion compensation on the power amplifier based on the trained PDTRANS pre-distortion model. According to the invention, accurate pre-distortion modeling of the power amplifier is realized.

Description

Technical Field

[0001] The present application belongs to the field of signal processing technology, and more specifically, relates to a power amplifier predistortion method driven by deep learning. Background Art

[0002] The power amplifier (PA) can amplify the power of the input signal to enhance the signal strength and coverage capability. The nonlinear characteristics of the PA will cause nonlinear distortion of the signal, and the memory effect will further deteriorate the linearity of the PA and reduce the quality of the output signal sequence. By building an accurate mathematical model through behavioral modeling and digital pre-distortion technology to track the nonlinear characteristics and memory effects of the PA, and correcting the signal through a digital pre-distorter, the output distortion can be effectively reduced and the signal quality can be improved. Traditional digital pre-distortion is mostly based on polynomial models or lookup table models. When dealing with PAs with strong nonlinearity and memory effects, it is often difficult to achieve an ideal linearization effect.

[0003] In recent years, neural networks have been widely used in various fields due to their powerful nonlinear fitting and global approximation capabilities. In the related technology, a RF channel modeling and pre-distortion method based on an extended real-valued time-delayed neural network (RVTDNN) is proposed. The nonlinear characteristics and memory effects in the RF channel are captured through a deep neural network, showing a certain modeling accuracy and robustness. Furthermore, a modeling method based on a graph convolutional neural network (GCN) is proposed, which converts the one-dimensional time series into a two-dimensional graph structure for processing, making full use of the time correlation and local characteristics of the signal, significantly improving the accuracy and performance of the pre-distorter modeling, and providing a new idea for the modeling of complex nonlinear systems.

[0004] Although various neural networks have demonstrated certain advantages in power amplifier modeling and pre-distortion processing, low-complexity neural networks often find it difficult to achieve ideal results when faced with complex and high-dimensional data processing. Summary of the invention

[0005] In view of the above-mentioned defects in the prior art, the present application provides a deep learning driven power amplifier predistortion method, which aims to solve the problem of poor performance of power amplifier predistortion processing based on neural network when facing complex and high-dimensional data processing.

[0006] In a first aspect, the present application provides a deep learning driven power amplifier predistortion method, comprising: Constructing a signal matrix of a target signal sequence, where the target signal sequence is obtained by preprocessing an original time signal sequence collected from a signal source; The signal matrix is ​​sequentially input into the PDTRANS predistortion model and the power amplifier to obtain the output signal sequence of the power amplifier; Determine a loss function value based on the output signal sequence and the target signal sequence, and adjust a hyperparameter of the PDTRANS predistortion model based on the loss function value; Repeat the steps of obtaining the output signal sequence of the power amplifier and adjusting the hyperparameters of the PDTRANS pre-distortion model until the loss function is minimized, thereby obtaining a trained PDTRANS pre-distortion model that matches the power amplifier; The power amplifier is predistorted and compensated based on the trained PDTRANS predistortion model.

[0007] In a second aspect, the present application also provides a deep learning driven power amplifier predistortion device, comprising: A matrix construction module, used for constructing a signal matrix of a target signal sequence, where the target signal sequence is obtained by preprocessing an original time signal sequence collected from a signal source; A model training module is used to input the signal matrix into the PDTRANS predistortion model and the power amplifier in sequence to obtain an output signal sequence of the power amplifier; determine a loss function value based on the output signal sequence and the target signal sequence, and adjust the hyperparameters of the PDTRANS predistortion model based on the loss function value; repeat the steps of obtaining the output signal sequence of the power amplifier and adjusting the hyperparameters of the PDTRANS predistortion model until the loss function is minimized, thereby obtaining a trained PDTRANS predistortion model that matches the power amplifier; The pre-distortion compensation module is used to perform pre-distortion compensation on the power amplifier based on the trained PDTRANS pre-distortion model.

[0008] In a third aspect, the present application also provides an electronic device, comprising: at least one memory for storing programs; and at least one processor for executing the programs stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method described in the first aspect or any possible implementation of the first aspect.

[0009] In a fourth aspect, the present application also provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method described in the first aspect or any possible implementation of the first aspect.

[0010] In a fifth aspect, the present application further provides a computer program product, which, when executed on a processor, enables the processor to execute the method described in the first aspect or any possible implementation manner of the first aspect.

[0011] The present application provides a deep learning-driven power amplifier predistortion method, which converts a one-dimensional time series signal into a two-dimensional signal matrix as a model input, so that the model can capture the dynamic changes of the time signal sequence and explain the complex interactive relationship within the signal; a Transformer-based predistortion model is used to perform power amplifier predistortion modeling, and the input signal is independently analyzed through a self-attention mechanism, and the correlation between different features is calculated, so as to deeply mine the complex dynamic behavior in the signal and more accurately model the nonlinear characteristics and memory effect of the power amplifier; the model training adopts a direct learning method, which can quickly adapt to the actual signal environment and shorten the predistortion modeling process of the power amplifier. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the present application or related technologies, the following is a brief introduction to the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0013] Figure 1 It is one of the flow charts of the deep learning driven power amplifier predistortion method provided in the embodiment of the present application; Figure 2 This is the second flow chart of the deep learning driven power amplifier predistortion method provided in the embodiment of the present application; Figure 3 It is a schematic diagram of a process of training a PDTRANS pre-distortion model based on a direct learning method provided in an embodiment of the present application; Figure 4 Schematic diagram of the architecture of the PDTRANS pre-distortion model provided in an embodiment of the present application; Figure 5 It is a structural schematic diagram of a power amplifier predistortion device driven by deep learning provided in an embodiment of the present application; Figure 6 It is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0015] When faced with complex and high-dimensional data processing, the Transformer model shows significant superiority in processing sequence data, self-attention mechanism and parallel computing. Therefore, this application considers introducing the Transformer into the power amplifier pre-distortion modeling and proposes a Transformer-based pre-distortion model (Pre-Distortion Transformer, PDTRANS).

[0016] Figure 1 is one of the flow charts of the deep learning driven power amplifier predistortion method provided in the embodiment of the present application, Figure 2 This is a second flow chart of the deep learning driven power amplifier predistortion method provided in the embodiment of the present application, combined with Figure 1 and Figure 2 , the method comprises at least the following steps: S101 . Construct a signal matrix of a target signal sequence, where the target signal sequence is obtained by preprocessing an original time signal sequence collected from a signal source.

[0017] S102: Input the signal matrix to the PDTRANS predistortion model and the power amplifier in sequence to obtain an output signal sequence of the power amplifier.

[0018] S103 . Determine a loss function value based on the output signal sequence and the target signal sequence, and adjust hyperparameters of the PDTRANS predistortion model based on the loss function value.

[0019] S104, repeating the steps of obtaining the output signal sequence of the power amplifier and adjusting the parameters of the PDTRANS pre-distortion model until the loss function is minimized, thereby obtaining a trained PDTRANS pre-distortion model that matches the power amplifier.

[0020] S105 . Perform predistortion compensation on the power amplifier based on the trained PDTRANS predistortion model.

[0021] The deep learning-driven power amplifier predistortion method provided in the embodiment of the present application converts a one-dimensional time series signal into a two-dimensional signal matrix as a model input, so that the model can capture the dynamic changes of the time signal sequence and explain the complex interactive relationship within the signal; a Transformer-based predistortion model is used to perform power amplifier predistortion modeling, and the input signal is independently analyzed through a self-attention mechanism, and the correlation between different features is calculated, so as to deeply mine the complex dynamic behavior in the signal and more accurately model the nonlinear characteristics and memory effects of the power amplifier; the model training adopts a direct learning method, which can quickly adapt to the actual signal environment and shorten the predistortion modeling process of the power amplifier.

[0022] For S101, a signal matrix of a target signal sequence is constructed as an input of a subsequent power amplifier predistortion model.

[0023] Preprocess the original time signal sequence to obtain the target signal sequence. The time signal sequence is a signal that is arranged in order in time, which provides a basis for the subsequent capture of the memory effect and nonlinear characteristics of the power amplifier. The main purpose of preprocessing is to remove unnecessary noise and interference, ensure the effectiveness and stability of the signal, and provide a high-quality data basis for subsequent power amplifier predistortion modeling and application.

[0024] Optionally, the preprocessing method includes filtering, such as Figure 2 The original time signal sequence is filtered by a bandpass filter as shown in Optionally, the preprocessing method also includes normalization processing to ensure that each sample signal in the constructed signal matrix has the same scale and to ensure that the signal amplitude is within a dynamic range suitable for the subsequent predistortion model.

[0025] Optionally, the original time signal sequence is sampled using a sliding window method, and the sequence length is the sliding window length. Sliding window sampling can convert long time series signals into short sequences, which is convenient for subsequent analysis, feature extraction, and model training, so that the model can be quickly trained under conditions of smaller data sets and limited hardware resources.

[0026] Furthermore, a signal matrix of the target signal sequence is constructed. The time signal sequence is one-dimensional data. By constructing the signal matrix of the target signal sequence, the one-dimensional time signal sequence is converted into a two-dimensional format, and the time series signal is converted into a form that is easy to process and analyze, so that the subsequent pre-distortion model can capture the dynamic changes of the time signal sequence and reveal the complex interactive relationship within the signal.

[0027] Optionally, S101 specifically includes: Extract the I / Q component characteristics and amplitude envelope characteristics of the sample signal in the target signal sequence at different times; The signal matrix of the target signal sequence is constructed with the number of extracted features and sequence length as dimensions.

[0028] Specifically, a signal matrix of the target signal sequence is constructed, which can convert the target signal sequence into a form that is easy to process and analyze. The dimensions of the signal matrix are the sequence length and the number of features. Feature pre-extraction, for example, pre-extracts features such as the frequency, phase, and amplitude of the signal, which can specifically be the I / Q component features and amplitude envelope features of the sample signal at different times.

[0029] I / Q components refer to the performance of the signal in the in-phase and quadrature components. By decomposing the signal into complex form in a rectangular coordinate system, the real part is the I component and the imaginary part is the Q component; in polar coordinates, it can be expressed by amplitude and phase. The amplitude envelope refers to the curve of the signal amplitude changing over time, which can reflect the change trend and fluctuation range of the signal.

[0030] For example, construct a signal matrix for the target signal sequence as follows:

[0031] in, represents the I component of the sample signal at time n, represents the Q component of the sample signal at time n, represents the amplitude envelope of the sample signal at time n, Indicates the memory depth, Corresponding to the length of the target signal sequence, that is, the number of sample signals.

[0032] For S102 to S104, a direct learning method is used to perform power amplifier pre-distortion modeling and training.

[0033] Figure 3 FIG. 1 is a flow chart of training a PDTRANS predistortion model based on a direct learning method according to an embodiment of the present application. Figure 3 As shown, at time n, the input of the PDTRANS predistortion model is the preprocessed target signal , the output is the predistortion signal The predistortion signal output by the PDTRANS predistortion model Input the power amplifier for pre-distortion compensation and obtain the output signal of the power amplifier .

[0034] The loss function value at time n is calculated using the output signal of the power amplifier and the target signal of the input PDTRANS predistortion model. The loss function is used to characterize the difference between the output signal sequence of the power amplifier and the target signal sequence. Specifically, the mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), etc. are used. The loss function value is further used to adjust the hyperparameters of the PDTRANS predistortion model.

[0035] The hyperparameters of the PDTRANS predistortion model are continuously adjusted using the loss function value until the loss function is minimized (for example, the loss function value is 0 or reaches a preset threshold close to 0), thereby obtaining a trained PDTRANS predistortion model that matches the power amplifier. After the PDTRANS predistortion model reaches convergence, the input signal of the power amplifier should be as close as possible to the target signal of the input PDTRANS predistortion model, thereby achieving predistortion compensation for the power amplifier.

[0036] The direct learning mode is simple and efficient, can quickly adapt to the actual signal environment, and shorten the pre-distortion modeling process of the power amplifier.

[0037] Figure 4 Schematic diagram of the architecture of the PDTRANS pre-distortion model provided in the embodiment of the present application, such as Figure 4 As shown, the PDTRANS predistortion model includes: (1) Input module The input module includes a convolution (Conv) layer, which is designed to perform preliminary feature extraction on the input signal matrix. The input signal matrix undergoes a 3×3 convolution operation to extract local time series features, including I / Q component features and amplitude envelope features. The convolution operation enhances the model's understanding of instantaneous characteristics and local patterns, laying the foundation for subsequent modules to capture the nonlinear characteristics of the signal.

[0038] (2) Transformer module The Transformer module is the core part of the PDTRANS pre-distortion model. Through the self-attention mechanism and parallel computing capabilities, it is responsible for modeling the complex nonlinear characteristics and memory effects of the signal and adapting to diverse signal characteristics and environmental requirements.

[0039] The Transformer module specifically includes an improved normalization layer, a parallel convolution layer, a self-attention layer, a short-circuit projection layer, and an output projection layer.

[0040] The Revised Layer Norm normalizes the input features (the extracted features output by the input module) to reduce the impact of signal amplitude changes on model performance and improve training stability.

[0041] Three parallel convolutional layers generate key features, query features, and value features based on the normalized input features:

[0042] in, represent query, key, and value features respectively, , , They are used to generate learnable weight matrices for the three features, Represents the normalized signal matrix. The query feature is used to calculate the attention weight, the key feature is used to calculate the attention weight together with the query feature, and the value feature is used to weight the attention output.

[0043] The self-attention layer uses the self-attention mechanism to model the dynamic dependencies of the signal and determines the self-attention results based on the key features, query features, and value features to meet the following requirements:

[0044] in, represents the scaling factor, Represent query, key and value features respectively. The self-attention mechanism captures the complex relationship between different parts of the input signal by dynamically assigning weights.

[0045] The short-circuit projection layer extracts the deep features of the value features through deep convolution, adds them to the self-attention results of the self-attention layer, and then performs convolution layer operations and normalization to further enhance the feature extraction capability of the model.

[0046] The output projection layer includes two fusion (Merge) layers and a deep convolution layer. The first fusion layer aggregates the output results of the short-circuit projection layer and the output results of the self-attention layer, and outputs the aggregated results, focusing on the combination of attention features and local features. After the aggregated results output by the first fusion layer pass through the deep convolution layer, they are again aggregated through the second fusion layer to output comprehensive features, thereby improving the expressiveness of the final features. This staged feature fusion strategy enables the model to more comprehensively model the dynamic characteristics and nonlinear behaviors in complex signals while taking into account the balance between global and local features.

[0047] (3) Feature Dimensionality Reduction Module The feature dimension after processing by the Transformer module is high, which is not conducive to subsequent processing. Therefore, feature dimension reduction is achieved through the feature dimension reduction module. The feature dimension reduction module includes a fully connected layer. Optionally, the feature dimension reduction module sets two fully connected layers to further improve the nonlinear modeling ability of the model and the expression ability of feature mapping. The two fully connected layers increase the model's expression ability of the nonlinear characteristics of the signal through a nonlinear activation function, while reducing the subsequent calculation complexity. Optionally, the nonlinear activation function is a Tanh activation function.

[0048] The fully connected layer performs feature dimensionality reduction through nonlinear transformation of the Tanh activation function, specifically satisfying:

[0049] in, represents the feature matrix after dimensionality reduction, represents the weight matrix of the fully connected layer, represents the bias vector of the fully connected layer, Represents the comprehensive features output by the Transformer module.

[0050] (4) Output module The output module further processes the feature matrix after dimension reduction and outputs the final predistortion signal, which specifically meets the following requirements:

[0051] in, represents the predistortion signal, represents the weight matrix of the output layer, represents the bias vector of the output layer, Represents the feature matrix after dimensionality reduction.

[0052] The predistortion signal output by the model is opposite to the nonlinear characteristics of the power amplifier. When input to the power amplifier, it can compensate for the nonlinear distortion of the power amplifier and achieve linear amplification.

[0053] Optionally, adjusting the hyper parameters of the PDTRANS pre-distortion model in S103 specifically includes: Adjust the learning rate, batch size, and number of network layers for the PDTRANS predistortion model.

[0054] Specifically, the learning rate, batch size, and number of network layers directly affect the training efficiency, generalization ability, and computational complexity of the model. Therefore, it is necessary to continuously use the loss function value to adjust the hyperparameters and find a balance between model performance improvement and resource consumption.

[0055] Different learning rate settings are used to balance the convergence speed and stability of the model; the adjustment of batch size focuses on the efficiency of memory usage and the accuracy of model gradient estimation; the optimization of the number of network layers focuses on the relationship between the model's expressiveness and computational complexity. Optionally, the adjustment of the number of network layers is the adjustment of the number of fully connected layers in the output feature dimensionality reduction module.

[0056] The embodiment of the present application can more efficiently learn the nonlinear characteristics and memory effects of the power amplifier through hyperparameter tuning of the model, and generate an accurate pre-distortion signal to compensate for the distortion behavior of the power amplifier.

[0057] Optionally, the training of the PDTRANS pre-distortion model uses standardized processes and optimization strategies to ensure the stability and reliability of the model performance. Specifically, 100,000 sets of signal data are used for modeling during the training process, and the data set is divided into training set, validation set and test set in a ratio of 3:1:1. All input signals are normalized before training to unify the data range to the same scale to optimize the convergence speed and performance of the model training. The normalized data is then input into the PDTRANS pre-distortion model for learning.

[0058] Optionally, the training uses the PyTorch framework and runs on a graphics processing unit (GPU) environment to accelerate model calculations. The specific hardware configuration includes three NVIDIA GTX 2080 GPUs and an Intel Corei7 CPU. The Adam optimizer is used in training, and the initial learning rate is set to , and dynamically adjust the learning rate during training to ensure fast convergence and robust performance.

[0059] Optionally, a mean square error (MSE) is used to measure the difference between the output signal sequence of the power amplifier and the target signal sequence after the predistortion compensation by the PDTRANS predistortion model, that is, the loss function of the PDTRANS predistortion model satisfies:

[0060] in, represents the loss function, represents the nth output signal in the output signal sequence of the power amplifier after predistortion compensation by the PDTRANS predistortion model, represents the nth target signal in the target signal sequence, and N represents the sequence length.

[0061] Optionally, the gradient is calculated by the error back propagation method, and the model parameters are optimized iteratively in combination with the gradient descent strategy. The training process continues until the performance indicators on the validation set reach the optimal level, ensuring that the model can accurately model the nonlinear characteristics of the power amplifier.

[0062] For S105, the trained PDTRANS pre-distortion model is used to perform pre-distortion compensation on the power amplifier, and different power amplifiers have different PDTRANS pre-distortion models obtained through direct learning.

[0063] A PDTRANS pre-distortion model is introduced at the front end of the power amplifier. The PDTRANS pre-distortion model is trained by direct learning until the model converges. Then, a target signal sequence containing the signal information at the current moment is input into the PDTRANS pre-distortion model to obtain a pre-distortion signal output by the model. The nonlinear distortion generated by the power amplifier at the current moment is compensated to realize the linearization of the power amplifier, thereby improving the adjacent channel power ratio of the power amplifier output signal and enhancing the linearity, anti-interference ability and overall performance of the communication system.

[0064] Optionally, S105 specifically includes: The time series signals at the current moment and the near historical moment are collected from the signal source, and the input matrix is ​​constructed after preprocessing; the input matrix includes the I / Q components and amplitude envelope of the time series signals at the current moment and the near historical moment; Input the input matrix to the trained PDTRANS predistortion model and output the predistortion signal at the current moment; The predistortion signal is input to the power amplifier for signal linearization compensation at the current moment.

[0065] The PDTRANS pre-distortion model provided in the embodiment of the present application can realize real-time, high-precision power amplifier pre-distortion compensation, significantly improve the linearity and anti-interference ability of the power amplifier, ensure the stability and efficiency of the system in complex environments, is suitable for complex power amplifier nonlinear environments, and can effectively respond to dynamically changing signals and working conditions; it has broad application prospects in military and civilian fields, especially in satellite communications, adaptive modulation, spectrum management and other applications.

[0066] The deep learning driven power amplifier predistortion device provided in the present application is described below. The deep learning driven power amplifier predistortion device described below and the deep learning driven power amplifier predistortion method described above can be referenced to each other.

[0067] Figure 5 is a structural diagram of a power amplifier predistortion device driven by deep learning provided in an embodiment of the present application, such as Figure 5 As shown, the device at least includes: A matrix construction module 501 is used to construct a signal matrix of a target signal sequence, where the target signal sequence is obtained by preprocessing an original time signal sequence collected from a signal source; The model training module 502 is used to input the signal matrix into the PDTRANS predistortion model and the power amplifier in sequence to obtain the output signal sequence of the power amplifier; determine the loss function value based on the output signal sequence and the target signal sequence, and adjust the hyperparameters of the PDTRANS predistortion model based on the loss function value; repeat the steps of obtaining the output signal sequence of the power amplifier and adjusting the hyperparameters of the PDTRANS predistortion model until the loss function is minimized, thereby obtaining a trained PDTRANS predistortion model that matches the power amplifier; The pre-distortion compensation module 503 is used to perform pre-distortion compensation on the power amplifier based on the trained PDTRANS pre-distortion model.

[0068] Optionally, the matrix construction module 501 is specifically used for: Extract the I / Q component characteristics and amplitude envelope characteristics of the sample signal in the target signal sequence at different times; The signal matrix of the target signal sequence is constructed with the number of extracted features and sequence length as dimensions.

[0069] Optionally, the PDTRANS pre-distortion model includes an input module, a Transformer module, a feature dimension reduction module and an output module; The input module includes a convolution layer; the convolution layer extracts local time series features of the signal matrix, including I / Q component features and amplitude envelope features; The Transformer module includes an improved normalization layer, a parallel convolution layer, a self-attention layer, a short-circuit projection layer, and an output projection layer; the improved normalization layer normalizes the input features; the parallel convolution layer generates key features, query features, and value features based on the normalized input features; the self-attention layer determines the self-attention results based on the key features, query features, and value features; the short-circuit projection layer extracts the deep features of the value features, adds the deep features and the self-attention results, and then performs convolution and normalization; the output projection layer aggregates the output features of the short-circuit projection layer and the self-attention results, performs deep convolution on the aggregated results, and then aggregates the output comprehensive features again; The feature dimension reduction module includes a fully connected layer, which performs feature dimension reduction through nonlinear transformation of the Tanh activation function; The output module outputs the predistortion signal based on the feature matrix after dimension reduction.

[0070] Optionally, the model training module 502 is specifically used for: The learning rate, batch size, and number of network layers of the PDTRANS pre-distortion model are adjusted based on the loss function value.

[0071] Optionally, the loss function of the PDTRANS predistortion model satisfies:

[0072] in, represents the loss function, represents the nth output signal in the output signal sequence of the power amplifier after predistortion compensation by the PDTRANS predistortion model, represents the nth target signal in the target signal sequence, and N represents the sequence length.

[0073] Optionally, the pre-distortion compensation module 503 is specifically used for: The time series signals at the current moment and the near historical moment are collected from the signal source, and the input matrix is ​​constructed after preprocessing; the input matrix includes the I / Q components and amplitude envelope of the time series signals at the current moment and the near historical moment; Input the input matrix to the trained PDTRANS predistortion model and output the predistortion signal at the current moment; The predistortion signal is input to the power amplifier for signal linearization compensation at the current moment.

[0074] It can be understood that the detailed functional implementation of each of the above-mentioned units / modules can be found in the introduction of the aforementioned method embodiment, and will not be repeated here.

[0075] It should be understood that the above-mentioned device is used to execute the method in the above-mentioned embodiment. The implementation principle and technical effect of the corresponding program module in the device are similar to those described in the above-mentioned method. The working process of the device can refer to the corresponding process in the above-mentioned method, which will not be repeated here.

[0076] Based on the method in the above embodiment, an embodiment of the present application provides an electronic device. The device may include: at least one memory for storing programs and at least one processor for executing the programs stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method described in the above embodiment.

[0077] Figure 6 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application, such as Figure 6 As shown, the electronic device may include: a processor (Processor) 601, a communication interface (Communications Interface) 602, a memory (Memory) 603 and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604. The processor 601 can call the software instructions in the memory 603 to execute the method described in the above embodiment.

[0078] In addition, the logic instructions in the above-mentioned memory 603 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present application.

[0079] Based on the method in the above embodiment, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method in the above embodiment.

[0080] Based on the method in the above embodiment, an embodiment of the present application provides a computer program product. When the computer program product runs on a processor, the processor executes the method in the above embodiment.

[0081] It is understandable that the processor in the embodiment of the present application may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0082] The method steps in the embodiments of the present application can be implemented by hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, and the software modules can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, mobile hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0083] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from a website site, a computer, a server or a data center to another website site, a computer, a server or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or a data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

[0084] It should be understood that the various numerical numbers involved in the embodiments of the present application are only used for the convenience of description and are not used to limit the scope of the embodiments of the present application.

[0085] It is easy for those skilled in the art to understand that the above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A deep learning driven power amplifier predistortion method, characterized in that: include: Constructing a signal matrix of a target signal sequence, wherein the target signal sequence is obtained by preprocessing an original time signal sequence collected from a signal source; Inputting the signal matrix into the PDTRANS predistortion model and the power amplifier in sequence to obtain an output signal sequence of the power amplifier; Determining a loss function value based on the output signal sequence and the target signal sequence, and adjusting a hyperparameter of the PDTRANS predistortion model based on the loss function value; Repeating the steps of obtaining the output signal sequence of the power amplifier and adjusting the hyperparameters of the PDTRANS pre-distortion model until the loss function is minimized, thereby obtaining a trained PDTRANS pre-distortion model matching the power amplifier; Predistortion compensation is performed on the power amplifier based on the trained PDTRANS predistortion model.

2. The power amplifier predistortion method according to claim 1, characterized in that: The signal matrix of the target signal sequence is constructed, comprising: Extracting I / Q component characteristics and amplitude envelope characteristics of sample signals in the target signal sequence at different times; A signal matrix of the target signal sequence is constructed with the number of extracted features and sequence length as dimensions.

3. The power amplifier predistortion method according to claim 1 or 2, characterized in that: The PDTRANS pre-distortion model includes an input module, a Transformer module, a feature dimension reduction module and an output module; The input module includes a convolution layer; the convolution layer extracts local time series features of the signal matrix, including I / Q component features and amplitude envelope features; The Transformer module includes an improved normalization layer, a parallel convolution layer, a self-attention layer, a short-circuit projection layer and an output projection layer; the improved normalization layer normalizes the input features; the parallel convolution layer generates key features, query features and value features based on the normalized input features; The self-attention layer determines the self-attention result based on the key feature, the query feature and the value feature; the short-circuit projection layer extracts the deep feature of the value feature, adds the deep feature and the self-attention result, and then performs a convolution operation and a normalization process; the output projection layer aggregates the output feature of the short-circuit projection layer and the self-attention result, performs a deep convolution on the aggregated result, and then aggregates the output comprehensive feature again; The feature dimension reduction module includes a fully connected layer, and the fully connected layer performs feature dimension reduction through a nonlinear transformation of a Tanh activation function; The output module outputs a predistortion signal based on the feature matrix after dimension reduction.

4. The power amplifier predistortion method according to claim 3, characterized in that: The step of adjusting the hyper parameters of the PDTRANS pre-distortion model based on the loss function value comprises: The learning rate, batch size and number of network layers of the PDTRANS pre-distortion model are adjusted based on the loss function value.

5. The power amplifier predistortion method according to claim 1, characterized in that: The loss function of the PDTRANS predistortion model satisfies: in, represents the loss function, represents the nth output signal in the output signal sequence of the power amplifier after predistortion compensation by the PDTRANS predistortion model, represents the nth target signal in the target signal sequence, and N represents the sequence length.

6. The power amplifier predistortion method according to claim 1, characterized in that: The performing predistortion compensation on the power amplifier based on the trained PDTRANS predistortion model comprises: Collecting the time series signals at the current moment and the near historical moment from the signal source, and constructing an input matrix after preprocessing; the input matrix includes the I / Q components and amplitude envelopes of the time series signals at the current moment and the near historical moment; Inputting the input matrix into the trained PDTRANS pre-distortion model, and outputting the pre-distortion signal at the current moment; The predistortion signal is input into the power amplifier to perform signal linearization compensation at the current moment.

7. A power amplifier predistortion device driven by deep learning, characterized in that: include: A matrix construction module, used to construct a signal matrix of a target signal sequence, wherein the target signal sequence is obtained by preprocessing an original time signal sequence collected from a signal source; A model training module, used for inputting the signal matrix into the PDTRANS predistortion model and the power amplifier in sequence to obtain the output signal sequence of the power amplifier; determining the loss function value based on the output signal sequence and the target signal sequence, and adjusting the hyperparameters of the PDTRANS predistortion model based on the loss function value; repeating the steps of obtaining the output signal sequence of the power amplifier and adjusting the hyperparameters of the PDTRANS predistortion model until the loss function is minimized, thereby obtaining a trained PDTRANS predistortion model matching the power amplifier; The predistortion compensation module is used to perform predistortion compensation on the power amplifier based on the trained PDTRANS predistortion model.

8. An electronic device, characterized in that: include: at least one memory for storing a computer program; At least one processor is used to execute the program stored in the memory. When the program stored in the memory is executed, the processor is used to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program runs on a processor, the processor is caused to execute the method according to any one of claims 1 to 6.

10. A computer program product, characterized in that When the computer program product runs on a processor, the processor is caused to execute the method according to any one of claims 1 to 6.

Images