Digital pre-distortion method and system based on hierarchical iterative neural network

Abstract

The invention relates to a digital pre-distortion method and system based on a hierarchical iterative neural network, and belongs to the field of digital pre-distortion of power amplifiers. According to the method, the pre-distortion process is decoupled into two stages of off-line generalization training and on-line specialization optimization. In the offline stage, a mixed data set containing multiple modulation modes and bandwidths is adopted to train an offline neural network, and a basic model with high generalization ability is constructed; in the online stage, aiming at a signal with a specific modulation mode and bandwidth, a single configuration data set is utilized to train an online neural network, and the accurate compensation capability of the model on a real-time input signal is enhanced. In practical application, a baseband signal is firstly subjected to targeted preprocessing through the online neural network, then is subjected to deep optimization through the offline neural network, and finally is input into the power amplifier PA. According to the method, through a generalization-specialization double-layer structure, the dynamic response capability and compensation precision of a digital pre-distortion system to a complex signal environment are remarkably improved.

Description

Technical Field

[0001] The invention belongs to the field of digital predistortion of power amplifiers and relates to a digital predistortion method and system based on a layered iterative neural network. Background Art

[0002] With the advent of fifth-generation wireless communication systems, communication systems face unprecedented challenges and opportunities. The development of 5G technology not only brings revolutionary improvements in transmission speed, capacity, and latency, but also heralds the need for wireless transmission at higher frequencies and wider bandwidths. With the increase in carrier frequency and the expansion of signal bandwidth, the designed high-capacity communication systems will face severe challenges in key components such as signal processing and power amplifiers (PAs). As one of the core components in wireless communication systems, the performance of the PA directly impacts system quality and efficiency. To achieve higher system performance, it is necessary to strike an appropriate balance between PA linearity and efficiency. However, this is particularly challenging in scenarios with high peak-to-average power ratio (PAPR).

[0003] Signals with high peak-to-average power ratios are often accompanied by significant nonlinear characteristics, placing higher demands on RF power amplifier performance. Power amplifiers must not only provide sufficient gain but also maintain high gain linearity. High-efficiency power amplifiers often exhibit significant nonlinear behavior, causing both nonlinear distortion (such as harmonics, intermodulation, and clipping) and linear distortion (amplitude-frequency distortion and phase-frequency distortion) in the signal. This results in signal spectrum broadening, waveform distortion, and frequency response deviation, which in turn affects communication performance.

[0004] Conventional power amplifier digital pre-distortion methods construct a pre-distorter and reversely model the nonlinear characteristics of the power amplifier based on algorithms such as lookup tables, polynomial fitting, and neural networks. This pre-distorts the input signal to offset the nonlinear distortion generated by the power amplifier and improve signal quality and power amplifier efficiency. However, these methods have limitations. For example, the lookup table method requires a large amount of data training and has a large memory footprint. The polynomial fitting method has limited processing capabilities for high-order nonlinearities. The neural network method has high computational complexity, long training time, and poor generalization. At the same time, these methods have poor adaptability to the pre-distortion models when facing temperature changes, aging, and changes in dynamic signal characteristics of the power amplifier, making it difficult to accurately compensate for distortion in real time, which affects the pre-distortion effect. Summary of the Invention

[0005] In light of this, the present invention aims to provide a digital predistortion (DPD) method and system based on a layered iterative neural network, dedicated to improving DPD performance in complex baseband signal scenarios. This innovative approach proposes a two-layer neural network predistortion architecture suitable for signal environments with varying modulation schemes and bandwidths. This method decouples the predistortion process into two phases: offline generalization training and online specialized optimization. In the offline phase, the neural network MultiOfflineNet is trained using a mixed dataset encompassing multiple modulation schemes and bandwidths to construct a basic model with high generalization capabilities, effectively addressing the limited adaptability of traditional predistortion models. In the online phase, MultiOnlineNet is trained using a single configured dataset for signals with specific modulation schemes and bandwidths, enhancing the model's ability to accurately compensate for real-time input signals and avoiding performance losses associated with generalization. In practical applications, the baseband signal InSignal is first targetedly preprocessed by MultiOnlineNet, then deeply optimized by MultiOfflineNet before being input to the power amplifier (PA). This dual-layer generalization-specialization architecture significantly improves the dynamic response and compensation accuracy of the DPD system in complex signal environments.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A digital predistortion method of a layered iterative neural network, the method comprising:

[0008] Generate baseband OFDM signals with various modulation modes and bandwidth configurations, input them into the power amplifier (PA) to obtain output data, and combine them to form a training data set;

[0009] Based on diverse OFDM signals and corresponding PA output data, the offline neural network model is trained offline to build a pre-distortion model with generalization capabilities.

[0010] Based on the completed offline training results, the test set in the single-configuration training data set is used to construct an online training data set. The online neural network model is trained using the online training data set to optimize the adaptability of the online neural network model to specific signals.

[0011] The performance of the offline and online training models is evaluated using the evaluation data set. If the normalized mean square error threshold between the OFDM signal and the PA output data is greater than the expected set value, the online neural network model training continues.

[0012] Based on the tested model, real-time baseband transmission data is obtained and input into the trained online neural network for targeted preprocessing. The overall nonlinear characteristics of the power amplifier (PA) are then compensated through the offline neural network. Finally, the power amplifier outputs the signal and transmits it to the antenna system for information transmission.

[0013] Furthermore, the construction process of the training dataset includes:

[0014] Based on orthogonal frequency division multiplexing technology, different OFDM time domain signals txWave are generated according to different bandwidths and modulations. During the generation process, different bandwidths and modulation modes are configured, and OFDM parameters are set accordingly. OFDM parameters include: subcarrier spacing, number of fast Fourier transform points, cyclic prefix length, upsampling factor, and sampling frequency;

[0015] The bit length of the transmission data block is calculated based on the modulation order and the number of OFDM subcarriers carrying data. Then, a long transmission data block is randomly generated. The transmission data block is modulated to generate an OFDM time-domain baseband signal. After upsampling, the OFDM baseband signal used in the training set is obtained. This step is repeated to obtain the baseband signal txWaveTrain of the training set, the baseband signal txWaveVal of the validation set, the baseband signal txWaveTest of the test set, and the baseband signal txWaveEval of the evaluation set under different modulation modes and bandwidths.

[0016] The txWaveTrain, txWaveVal, txWaveTest, and txWaveEval of different modulation modes and bandwidths are respectively input into the power amplifier PA to obtain the corresponding output signals paOutputTrain, paOutputVal, paOutputTest, and paOutputEval; the input signals are paired with the output signals to generate the training set Train = (txWaveTrain, paOutputTrain), the verification set Val = (txWaveVal, paOutputVal), the test set Test = (txWaveTest, paOutputTest), and the evaluation set Eval = (txWaveEval, paOutputEval).

[0017] Furthermore, the training process for the offline neural network model MultiOfflineNet is as follows:

[0018] The training set signal paOutputTrain, validation set signal paOutputVal, and test set signal paOutputTest output by the power amplifier corresponding to different modulation modes and bandwidths are processed by the linear gain factor lineGainFactor of the power amplifier, that is, each signal is divided by the linear gain factor lineGainFactor, and then combined with the corresponding configured txWaveTrain, txWaveVal, and txWaveTest to form a mixed training data set of the MultiOfflineNet neural network, where the mixed training data set includes a mixed training set MixTrain = (txMixWaveTrain, paMixOutputTrain), a mixed validation set MixVal = (txMixWaveVal, paMixOutputVal), and a mixed test set MixTest = (txMixWaveTest, paMixOutputTest);

[0019] Perform normalization preprocessing on each data set in the mixed training data set to obtain corresponding data, where the data includes training data MtxTrain = (inputMtxTrain, outputMtxTrain), verification data MtxVal = (inputMtxVal, outputMtxVal), and test data MtxTest = (inputMtxTest, outputMtxTest);

[0020] Then, the offline neural network MultiOfflineNet is trained using the training data MtxTrain = (inputMtxTrain, outputMtxTrain). During the training process, the verification data MtxVal = (inputMtxVal, outputMtxVal) is used for verification, and the test data MtxTest = (inputMtxTest, outputMtxTest) is used for testing.

[0021] Furthermore, the normalization preprocessing process for each data set in the mixed training data set is as follows:

[0022] Split the normalized paMixOutputTrain and paMixOutputVal into an in-phase signal I and a quadrature signal Q;

[0023] Then, according to the size of the memory depth memDepth and the nonlinear order nonlinearDegree, a matrix of corresponding size is formed. The matrix obtained by preprocessing is as follows:

[0024]

[0025] Where m represents the memory depth and q represents the nonlinear order;

[0026] Split the normalized txMixWaveTrain and txMixWaveVal into the in-phase signal I and the quadrature signal Q, and then combine them to obtain outputMtxTrain and outputMtxVal; inputMtxTrain and outputMtxTrain constitute the training data;

[0027] Similarly, inputMtxVal and outputMtxVal constitute the verification data.

[0028] Furthermore, the training process for the online neural network model MultiOnlineNet is as follows:

[0029] Select txWaveTest data corresponding to one of the modulation modes and bandwidths. After the data txWaveTest passes through the power amplifier PA, paOutputOnTest is generated.

[0030] Divide paOutputOnTest by the linear gain factor lineGainFactor of the PA, normalize txWaveTest and paOutputOnTest, and then preprocess them. Generate inputMtxTest data from paOutputOnTest as the input of the neural network, and generate paOutputTestMtx as the output of the neural network by normalizing the txWaveTest data.

[0031] Use inputMtxTest and paOutputTestMtx to train the online neural network MultiOnlineNet; when the loss value of the online neural network is less than the specified threshold, stop training and save the online deep neural network MultiOnlineNet.

[0032] Furthermore, the training set data of the online neural network is obtained in small batches, the gradient and loss of the online neural network model are evaluated, and the network parameters are updated. The termination condition of online training is defined as the maximum number of training rounds or the loss value reaches the threshold. The specific process of calculating the online neural network loss value is as follows: the small batch inputMtx data is passed through the current online training neural network MultiOnlineNet to obtain the predicted value Y of the online neural network. The online neural network loss value loss is expressed as the mean square error:

[0033]

[0034] Among them, N represents the number of elements in Y and T, Y i represents the predicted value of the online neural network, T i Represents the true value.

[0035] The preset normalized mean square error NMSE value is used as the loss threshold, where the relationship between the NMSE value and the loss value loss is NMSE = 10*log 10 (loss).

[0036] Furthermore, the process of performance testing and evaluation for the completed offline training model and online training model is as follows:

[0037] The evaluation signals txWaveEval with different modulation modes and bandwidths are passed through the online neural network MultiOnlineNet and the offline neural network MultiOfflineNet, and finally through the power amplifier PA to obtain paOutputOnEval;

[0038] The adjacent channel power ratio (ACPR), normalized mean square error (NMSE), and error vector magnitude (EVM) of txWaveEval and paOutputOnEval for different modulation modes and bandwidths are calculated in turn, and the model performance is evaluated based on the above evaluation indicators.

[0039] Furthermore, the offline neural network MultiOfflineNet includes a feature input layer, a fully connected output layer, and three fully connected layers between the input layer and the output layer. Each fully connected layer includes several neurons, and each fully connected layer uses the leakyRelu function as the activation function. The number of neurons in each fully connected layer is 64.

[0040] Furthermore, the online neural network MultiOnlineNet includes a feature input layer, an FC output layer, and two fully connected layers between the input layer and the output layer. Each fully connected layer uses the leakyRelu function as the activation function, and the number of neurons in each fully connected layer is 30.

[0041] On the other hand, a system for executing the digital predistortion method of the aforementioned hierarchical iterative neural network is also proposed, the system comprising: a data preparation module, an offline training module, an online training module, a performance evaluation module and a predistortion processing module, wherein:

[0042] The data preparation module is used to generate OFDM time-domain signals with different modulation modes and bandwidths and establish corresponding data sets. The offline training module trains the offline neural network based on the established data sets to obtain a deeply optimized offline neural network.

[0043] The online training module processes the data set according to the trained offline neural network and trains the online neural network based on the data set to obtain an online neural network for specific signal processing. The performance evaluation module performs performance testing and evaluation on the trained offline training model and online training model based on the evaluation data set. If the normalized mean square error threshold of the OFDM signal and PA output data is greater than the expected set value, the online neural network model training will continue.

[0044] The pre-distortion processing module includes a baseband processor, an online neural network, an offline neural network, and a power amplifier PA. The baseband signal sent by the baseband processor first passes through the trained online neural network MultiOnlineNet and the offline neural network MultiOfflineNet, then passes through the power amplifier PA to couple the PA output signal, undergoes down-conversion, and obtains the paOutputOnEval through analog-to-digital conversion sampling ADC, and finally transmits the paOutputOnEval signal to the antenna system.

[0045] The beneficial effects of the present invention are:

[0046] In mobile communication systems, the nonlinearity of radio frequency power amplifiers has long restricted the improvement of signal transmission power. Currently, while conventional digital pre-distortion methods can alleviate nonlinear distortion to a certain extent, they still face performance bottlenecks. With the increasing maturity of deep learning technology, it has demonstrated superior performance to traditional methods in single-configuration scenarios. Against this backdrop, the present invention proposes a deep learning-based digital pre-distortion method that closely fits actual application scenarios and has significant advantages:

[0047] First, an innovative dual neural network architecture is used to achieve hierarchical compensation: an offline neural network is constructed by using mixed configuration training data to compensate for the overall nonlinear characteristics of the power amplifier (PA); an online neural network is generated using single configuration training data to accurately optimize the unique nonlinear problems of the PA under specific configurations, effectively meeting the digital pre-distortion requirements of the PA in actual scenarios.

[0048] Secondly, the layered iterative dual neural network is effective in improving the nonlinear performance of PA. Measured data show that: under the QPSK signal configuration with a bandwidth of 5M, the normalized mean square error (NMSE) is significantly reduced from -20.7124dB to -32.5217dB, and the error vector magnitude is reduced from 0.087203 to 0.016026; in the 64QAM signal scenario with a bandwidth of 100M, the NMSE is optimized from -20.0994dB to -32.9132dB, and the error vector magnitude is reduced from 0.093981 to 0.015636, significantly improving the signal transmission quality and system performance.

[0049] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0051] Figure 1 Schematic diagram of the overall process of a digital predistortion method based on a layered iterative neural network according to an embodiment of the present invention;

[0052] Figure 2 A schematic diagram of the data preparation process according to an embodiment of the present invention;

[0053] Figure 3 Schematic diagram of an offline training process of an offline neural network according to an embodiment of the present invention;

[0054] Figure 4 Schematic diagram of an online training process of an online neural network according to an embodiment of the present invention;

[0055] Figure 5 Schematic diagram of the performance evaluation test process under an embodiment of the present invention;

[0056] Figure 6 Schematic diagram of the data processing flow of an offline training module according to another embodiment of the present invention;

[0057] Figure 7 Schematic diagram of the data processing flow of an online training module according to another embodiment of the present invention;

[0058] Figure 8 Schematic diagram of the data processing flow of a performance evaluation module according to another embodiment of the present invention;

[0059] Figure 9 Schematic diagram of the structure of a digital predistortion module according to another embodiment of the present invention;

[0060] Figure 10 Schematic diagram of the structure of an offline neural network according to another embodiment of the present invention;

[0061] Figure 11 Schematic diagram of the structure of an online neural network according to another embodiment of the present invention;

[0062] Figure 12 is a constellation diagram without using pre-distortion compensation according to yet another embodiment of the present invention;

[0063] Figure 13 This is a constellation diagram after using MultiOfflineNet pre-distortion compensation alone according to yet another embodiment of the present invention;

[0064] Figure 14 FIG. 1 is a constellation diagram after compensation using a layered iterative neural network according to yet another embodiment of the present invention. DETAILED DESCRIPTION

[0065] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0066] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0067] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0068] See also Figures 1 to 14 , which is a digital predistortion method and system based on a hierarchical iterative neural network.

[0069] Example 1

[0070] This embodiment first describes in detail the specific process of a layered iterative neural network digital predistortion method. Figure 1 As shown, it includes the following steps:

[0071] S1. Generate baseband OFDM signals with multiple modulation modes and bandwidth configurations, input them into the power amplifier PA to obtain output data, and combine them to form a training data set;

[0072] S2. Based on the diverse OFDM signals and corresponding PA output data, the offline neural network model is trained offline to build a pre-distortion model with generalization capabilities.

[0073] S3. Based on the completed offline training results, the test set in the single-configuration training data set is taken to construct an online training data set, and the online neural network model is trained using the online training data set to optimize the adaptability of the online neural network model to specific signals;

[0074] S4. Perform performance testing and evaluation on the offline training model and the online training model after training using the evaluation dataset.

[0075] S5. Based on the tested model, real-time baseband transmission data is obtained and input into the trained online neural network for targeted preprocessing. Then, it is deeply optimized through the offline neural network. Finally, the power amplifier outputs the signal and transmits it to the antenna system for information transmission.

[0076] In step S1 of this embodiment, in order to generate baseband OFDM signals with multiple modulation modes and bandwidth configurations, based on orthogonal frequency division multiplexing (OFDM) technology, the configuration parameters are flexibly adjusted to generate OFDM time domain signals with different modulation modes and bandwidths, which are recorded as txWave. The txWave is input into the power amplifier PA to obtain the corresponding output signal paOutput. Based on the above signal, a series of single-configuration training data sets (txWave, paOutput) are generated, and are scientifically divided into training sets (txWaveTrain, paOutputTrain), verification sets (txWaveVal, paOutputVal), test sets (txWaveTest, paOutputTest) and evaluation sets (txWaveEval, paOutputEval), providing complete data support for subsequent offline training and online training links. The specific process is as follows: Figure 2 Shown, including:

[0077] Step 11: Generate different OFDM time-domain signals based on different bandwidths and modulations. In addition to configuring different bandwidths and modulation methods, this process also requires configuring other OFDM parameters, including subcarrier spacing, number of fast Fourier transform (FFT) points, cyclic prefix (CP) length, upsampling factor, and sampling frequency.

[0078] Step 12: Based on the modulation order and the number of OFDM subcarriers carrying data, the transmission data block bit length is calculated and a random transmission data block of that bit length is generated. This random transmission data block of equal length is then generated. This data block is modulated to generate an OFDM time-domain baseband signal, which is then upsampled to obtain the OFDM baseband signal used in the training set. By repeating this process, the baseband signals txWaveTrain for the training set, txWaveVal for the validation set, txWaveTest for the test set, and txWaveEval for the evaluation set are obtained for different modulation modes and bandwidths.

[0079] Step 13: Input txWaveTrain, txWaveVal, txWaveTest, and txWaveEval of different modulation modes and bandwidths into the power amplifier (PA) to generate the corresponding output signals paOutputTrain, paOutputVal, paOutputTest, and paOutputEval. Pair the input and output signals to generate the training set Train = (txWaveTrain, paOutputTrain), the validation set Val = (txWaveVal, paOutputVal), the test set Test = (txWaveTest, paOutputTest), and the evaluation set Eval = (txWaveEval, paOutputEval).

[0080] In step S2 of this embodiment, the training set (txWaveTrain, paOutputTrain), validation set (txWaveVal, paOutputVal) and test set (txWaveTest, paOutputTest) in each single configuration training data set are combined according to the modulation mode and bandwidth to construct a mixed configuration training data set (txMixWave, paMixOutput). The offline neural network MultiOfflineNet is trained using the mixed training set, and the training effect is verified by the mixed validation set during the training process. After the training is completed, the performance of the MultiOfflineNet model is evaluated using the mixed test set. Figure 3 As shown, the offline training process is:

[0081] Step 21: Divide paOutputTrain, paOutputVal, and paOutputTest of different modulation modes and bandwidths by the PA linear gain factor lineGainFactor. Then combine them with the corresponding configured txWaveTrain, txWaveVal, and txWaveTest to form a mixed training dataset (txMixWave, paMixOutput) of the MultiOfflineNet neural network. This dataset also includes a mixed training set MixTrain = (txMixWaveTrain, paMixOutputTrain), a mixed validation set MixVal = (txMixWaveVal, paMixOutputVal), and a mixed test set MixTest = (txMixWaveTest, paMixOutputTest);

[0082] Step 22: Perform normalization preprocessing on the training set (txMixWaveTrain, paMixOutputTrain) in the mixed training dataset, where the training set generates training data MtxTrain = (inputMtxTrain, outputMtxTrain). Similarly, the validation set generates validation data MtxVal = (inputMtxVal, outputMtxVal) and the test set generates test data MtxTest = (inputMtxTest, outputMtxTest).

[0083] Step 23: Use the training data (inputMtxTrain, paOutputMtxTrain) to train the offline neural network MultiOfflineNet. During the training process, use the validation data (inputMtxVal, outputMtxVal) for verification and the test data (inputMtxTest, paOutputTest) for testing.

[0084] In step S3 of this embodiment, based on the completed offline training results, the test set (txWaveTest, paOutputTest) in the single configuration training data set is taken to construct an online training data set. The specific process is: first input txWaveTest into the trained offline neural network MultiOfflineNet, and obtain txWaveTestNN after pre-distortion processing; then input txWaveTestNN into the power amplifier PA to obtain the output signal paOutputOnTest. Replace paOutputTest in the original test set with paOutputOnTest to generate a new training data set (txWaveTest, paOutputOnTest). Use this data set to train the online neural network MultiOnlineNet, and dynamically determine whether to terminate the training based on the loss value during the training process to ensure the model optimization effect. Figure 4 As shown in Figure 2, the specific process of online training includes:

[0085] Step 31: Select txWaveTest data with a modulation mode and bandwidth. After the data txWaveTest passes through the power amplifier PA, paOutputOnTest is generated.

[0086] Step 32: Divide paOutputOnTest by the linear gain factor lineGainFactor of the PA, normalize txWaveTest and paOutputOnTest, and then preprocess them. Generate inputMtxTest data from paOutputOnTest as the input of the neural network, and generate paOutputTestMtx as the output of the neural network by normalizing the txWaveTest data.

[0087] Step 33: Use inputMtxTest and paOutputTestMtx to train the online neural network MultiOnlineNet. Once the loss value is less than the specified threshold, the training is stopped and the online deep neural network MultiOnlineNet is saved.

[0088] In step S4 of this embodiment, the performance of the final DPD system is evaluated using the evaluation set (txWaveEval, paOutputEval). The specific process is as follows: txWaveEval is passed through the online-trained MultiOnlineNet and the offline-trained MultiOfflineNet, and then input into the power amplifier PA to obtain the output signal paOutputOnEval. Based on txWaveEval and paOutputOnEval, the core performance indicators of normalized mean square error (NMSE), adjacent channel power ratio (ACPR), and error vector magnitude (EVM) are calculated to quantify the actual compensation effect and performance of the DPD system.

[0089] like Figure 5 As shown, the performance test evaluation process is:

[0090] Step 1: The evaluation signal txWaveEval of different modulation modes and bandwidths passes through the online neural network MultiOnlineNet and the offline neural network MultiOfflineNet, and finally passes through the power amplifier PA to obtain paOutputOnEval.

[0091] Step 2: Calculate ACPR, NMSE, and EVM of txWaveEval and paOutputOnEval for different modulation modes and bandwidths respectively.

[0092] In this example, the NMSE, ACPR, and EVM of txWaveEval and paOutputOnEval are analyzed and compared.

[0093] Normalized Mean Squared Error (NMSE): A normalized form of the Mean Squared Error (MSE) that evaluates the difference between the model's predicted value and the true value. The smaller the NMSE value, the better the model's prediction effect. The specific formula for calculating NMSE is as follows:

[0094]

[0095] where y r (n) represents the actual output signal paOutputOnEval of the power amplifier PA divided by the linear gain factor of the power amplifier, y i (n) represents the ideal output signal txWaveEval of the power amplifier PA.

[0096] Adjacent Channel Power Ratio (ACPR): The ratio of the primary channel power to the adjacent channel power. A larger ACPR value indicates less interference from the primary channel to the adjacent channel, and better signal spectral purity. The specific formula for ACPR is as follows:

[0097]

[0098] Among them, P a Represents the power within the adjacent channel bandwidth, P m Represents the signal power in the main channel.

[0099] Error Vector Magnitude (EVM): This is used to measure the deviation between the actual signal and the ideal signal. The smaller the EVM value, the better the signal quality. The specific formula for EVM is as follows:

[0100]

[0101] where e r (n) represents the constellation point of the actual signal, e i (n) represents the constellation point of the ideal signal.

[0102] In step S5 of this embodiment, after the training is completed, the baseband processor sends the baseband evaluation signal txWaveEval again to simulate real-time baseband data transmission, simulate the baseband real-time transmission of evaluation signals with different bandwidths and different modulation combinations, and pass through the trained online neural network MultiOnlineNet and offline neural network MultiOfflineNet, through the power amplifier PA, and couple the PA output signal.

[0103] Example 2

[0104] This embodiment establishes a digital predistortion system based on a hierarchical iterative neural network based on a digital predistortion method proposed in Example 1, which includes: a data preparation module, an offline training module, an online training module, a performance evaluation module and a predistortion processing module, wherein the data preparation module is used to generate OFDM time domain signals with different modulation modes and bandwidths and establish corresponding data sets; the offline training module trains the offline neural network based on the established data set to obtain a deeply optimized offline neural network; the online training module processes the data set according to the trained offline neural network, and trains the online neural network based on the trained data set to obtain an online neural network for specific signal processing; the performance evaluation module performs performance testing and evaluation on the trained offline neural network and the online neural network, and deploys the evaluated offline neural network and the online neural network in the predistortion processing module to perform real-time signal digital predistortion processing.

[0105] Specifically, in this embodiment, the data preparation module is based on orthogonal frequency division multiplexing (OFDM) technology. By flexibly adjusting the configuration parameters, it generates an OFDM time domain signal with different modulation modes and bandwidths, which is recorded as txWave. The txWave is input into the power amplifier PA to obtain the corresponding output signal paOutput. Based on the above signal, the module generates a series of single-configuration training data sets (txWave, paOutput), and scientifically divides them into a training set (txWaveTrain, paOutputTrain), a validation set (txWaveVal, paOutputVal), a test set (txWaveTest, paOutputTest), and an evaluation set (txWaveEval, paOutputEval), providing complete data support for subsequent offline training and online training.

[0106] The offline training module combines the training set (txWaveTrain, paOutputTrain), validation set (txWaveVal, paOutputVal) and test set (txWaveTest, paOutputTest) in each single configuration training data set according to the modulation mode and bandwidth to construct a mixed configuration training data set (txMixWave, paMixOutput). The offline neural network MultiOfflineNet is trained using this mixed training set, and the training effect is verified by the mixed validation set during the training process. After the training is completed, the performance of the MultiOfflineNet model is evaluated using the mixed test set. Figure 6 The data processing flow diagram of the offline training module shown in the figure shows that in the offline training link of this embodiment, the offline training set is composed of (txMixWaveTrain, PaMixOutputTrain), which integrates the training sets (txWaveTrain, paOutputTrain) under different modulation modes and bandwidth configurations. The specific processing flow is as follows: First, txMixWaveTrain is normalized to obtain paOutputMtx data; at the same time, PaMixOutputTrain is divided by the power amplifier linear gain factor lineGainFactor to complete linear scaling, and then normalization is performed. After the above operations, a data set (inputMtx, paOutputMtx) for offline neural network training is generated. Finally, this data set is used to train the offline neural network MultiOfflineNet and construct a pre-distortion model with generalization capabilities.

[0107] The online training module takes the test set (txWaveTest, paOutputTest) in the single configuration training data set based on the completed offline training results to construct an online training data set. The specific process is: first input txWaveTest into the trained offline neural network MultiOfflineNet, and obtain txWaveTestNN after pre-distortion processing; then input txWaveTestNN into the power amplifier PA to obtain the output signal paOutputOnTest. Replace paOutputTest in the original test set with paOutputOnTest to generate a new training data set (txWaveTest, paOutputOnTest). Use this data set to train the online neural network MultiOnlineNet, and dynamically determine whether to terminate the training based on the loss value during the training process to ensure the model optimization effect. Figure 7 The data processing process of the online training module is shown. In this embodiment, the online training module selects a test set (txWaveTest, paOutputTest) from the training data, inputs txWaveTest into the trained offline neural network MultiOfflineNet for digital pre-distortion processing, and then passes the processed signal through the power amplifier PA to generate paOutputOnTest; then, paOutputOnTest replaces paOutputTest in the original test set to form a training set (txWaveTest, paOutputOnTest) for online training; then, the test set is normalized, txWaveTest is normalized to obtain paOutputMtx, and paOutputOnTest is normalized to obtain inputMtx, thereby constructing a new training set (inputMtx, paOutputMtx), and this training set is used to train the online neural network MultiOnlineNet.

[0108] The performance evaluation module uses the evaluation set (txWaveEval, paOutputEval) to evaluate the performance of the final digital pre-distortion system. The specific process is: txWaveEval is passed through the MultiOnlineNet completed by online training and the MultiOfflineNet completed by offline training, and then input into the power amplifier PA to obtain the output signal paOutputOnEval. Based on txWaveEval and paOutputOnEval, the core performance indicators of normalized mean square error (NMSE), adjacent channel power ratio (ACPR), and error vector magnitude (EVM) are calculated respectively to quantify the actual compensation effect and performance of the digital pre-distortion system. Figure 8The performance evaluation module's data processing process is shown in Figure 1. In this embodiment, for the txWaveEval in the training dataset, the performance evaluation module first performs digital predistortion processing on it through the online neural network MultiOnlineNet and the offline neural network MultiOfflineNet. The digital predistortion output signal paOutputOnEval is then obtained through the power amplifier PA. The adjacent channel power ratio (ACPR) is then calculated based on paOutputOnEval. The normalized mean square error (NMSE) and error vector magnitude (EVM) are calculated in combination with txWaveEval and paOutputOnEval to evaluate the performance of the digital predistortion system.

[0109] In this embodiment, if Figure 9 Figure 1 shows the structure of the predistortion processing module, which includes a baseband processor, an online neural network, an offline neural network, and a power amplifier (PA). The baseband processor sends a baseband signal, denoted as txWaveTest, through the deployed offline neural network MultiOfflineNet. The signal then passes through a digital-to-analog converter (DAC) and an upconverter, and then through the power amplifier (PA). The PA output signal is coupled, down-converted, and sampled using an analog-to-digital converter (ADC). The resulting feedback digital signal, denoted as paOutputOnTest, is used for training the online neural network MultiOnlineNet.

[0110] When the online neural network MultiOnlineNet is initially deployed, initial neural network parameters are loaded. The training set for the MultiOnlineNet network is generated using the real-time collected power amplifier (PA) feedback signal (paOutputOnTest) and baseband signal (txWaveTest). During training, the terminal device repeatedly fine-tunes the online neural network using small batches of training data until the loss threshold is met.

[0111] After training, the baseband processor sends the baseband evaluation signal txWaveEval again to simulate real-time baseband data transmission. It simulates the baseband sending evaluation signals with different bandwidths and modulation combinations in real time. The signals are passed through the trained online neural network MultiOnlineNet and offline neural network MultiOfflineNet, coupled with the power amplifier PA, and then down-converted and sampled by analog-to-digital conversion (ADC) to obtain paOutputOnEval.

[0112] Example 3

[0113] This embodiment describes the actual application of the predistortion processing module finally obtained based on Embodiment 1 and Embodiment 2.

[0114] With the increasing integration of terminal devices and the rapid development of high-efficiency power amplifiers (PAs), there is an urgent need to research predistortion technology in these devices. Terminal devices are very complex in transmission scenarios, with dynamically changing signal waveforms and bandwidths. The digital predistortion method based on a layered iterative neural network proposed in this paper supports high-efficiency power amplifiers in terminals to meet performance requirements under different waveforms and bandwidths.

[0115] The digital pre-distortion system based on hierarchical iterative neural networks is as follows: the baseband processor sends the baseband signal through the online neural network MultiOnlineNet to complete targeted and precise pre-processing, and then uses the offline neural network MultiOfflineNet for deep optimization. The signal is then transmitted to the air interface through the digital-to-analog converter (DAC) and up-converter, and finally through the power amplifier PA and antenna system. The specific block diagram is as follows Figure 9 shown.

[0116] The specific applications of the digital predistortion system based on the hierarchical iterative neural network in actual terminals are as follows:

[0117] The offline neural network MultiOfflineNet is trained in the cloud or on a server using a mixed training set (including datasets with different bandwidths and debugging methods) prepared in the offline training data. The trained MultiOfflineNet model weights and architecture are solidified into a lightweight format and embedded in the terminal device.

[0118] In this embodiment, the baseband processor sends a baseband signal denoted as txWaveTest through the deployed offline neural network MultiOfflineNet, through the digital-to-analog converter DAC and the up-converter, and then through the power amplifier PA, coupling the PA output signal, down-converting, and sampling the analog-to-digital converter (ADC) to obtain a feedback digital signal denoted as paOutputOnTest, which is used for online neural network MultiOnlineNet training.

[0119] When the online neural network MultiOnlineNet is initially deployed, initial neural network parameters are loaded. The training set for the MultiOnlineNet network is generated using the real-time collected power amplifier (PA) feedback signal (paOutputOnTest) and baseband signal (txWaveTest). During training, the terminal device repeatedly fine-tunes the online neural network using small batches of training data until the loss threshold is met.

[0120] After training, the baseband processor sends the baseband evaluation signal txWaveEval again to simulate real-time baseband data transmission. It simulates the baseband sending evaluation signals with different bandwidths and modulation combinations in real time. The signals are passed through the trained online neural network MultiOnlineNet and offline neural network MultiOfflineNet, coupled with the power amplifier PA, and then down-converted and sampled by analog-to-digital conversion (ADC) to obtain paOutputOnEval.

[0121] In this embodiment, the key steps are described as follows:

[0122] There are many signal characteristics that characterize baseband signals. In this embodiment, bandwidth and modulation are primarily used for illustration. Modulation methods include QPSK, 16QAM, and 64QAM. Bandwidth is expressed using the number of FFT points and the number of allocated subcarriers. A higher modulation order and a larger bandwidth increase the peak-to-average power ratio (PAPR) of the OFDM time-domain signal. This high PAPR of the baseband signal places high demands on the linearity of the power amplifier (PA), requiring the PA to maintain a wider linear range.

[0123] In this embodiment, the analog baseband signal is based on an orthogonal frequency division multiplexing (OFDM) waveform modulation scheme to generate OFDM time-domain signals of varying bandwidths and modulations. Parameters such as the subcarrier spacing, FFT count, CP length, upsampling factor, and sampling frequency must be configured. Based on the modulation order and number of subcarriers, the transmission data block bit length is calculated. Transmitted data blocks are randomly generated and modulated, and an OFDM time-domain waveform is generated. A training set is obtained through upsampling, and different configurations are used to generate the training set dsDataTrain, the validation set dsDataVal, the test set dsDataTest, and the evaluation set dsDataEval.

[0124] A power amplifier (PA) converts a small input signal into a high-power output through active devices such as transistors. In the linear region, the output is proportional to the input. However, in practice, inherent nonlinearity (described by nonlinearDegree) and memory effects (described by memDepth) in these devices can cause harmonic and intermodulation distortion in the output signal, resulting in the third-order intercept point (IP3) and 1dB compression point.

[0125] In this embodiment, the main parameters characterizing the power amplifier PA include memory depth (memDepth), nonlinear degree (nonlinearDegree), linear gain factor (lineGainFactor), amplitude-to-phase conversion (AMPMConversion), third-order intermodulation intercept point (IP3), 1dB compression point, and cubic polynomial model. These parameters are described below:

[0126] Memory depth (memDepth): refers to the depth of the memory effect of the power amplifier PA, which describes the time range or number of sample points over which the current output is affected by historical inputs. The memory effect originates from the dynamic characteristics of the circuit (such as thermal effects, bias circuit response, etc.), causing the nonlinear characteristics to change over time. In OFDM signals, the memory effect will introduce inter-symbol interference and affect the signal quality. A larger memDepth indicates that the system nonlinearity has a longer delay dependence. The memory depth of the power amplifier PA in this embodiment is 5.

[0127] Nonlinearity order (nonlinear degree): describes the order of the nonlinear characteristics of a power amplifier (PA), typically referring to the highest order of a Taylor expansion or Volterra series. The nonlinearity order most directly describes the degree of nonlinearity in a power amplifier (PA). In this embodiment, the nonlinearity order of the power amplifier (PA) is 5.

[0128] Linear gain factor (lineGainFactor): Linear gain factor refers to the linear conversion of the dB gain of the power amplifier PA. The conversion formula between linear gain factor (lineGainFactor) and dB gain (dBGain) is:

[0129]

[0130] In this embodiment, the power amplifier PA gain is set to 24.5 dB, and the linear gain factor is set to 16.788.

[0131] Amplitude-to-Phase Conversion (AMPMConversion): This parameter describes how changes in the input signal's amplitude affect the output signal's phase, expressed in degrees / dB. In this example, the amplitude-to-phase conversion is set to 1 degree / dB, meaning that for every 1dB change in the PA input signal's amplitude, the output signal's phase will change by 1 degree.

[0132] Third-Order Intermodulation Intercept Point (IP3): This refers to the intersection of the ideal linear gain curve of the power amplifier (PA) and the third-order intermodulation products. A higher IP3 indicates better linearity and improved intermodulation distortion immunity. In this embodiment, the IP3 is set to 35dBm, indicating that the ideal linear gain curve of the PA intersects the third-order intermodulation products when the input power of the PA is 35dBm.

[0133] 1dB compression point: The point at which the actual gain curve differs from the ideal gain curve by 1dB at output power. The 1dB compression point marks the critical point at which the power amplifier enters the nonlinear compression region. Beyond this point, gain compression occurs, and the signal becomes severely distorted. In this embodiment, the 1dB compression point is set to 25.4dBm, indicating that if the input power of the power amplifier PA exceeds 25.4dBm, the output signal will become severely distorted.

[0134] In this embodiment, the specific configuration parameters of the power amplifier PA are as follows:

[0135] Method:"Cubic polynomial"

[0136] LinearGain:24.5000

[0137] TOISpecification:'IIP3'

[0138] IIP3:35

[0139] AMPMConversion:1

[0140] PowerLowerLimit:10

[0141] PowerUpperLimit:Inf

[0142] ReferenceImpedance:50

[0143] In this embodiment, the offline neural network MultiOfflineNet in the pre-distortion processing module consists of a feature input layer, an FC output layer, and three fully connected layers between the input and output layers. The number of neurons in the three fully connected layers is 64. The activation function is leakyRelu. The offline neural network structure is as follows: Figure 10 shown.

[0144] In this embodiment, the number of columns of input data of the offline neural network MultiOfflineNet is 30, which matches the input dimension of the offline neural network. The number of columns of output data is 2, which matches the output dimension of the offline neural network.

[0145] The offline neural network is trained by collecting data before the power amplifier PA works. In order to realize the compensation of macro nonlinear characteristics of the offline neural network, the offline neural network training data includes the impact of various configurations and various working scenario factors on the power amplifier PA.

[0146] In this embodiment, the baseband generates different OFDM time-domain signals based on different bandwidths and modulations. In this embodiment, the bandwidths selected are 5 MHz, 15 MHz, 40 MHz, and 100 MHz. The modulation methods include QPSK, 16QAM, and 64QAM. This results in a total of 12 OFDM time-domain signals. These 12 OFDM signals are shown in Table 1.

[0147] Table 1

[0148]

[0149] In addition to configuring different bandwidths and modulation methods, generating OFDM time-domain signals also requires configuring other parameters, as shown in Table 2.

[0150] Table 2

[0151]

[0152]

[0153] According to the modulation order and the number of OFDM subcarriers carrying data, the bit length of the transmission data block is calculated, and a transmission data block of the corresponding length is randomly generated.

[0154] Based on this, a random block of transmit data of equal length is generated. The block size is 131*96. This block is modulated to generate an OFDM time-domain baseband signal. This is then upsampled to obtain the OFDM baseband signal used for training. The resulting OFDM baseband signal size is 131520*1.

[0155] Repeat this process to generate the training set txWaveTrain and the validation set txWaveVal for different modulation modes and bandwidths. The size of txWaveTrain and txWaveVal is 131520*1.

[0156] txWaveTrain and txWaveVal of different modulation modes and bandwidths are passed through the power amplifier (PA) to generate the corresponding output signals paOutputTrain and paOutputVal. The input and output signals are paired and combined to generate the training set (txWaveTrain, paOutputTrain) and validation set (txWaveVal, paOutputVal) for the offline neural network.

[0157] In this embodiment, txWaveTrain and txWaveVal for 12 different modulation modes and bandwidth combinations are combined to generate txMixWaveTrain and txMixWaveVal. The resulting txMixWaveTrain and txMixWaveVal values ​​are 1578240*1, or (131520*12)*1. After eliminating the nonlinear gain factor, paOutputTrain and paOutputVal are combined to generate paMixOutputTrain and paMixOutputVal. The resulting paMixOutputTrain and paMixOutputVal values ​​are also 1578240*1.

[0158] After normalizing txMixWaveTrain, txMixWaveVal, paMixOutputTrain, and paMixOutputVal, they are preprocessed to obtain outputMtxTrain, outputMtxVal, inputMtxTrain, and inputMtxVal, respectively. The preprocessing process is as follows: the normalized paMixOutputTrain and paMixOutputVal are split into in-phase signals (I signals) and quadrature signals (Q signals). Then, they are combined into matrices of corresponding sizes according to the memory depth (memDepth) and nonlinear degree (nonlinearDegree). The preprocessed matrices are as follows:

[0159]

[0160] Where m represents the memory depth (memDepth), q represents the nonlinear degree (nonlinearDegree), and in this embodiment, the memory depth is set to 5 and the nonlinear degree is also set to 5. The size of inputMtxTrain and outputMtxTrain is 131520*30.

[0161] Split the normalized txMixWaveTrain and txMixWaveVal into an in-phase signal (I signal) and a quadrature signal (Q signal), then combine them to generate outputMtxTrain and outputMtxVal. The size of outputMtxTrain and outputMtxVal is 131520 x 2. InputMtxTrain and outputMtxTrain constitute the training data. Similarly, inputMtxVal and outputMtxVal constitute the validation data.

[0162] The offline neural network MultiOfflineNet is trained using the training data inputMtxTrain and outputMtxTrain. During training, the validation data inputMtxVal and outputMtxVal are used for verification. inputMtxTrain serves as the input of the offline neural network MultiOfflineNet, and outputMtxTrain serves as the output of the offline neural network MultiOfflineNet. The specific offline neural network training parameters are shown in Table 3.

[0163] Table 3

[0164]

[0165] The condition for stopping training is that the validation loss does not improve in the validation of the validation patience round (here 5), then the training is terminated early or trained to the maximum number of rounds.

[0166] After training, the offline neural network MultiOfflineNet is deployed on the terminal device. The offline neural network mainly compensates for the overall nonlinear characteristics of the power amplifier PA and solves the macro nonlinear characteristic compensation of the power amplifier PA.

[0167] In this embodiment, the online neural network MultiOnlineNet consists of a feature input layer, an FC output layer, and two fully connected layers between the input and output layers. The number of neurons in the two fully connected layers is 30. The activation function is leakyRelu. The online neural network structure is as follows: Figure 11 shown.

[0168] The output data has a row size of 2, which matches the output dimension of the online neural network.

[0169] The online neural network is completed during normal operation, when the modulation mode and bandwidth, as well as relevant power amplifier operating conditions such as temperature, are already configured. The online neural network solves the problem of nonlinear compensation of the power amplifier under specific baseband signal configurations and operating scenarios.

[0170] In this embodiment, a single-configuration modulation scheme and bandwidth are used to regenerate an OFDM time-domain signal. The specific generation method is similar to the offline neural network training data generation process. The resulting OFDM baseband time-domain signal is txWaveTest, whose size is 131520*1. txWaveTest is passed through the trained offline neural network MultiOfflineNet and then through the power amplifier PA. The PA output signal is coupled to produce the corresponding output signal paOutputOnTest.

[0171] First, the linear gain factor of paOutputOnTest is eliminated, and then txWaveTest and paOutputOnTest are normalized and preprocessed to generate the training data set (inputMtx, paOutputMtx) of the online neural network. The production process is the same as the offline neural network training data set generation process.

[0172] Online neural network training data (inputMtx, paOutputMtx) is acquired in small batches, model gradients and losses are evaluated, and network parameters are updated. The termination condition for online training is defined as the maximum number of training rounds or the loss value reaches the threshold. It mainly simulates the process of the baseband sending time-domain OFDM signals through the power amplifier PA.

[0173] The specific process of calculating the online neural network loss value is as follows: a small batch of inputMtx data passes through the current online training neural network MultiOnlineNet to obtain the predicted value Y of the online neural network.

[0174] The loss value of the online neural network is expressed using the mean square error (MSE function). The loss value threshold of this embodiment is NMSE = -35, and the calculation formula between NMSE and loss is:

[0175] NMSE = 10*log 10 (loss)

[0176] The online neural network training parameters in this embodiment are shown in Table 4.

[0177] Table 4

[0178]

[0179] The online neural network training process simulates the baseband data transmission process in real time. If the maximum number of rounds of training is not reached or the loss value threshold is not reached during online training, training will continue to calculate the loss value and update the network parameters.

[0180] During the online neural network training process, if the loss value threshold is reached, the online neural network training is stopped, indicating that the online neural network has learned the microscopic nonlinear characteristics of the power amplifier PA of the fixed configuration.

[0181] In this embodiment, the performance of both online and offline neural networks is evaluated, essentially assessing the neural network's ability to compensate for the nonlinear distortion of the power amplifier (PA). Digital predistortion performance can be evaluated across three dimensions: normalized mean square error (NMSE), adjacent channel power ratio (ACPR), and error vector magnitude (EVM).

[0182] In the terminal baseband link transmission, it is assumed that the baseband sends an OFDM time domain signal txWaveEval for performance evaluation. According to the invention, in this embodiment, the OFDM time domain signal configuration parameters for generating txWaveEval are the same as the txWaveTest parameters.

[0183] The time-domain signal txWaveEval passes through the online neural network MultiOnlineNet and the offline neural network MultiOfflineNet, and then passes through the power amplifier PA to obtain paOutputOnEval. In this embodiment, the NMSE, ACPR, and EVM of txWaveEval and paOutputOnEval are analyzed and compared.

[0184] In this embodiment, the NMSE, ACPR, EVM and constellation diagram of three methods are compared to evaluate the performance of digital pre-distortion: no digital pre-distortion, independent offline training neural network MultiOfflineNet, and layered iterative neural network. Figure 12 、 Figure 13 、 Figure 14 shown.

[0185] Table 5

[0186]

[0187] In this embodiment, Table 5 compares the performance of various bandwidths (5M, 10M, 40M, and 100M), modulation modes (QPSK, 16QAM, and 64QAM), and various combinations of bandwidth and modulation. Overall, using an independent offline neural network for compensation outperforms not using any digital pre-distortion processing. Using a layered iterative neural network also outperforms independent offline neural network compensation.

[0188] In this embodiment, Figure 12 , Figure 13 and Figure 14 , giving the constellation diagram after the power amplifier PA, Figure 12 This is the constellation diagram output without any digital pre-distortion compensation, with obvious noise and phase rotation problems. Figure 13 The constellation diagram after compensation by the separate offline neural network MultiOfflineNet basically solves the phase rotation problem, but the influence of noise is still obvious. Figure 14 This is the constellation diagram after the complete layered iterative neural network compensation in the present invention, and there is obviously no phase rotation and power amplifier noise influence.

[0189] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A digital predistortion method based on a layered iterative neural network, characterized in that: The method comprises: Generate baseband OFDM signals with various modulation modes and bandwidth configurations, input them into the power amplifier (PA) to obtain output data, and combine them to form a training data set; Based on diverse OFDM signals and corresponding PA output data, the offline neural network model is trained offline to build a pre-distortion model with generalization capabilities. Based on the completed offline training results, the test set in the single-configuration training data set is used to construct an online training data set. The online neural network model is trained using the online training data set to optimize the adaptability of the online neural network model to specific signals. The performance of the offline and online training models is evaluated using the evaluation data set. If the normalized mean square error threshold between the OFDM signal and the PA output data is greater than the expected set value, the online neural network model training continues. Based on the tested model, real-time baseband transmission data is obtained and input into the trained online neural network for targeted preprocessing. The overall nonlinear characteristics of the power amplifier (PA) are then compensated through the offline neural network. Finally, the power amplifier outputs the signal and transmits it to the antenna system for information transmission.

2. The digital predistortion method based on a layered iterative neural network according to claim 1, wherein: The process of constructing the training dataset includes: Based on orthogonal frequency division multiplexing technology, different OFDM time domain signals txWave are generated according to different bandwidths and modulations. During the generation process, different bandwidths and modulation modes are configured, and OFDM parameters are set accordingly. OFDM parameters include: subcarrier spacing, number of fast Fourier transform points, cyclic prefix length, upsampling factor, and sampling frequency; The bit length of the transmission data block is calculated based on the modulation order and the number of OFDM subcarriers carrying data. Then, a long transmission data block is randomly generated. The transmission data block is modulated to generate an OFDM time-domain baseband signal. After upsampling, the OFDM baseband signal used in the training set is obtained. This step is repeated to obtain the baseband signal txWaveTrain of the training set, the baseband signal txWaveVal of the validation set, the baseband signal txWaveTest of the test set, and the baseband signal txWaveEval of the evaluation set under different modulation modes and bandwidths. The txWaveTrain, txWaveVal, txWaveTest, and txWaveEval of different modulation modes and bandwidths are respectively input into the power amplifier PA to obtain the corresponding output signals paOutputTrain, paOutputVal, paOutputTest, and paOutputEval; the input signals are paired with the output signals to generate the training set Train = (txWaveTrain, paOutputTrain), the verification set Val = (txWaveVal, paOutputVal), the test set Test = (txWaveTest, paOutputTest), and the evaluation set Eval = (txWaveEval, paOutputEval).

3. The digital predistortion method based on a layered iterative neural network according to claim 2, wherein: The training process for the offline neural network model MultiOfflineNet is as follows: The training set signal paOutputTrain, validation set signal paOutputVal, and test set signal paOutputTest output by the power amplifier corresponding to different modulation modes and bandwidths are processed by the linear gain factor lineGainFactor of the power amplifier, that is, each signal is divided by the linear gain factor lineGainFactor, and then combined with the corresponding configured txWaveTrain, txWaveVal, and txWaveTest to form a mixed training data set of the MultiOfflineNet neural network, where the mixed training data set includes a mixed training set MixTrain = (txMixWaveTrain, paMixOutputTrain), a mixed validation set MixVal = (txMixWaveVal, paMixOutputVal), and a mixed test set MixTest = (txMixWaveTest, paMixOutputTest); Perform normalization preprocessing on each data set in the mixed training data set to obtain corresponding data, where the data includes training data MtxTrain = (inputMtxTrain, outputMtxTrain), verification data MtxVal = (inputMtxVal, outputMtxVal), and test data MtxTest = (inputMtxTest, outputMtxTest); Then, the offline neural network MultiOfflineNet is trained using the training data MtxTrain = (inputMtxTrain, outputMtxTrain). During the training process, the verification data MtxVal = (inputMtxVal, outputMtxVal) is used for verification, and the test data MtxTest = (inputMtxTest, outputMtxTest) is used for testing.

4. The digital predistortion method based on a layered iterative neural network according to claim 3, wherein: The normalization preprocessing process for each data set in the mixed training data set is: Split the normalized paMixOutputTrain and paMixOutputVal into an in-phase signal I and a quadrature signal Q; Then, according to the size of the memory depth memDepth and the nonlinear order nonlinearDegree, a matrix of corresponding size is formed. The matrix obtained by preprocessing is as follows: Where m represents the memory depth and q represents the nonlinear order; Split the normalized txMixWaveTrain and txMixWaveVal into the in-phase signal I and the quadrature signal Q, and then combine them to obtain outputMtxTrain and outputMtxVal; inputMtxTrain and outputMtxTrain constitute the training data; Similarly, inputMtxVal and outputMtxVal constitute the verification data.

5. The digital predistortion method based on a layered iterative neural network according to claim 2, wherein: The training process for the online neural network model MultiOnlineNet is as follows: Select txWaveTest data corresponding to one of the modulation modes and bandwidths. After the data txWaveTest passes through the power amplifier PA, paOutputOnTest is generated. Divide paOutputOnTest by the linear gain factor lineGainFactor of the PA, normalize txWaveTest and paOutputOnTest, and then preprocess them. Generate inputMtxTest data from paOutputOnTest as the input of the neural network, and generate paOutputTestMtx as the output of the neural network by normalizing the txWaveTest data. Use inputMtxTest and paOutputTestMtx to train the online neural network MultiOnlineNet; when the loss value of the online neural network is less than the specified threshold, stop training and save the online deep neural network MultiOnlineNet.

6. The digital predistortion method based on a layered iterative neural network according to claim 5, characterized in that: The training data of the online neural network is obtained in small batches, the gradient and loss of the online neural network model are evaluated, and the network parameters are updated. The termination condition of online training is defined as the maximum number of training rounds or the loss value reaches the threshold. The specific process of calculating the online neural network loss value is as follows: the small batch inputMtx data is passed through the current online training neural network MultiOnlineNet to obtain the predicted value Y of the online neural network. The online neural network loss value loss is expressed as the mean squared error: Among them, N represents the number of elements in Y and T, Y i represents the predicted value of the online neural network, T i represents the true value; The preset normalized mean square error NMSE value is used as the loss threshold, where the relationship between the NMSE value and the loss value loss is NMSE = 10*log 10 (loss).

7. The digital predistortion method based on a layered iterative neural network according to claim 2, wherein: The process of performance testing and evaluation for the completed offline training model and online training model is as follows: The evaluation signals txWaveEval with different modulation modes and bandwidths are passed through the online neural network MultiOnlineNet and the offline neural network MultiOfflineNet, and finally through the power amplifier PA to obtain paOutputOnEval; The adjacent channel power ratio (ACPR), normalized mean square error (NMSE), and error vector magnitude (EVM) of txWaveEval and paOutputOnEval for different modulation modes and bandwidths are calculated in turn. The model performance is evaluated based on these evaluation indicators. If the normalized mean square error (NMSE) is greater than the preset threshold during the evaluation, the online neural network model training process is continued.

8. The digital predistortion method based on a layered iterative neural network according to claim 1, characterized in that: The offline neural network MultiOfflineNet consists of a feature input layer, a fully connected output layer, and three fully connected layers between the input and output layers. Each fully connected layer includes several neurons, and each fully connected layer uses the leakyRelu function as the activation function. The number of neurons in each fully connected layer is 64.

9. The digital predistortion method based on a layered iterative neural network according to claim 1, characterized in that: The online neural network MultiOnlineNet consists of a feature input layer, an FC output layer, and two fully connected layers between the input layer and the output layer. Each fully connected layer uses the leakyRelu function as the activation function, and the number of neurons in each fully connected layer is 30.

10. A system for executing the digital predistortion method based on a layered iterative neural network according to any one of claims 1 to 9, characterized in that: The system includes: a data preparation module, an offline training module, an online training module, a performance evaluation module and a pre-distortion processing module, wherein: The data preparation module is used to generate OFDM time-domain signals with different modulation modes and bandwidths and establish corresponding data sets. The offline training module trains the offline neural network based on the established data sets to obtain a deeply optimized offline neural network. The online training module processes the data set according to the trained offline neural network, and trains the online neural network based on the data set to obtain an online neural network for specific signal processing; The performance evaluation module uses the evaluation data set to perform performance testing on the offline and online training models. If the normalized mean square error threshold between the OFDM signal and the PA output data is greater than the expected set value, the online neural network model training will continue. The pre-distortion processing module includes a baseband processor, an online neural network, an offline neural network, and a power amplifier PA. The baseband signal sent by the baseband processor first passes through the trained online neural network MultiOnlineNet and the offline neural network MultiOfflineNet, then passes through the power amplifier PA to couple the PA output signal, undergoes down-conversion, and obtains the paOutputOnEval through analog-to-digital conversion sampling ADC, and finally transmits the paOutputOnEval signal to the antenna system.

Images