A Digital Predistortion Optimization Method for Power Amplifiers Based on Complex Convolutional Neural Networks

Abstract

This invention discloses a method for optimizing digital pre-distortion (DPD) for power amplifiers (PAs) based on a complex convolutional neural network. The method comprises the following steps: S1. Collecting multiple sets of digital signals before and after a power amplifier (PA) to form a sample set; S2. Preprocessing the data in the sample set; S3. Constructing a complex time-delay convolutional neural network containing a linear auxiliary branch and a nonlinear model as the DPD model for the PA; S4. Training the DPD model using the preprocessed data in the sample set to obtain trained DPD model parameters; and S5. When performing DPD on the PA, the digital signal to be transmitted is first fed into the trained DPD model for processing. While maintaining low complexity, the method significantly improves linearization capability by introducing complex domain processing and an auxiliary linear branch.

Description

Technical Field

[0001] The present invention relates to the field of digital predistortion, and in particular to a power amplifier digital predistortion optimization method based on a complex convolutional neural network. Background Art

[0002] With the development of 5G and future 6G communication systems, the demand for high data rates and wide bandwidth continues to increase. However, when amplifying RF signals, power amplifiers (PAs) experience inherent nonlinearities and memory effects, which can cause signal distortion and adjacent channel interference, thereby reducing the transmission quality and spectral efficiency of the communication system. Digital predistortion (DPD) technology, due to its advantages in linearization performance and stable reliability, has attracted considerable attention in the field of amplifier linearization. Essentially, a DPD module is added to generate a predistorted signal with the opposite nonlinear characteristics of the PA to compensate for the PA's nonlinear distortion.

[0003] Building an accurate behavioral model of a power amplifier (PA) is fundamental to implementing digital predistortion (DPD) technology. Traditional models, such as the Volterra model and the generalized memory polynomial (GMP) model, experience significant increases in model complexity in ultra-wideband (UWB) applications, making it difficult to strike a balance between implementation difficulty and accuracy. In recent years, neural networks (NNs) have been introduced to PA modeling due to their powerful ability to approximate nonlinear functions. Recurrent neural networks (RNNs), due to their built-in memory structure, can effectively address PA memory effects and were first demonstrated to be viable for PA modeling. However, RNNs have limited linearization capabilities, complex network structures, and long training times. Subsequently, various deep neural networks (DNNs) have been introduced. These models, taking in-phase and quadrature (I / Q) components of a specific memory depth as input, along with other envelope-related terms, demonstrate excellent linearization capabilities. However, the fully connected (FC) network structure of DNNs results in a large number of model parameters and high computational complexity, especially when processing wideband signals. To reduce model complexity, convolutional neural networks (CNNs) were introduced. CNNs significantly reduce the number of model parameters through weight sharing in convolutional layers. The DPD method based on the real-valued time-delay convolutional neural network (RVTDCNN) extracts the nonlinear characteristics of PA through convolutional layers, reducing the number of parameters by 50% while maintaining comparable linearization performance to more complex models. However, existing RVTDCNN methods still have significant room for improvement in nonlinear compensation performance, especially when processing strongly nonlinear signals, where their linearization capabilities are limited. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a power amplifier digital predistortion optimization method based on complex convolutional neural network, which significantly improves the linearization capability by introducing complex domain processing and auxiliary linear branches while maintaining low complexity characteristics.

[0005] The objective of the present invention is achieved through the following technical solution: a power amplifier digital predistortion optimization method based on a complex convolutional neural network, comprising the following steps:

[0006] S1. Collect multiple sets of digital signals before and after the power amplifier PA to form a sample set;

[0007] In step S1, the acquisition process of any group of digital signals before and after passing through the power amplifier PA is as follows:

[0008] The digital signal x(n) to be sent is transmitted to the power amplifier PA after passing through the DAC and up-converter. The power amplifier PA amplifies the received signal and then transmits it. The signal output by the power amplifier is down-converted by the down-converter and transmitted to the ADC for conversion to obtain the digital signal y(n). x(n) and y(n) are a set of digital signals before and after passing through the power amplifier PA.

[0009] S2. Preprocess the data in the sample set, including:

[0010] S201. For any set of digital signals {x(n), y(n)} in the sample set, first normalize y(n) to obtain Where G is the ideal gain of the power amplifier PA;

[0011]

[0012] Among them I x [n] and Q x [n],I y [n] and Q y [n], are the in-phase and quadrature components of x(n) and y(n), respectively;

[0013] S202. Perform sliding window processing on y′(n), where the window length is L, to obtain window data Y n =[y′(n),y′(n-1),…,y′(n-L+1)] T ;

[0014] Preprocess y(n) to obtain the window data Y n =[y′(n),y′(n-1),…,y′(n-L+1)] T , as the window data corresponding to x(n);

[0015] S203. Repeat steps S201 to S202 for each group of digital signals in the sample set to complete preprocessing of various digital signals in the sample set.

[0016] S3. Construct a complex time-delay convolutional neural network containing a linear auxiliary branch and a nonlinear model as a digital predistortion model for the power amplifier;

[0017] In the complex time-delay convolutional neural network constructed in step S2;

[0018] The linear auxiliary branch consists of a fully connected layer with L input features and 1 output feature. It is used to perform a linear transformation on the complex input and generate the initial prediction value. The neuron parameters in the linear auxiliary branch are trained using the gradient descent algorithm.

[0019] The nonlinear model consists of multiple cascaded one-dimensional complex convolutional layers and fully connected layers, optimized using the Adam algorithm;

[0020] The outputs of the linear auxiliary branch and the nonlinear model are combined through an adder as the final output of the complex time-delay convolutional neural network.

[0021] S4. Use the pre-processed data from the sample set to train the digital pre-distortion model of the power amplifier to obtain the trained digital pre-distortion model parameters;

[0022] The step S4 comprises:

[0023] S401. Training linear auxiliary branch:

[0024] A1. Window data Y obtained by preprocessing any set of digital signals in the sample set n =[y′(n),y′(n-1),…,y′(n-L+1)] T , input it into the linear auxiliary branch, and the linear auxiliary branch outputs the linearly transformed result y linear (n);

[0025] A2. Using window data Y n The corresponding x(n) and y linear (n) Calculate the loss function loss1:

[0026] loss1=|x(n)-y linear (n)| 2

[0027] A3. Based on the loss function loss1, the gradient descent algorithm is used to update and optimize the linear auxiliary branch.

[0028] A4. Repeat steps A1 to A3 for different window data obtained by digital signal processing in the sample set until the linear auxiliary branch converges, that is, loss1 is less than a preset convergence threshold;

[0029] S402. Fix the parameters of the linear auxiliary branch and train the nonlinear model:

[0030] B1. Window data Y obtained by preprocessing any set of digital signals in the sample set n =[y′(n),y′(n-1),…,y′(n-L+1)] T , input them into the linear auxiliary branch and the nonlinear model respectively, and then the adder combines the output results of the linear auxiliary branch and the nonlinear model to obtain the output result y of the digital predistortion model DPD (n);

[0031] B2. Using window data Y n The corresponding x(n) and y DPD (n) Calculate the loss function loss2:

[0032] loss2=|x(n)-y DPD (n)| 2

[0033] B3. Based on the loss function loss2, the Adam algorithm is used to update and optimize the nonlinear model;

[0034] B4. Repeat steps B1 to B3 for different window data obtained by digital signal processing in the sample set until the model converges, that is, loss2 is less than a preset convergence threshold;

[0035] S5. When performing actual digital pre-distortion of a power amplifier, the digital signal to be transmitted is first sent to a well-trained digital pre-distortion model for processing, and then transmitted to the power amplifier PA through a DAC and an up-converter. The power amplifier PA amplifies the received signal and then transmits it.

[0036] The beneficial effect of the present invention is that the present invention significantly improves the linearization capability by introducing complex domain processing and auxiliary linear branches while maintaining the low complexity characteristic. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the overall principle of the present invention;

[0038] Figure 2 Schematic diagram of the spectrum comparison between CVTDCNN and other models;

[0039] Figure 3 Schematic diagram of the experimental platform setup. DETAILED DESCRIPTION

[0040] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0041] like Figure 1 As shown, a power amplifier digital predistortion optimization method based on a complex convolutional neural network is characterized by comprising the following steps:

[0042] S1. Collect multiple sets of digital signals before and after the power amplifier PA to form a sample set;

[0043] In step S1, the acquisition process of any group of digital signals before and after passing through the power amplifier PA is as follows:

[0044] The digital signal x(n) to be sent is transmitted to the power amplifier PA after passing through the DAC and up-converter. The power amplifier PA amplifies the received signal and then transmits it. The signal output by the power amplifier is down-converted by the down-converter and transmitted to the ADC for conversion to obtain the digital signal y(n). x(n) and y(n) are a set of digital signals before and after passing through the power amplifier PA.

[0045] S2. Preprocess the data in the sample set;

[0046] In step S2, preprocessing the data in the sample set includes:

[0047] S201. For any set of digital signals {x(n), y(n)} in the sample set, first normalize y(n) to obtain Where G is the ideal gain of the power amplifier PA;

[0048]

[0049] Among them I x [n] and Q x [n],I y [n] and Q y [n], are the in-phase and quadrature components of x(n) and y(n), respectively;

[0050] S202. Perform sliding window processing on y′(n), where the window length is L, to obtain window data Y n =[y′(n),y′(n-1),…,y′(n-L+1)] T ;

[0051] Preprocess y(n) to obtain the window data Y n =[y′(n),y′(n-1),…,y′(n-L+1)] T , as the window data corresponding to x(n);

[0052] S203. Repeat steps S201 to S202 for each group of digital signals in the sample set to complete preprocessing of various digital signals in the sample set.

[0053] S3. Construct a complex time-delay convolutional neural network containing a linear auxiliary branch and a nonlinear model as a digital predistortion model for the power amplifier;

[0054] This paper uses CVNNs as the core modeling tool. Compared to traditional real-valued neural networks (RVNNs), CVNNs can directly process complex inputs and outputs without decomposing the complex signal into real and imaginary parts. This not only reduces preprocessing steps but also improves computational efficiency and reduces the number of model parameters, significantly improving model training and operation efficiency. CVNNs also fully utilize the signal's amplitude and phase information, allowing for more efficient feature extraction and improved model performance.

[0055] In this paper, CVNNs are used to construct a complex time-delay convolutional neural network (CVTDCNN). Through the complex convolutional layer, CVTDCNN can directly extract the nonlinear features of the input signal, maintain the low complexity characteristics, and show superior performance when processing strong nonlinear signals.

[0056] In order to further improve the performance of the complex time-delay convolutional neural network (CVTDCNN), the present invention introduces the Boosting algorithm. Boosting is an integrated learning method that iteratively trains multiple weak learners and combines them into a strong learner, thereby significantly improving the predictive ability and robustness of the model. In the present invention, CVTDCNN combines a linear auxiliary branch and a nonlinear model, wherein the linear auxiliary branch acts as a weak learner to generate an initial prediction value; the nonlinear model further fits the residual of the linear auxiliary branch and focuses on processing the nonlinear part of the signal. The final output is the sum of the two. By introducing the Boosting algorithm, CVTDCNN significantly reduces the adjacent channel leakage ratio (ACLR) when processing strong nonlinear signals, thereby improving the transmission quality and spectrum efficiency of the communication system.

[0057] The network structure of CVTDCNN is as follows Figure 1 As shown in Figure 1, the proposed method consists of two main components: a linear auxiliary branch and a nonlinear model. The linear auxiliary branch uses a fully connected (FC) layer (complex linear layer) to generate initial predictions and process the linear portion of the signal. The nonlinear model consists of multiple cascaded one-dimensional complex convolutional layers and fully connected layers. This structure provides an efficient and reliable solution for power amplifier linearization in future 6G and higher-generation communication systems.

[0058] S4. Use the pre-processed data from the sample set to train the digital pre-distortion model of the power amplifier to obtain the trained digital pre-distortion model parameters;

[0059] The training process of the digital predistortion model adopts an indirect learning method, that is, the power amplifier output signal is converted into a digital signal, and after normalization and window processing, it is used as the input of the predistortion model, and the digital signal x(n) to be sent is used as the output of the model for training. Specifically, step S4 includes:

[0060] S401. Training linear auxiliary branch:

[0061] A1. Window data Y obtained by preprocessing any set of digital signals in the sample set n =[y′(n),y′(n-1),…,y′(n-L+1)] T , input it into the linear auxiliary branch, and the linear auxiliary branch outputs the linearly transformed result y linear (n);

[0062] A2. Using window data Y n The corresponding x(n) and y linear (n) Calculate the loss function loss1:

[0063] loss1=|x(n)-y linear (n)| 2

[0064] A3. Based on the loss function loss1, the gradient descent algorithm is used to update and optimize the linear auxiliary branch.

[0065] A4. Repeat steps A1 to A3 for different window data obtained by digital signal processing in the sample set until the linear auxiliary branch converges, that is, loss1 is less than a preset convergence threshold;

[0066] S402. Fix the parameters of the linear auxiliary branch and train the nonlinear model:

[0067] B1. Window data Y obtained by preprocessing any set of digital signals in the sample set n =[y′(n),y′(n-1),…,y′(n-L+1)] T , input them into the linear auxiliary branch and the nonlinear model respectively, and then the adder combines the output results of the linear auxiliary branch and the nonlinear model to obtain the output result y of the digital predistortion model DPD (n);

[0068] B2. Using window data Y n The corresponding x(n) and y DPD(n) Calculate the loss function loss2:

[0069] loss2=|x(n)-y DPD (n)| 2

[0070] B3. Based on the loss function loss2, the Adam algorithm is used to update and optimize the nonlinear model;

[0071] B4. Repeat steps B1 to B3 for different window data obtained by digital signal processing in the sample set until the model converges, that is, loss2 is less than a preset convergence threshold;

[0072] S5. When performing actual digital pre-distortion of a power amplifier, the digital signal to be transmitted is first sent to a well-trained digital pre-distortion model for processing, and then transmitted to the power amplifier PA through a DAC and an up-converter. The power amplifier PA amplifies the received signal and then transmits it.

[0073] In some embodiments, the sample set can be divided into a training set, a test set and a validation set. The training process in step S4 above can use the training set obtained by dividing the sample set, and the validation set and the test set are used to verify and test the training results.

[0074] In the embodiments of the present application, in the present invention, in order to verify the effectiveness of the CVTDCNN model in the application of power amplifier (PA) digital predistortion (DPD), we use two main performance evaluation indicators: normalized mean square error (NMSE) and adjacent channel leakage ratio (ACLR). The lower the NMSE value, the smaller the error between the output of the model and the ideal output, and the better the linearization performance of the model. The lower the ACLR value, the less signal leakage in the adjacent channel, the higher the spectrum efficiency, and the less interference to other channels. These two indicators can fully reflect the linearization performance of the model and the impact on the signal spectrum. NMSE and ACLR are defined as:

[0075]

[0076] ACLR (dB) =P Adj(dB) -P Main(dB) .

[0077] In machine learning, when the model size is small, the initialization of the parameters can have a significant impact on the final solution. To reduce the impact of random parameter initialization on performance evaluation results, this application trains the model three times, each time using different random initialization parameters, independently on the same dataset to reduce the chance of results caused by specific initialization. The mean and variance of the NMSE and ACLR values obtained from all training runs are calculated.

[0078] 1. Data preparation:

[0079] The data is sampled Orthogonal Frequency Division Multiplexing (OFDM) data, where the data sampled by the PA transmitter is x(n) and the data sampled by the receiver is y(n);

[0080] Normalize y(n) to get y′(n);

[0081] The data is divided into training set, validation set and test set in a ratio of 8:1:1;

[0082] Perform sliding window processing on y′(n), where the window length is L, and obtain the window data Y n =[y′(n),y′(n-1),…,y′(n-L+1)] T , as the network input.

[0083] 2. Model initialization:

[0084] First, create the network structure of the CVTDCNN model;

[0085] The Kaiming initialization method is used to initialize the neuron parameters in the model. This method helps prevent the gradient from disappearing or exploding in the early stages of training, thereby accelerating convergence and improving model performance.

[0086] 3. Calculate the linear auxiliary branch parameters:

[0087] The gradient descent (GD) algorithm is used to calculate and update the parameters of the neurons in the model.

[0088] 4. Training nonlinear models:

[0089] Fixed linear auxiliary branch coefficients ensure that the nonlinear model focuses on learning the nonlinear part of the signal, improving the overall performance of the model;

[0090] The nonlinear model is trained on the dataset, and the model parameters are updated through the back-propagation algorithm to further optimize the MSE loss function.

[0091] 5. Output:

[0092] After training is completed, the final digital pre-distortion (DPD) model is output.

[0093] like Figure 2As shown, in the embodiments of the present application, a specific hardware and software platform is used to verify the performance of the CVTDCNN model. The test signal used in the experiment is a 20MHz bandwidth OFDM signal with a peak-to-average power ratio (PAPR) of 13.2dB. The signal is generated in MATLAB on a personal computer (PC), using 8PSK modulation and 2048 subcarriers. The generated test signal is processed by a software-defined radio (SDR) platform and the baseband signal is up-converted to a center frequency of 2.3GHz, and then transmitted to a PA system consisting of two cascaded ZX60-V82-S+PAs with a frequency range of 20MHz-6GHz. The SDR captures 65536 output samples and trains the network on the PC using Python together with the transmitted data. In the embodiments of the present application, the digital signal can be power controlled, analog-to-digital converted, and up-converted by the SDR (software defined radio platform) before being transmitted to the power amplifier; by controlling the signal power before the PA input, corresponding to the degree of various nonlinear distortions, different nonlinear distortion scenarios can be simulated according to the method of the present application. The parameter settings in the CVTDCNN network are shown in the following table:

[0094] Table 1 Parameter settings in CVTDCNN network

[0095]

[0096] The CVTDCNN training parameter settings are described in the following table:

[0097] Table 2 CVTDCNN training parameter settings

[0098]

[0099] After testing the CVTDCNN model, it was noticed that ACLR did not significantly improve with the increase of memory depth M, but instead showed an increasing trend, as shown in the following table:

[0100] Table 3 Effect of different memory depths on NMSE and ACLR of CVTDCNN model

[0101]

[0102] Therefore, considering the balance between model performance and computational complexity, this patent sets M=4 as the memory depth for subsequent experiments.

[0103] When the transmit signal power is set to 7dBm to 11dBm, it can simulate the different transmission power requirements that may be encountered in actual communication systems. The CVTDCNN model including the linear auxiliary branch improves the ACLR by an average of 1.5dB, as shown in the following table:

[0104] Table 4 Impact of non-linear auxiliary branches on the NMSE and ACLR of the model

[0105]

[0106] The table shows that the linear auxiliary branch helps to further reduce the signal leakage in adjacent channels, thereby improving the spectrum efficiency and communication quality.

[0107] In order to verify the performance of the CVTDCNN model proposed in this invention, this patent conducts a detailed comparative analysis with several other neural network-based models. These models are divided into: traditional methods and neural network methods. Traditional methods include the generalized memory polynomial (GMP) model, and neural network methods include the real-valued time-delay convolutional neural network (RVTDCNN), the gated recurrent unit (GRU), and the deep gated recurrent unit (DGRU). The purpose of the comparative analysis is to evaluate the differences in digital pre-distortion (DPD) performance of different models, especially in the key indicator of ACLR.

[0108] As can be seen from Table 5, CVTDCNN shows excellent linearization performance at all tested signal powers, and its ACLR value is significantly lower than that of other models. Compared with RVTDCNN, the ACLR of the CVTDCNN model is reduced by 1.7dB on average, and the maximum reduction reaches 2.2dB when the transmit power is 9dBm, showing its superior performance in processing strong nonlinear signals. At this power level, Figure 3 The spectra of the received signals after various processing methods are shown, clearly demonstrating the superior ability of the CVTDCNN model in suppressing sideband power.

[0109] Table 5 Comparison of ALCR of CVTDCNN and other models

[0110]

[0111] The foregoing description is a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein and should not be construed as excluding other embodiments. Instead, the present invention can be used in other combinations, modifications, and environments and can be modified within the scope of the concept described herein through the above teachings or techniques or knowledge in the relevant field. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention are intended to be protected by the appended claims.

Claims

1. A power amplifier digital predistortion optimization method based on complex convolutional neural network, characterized by: The following steps are involved: S1. Collect multiple sets of digital signals before and after the power amplifier PA to form a sample set; S2. Preprocess the data in the sample set; In step S2, preprocessing the data in the sample set includes: S201. For any set of digital signals {x(n), y(n)} in the sample set, first normalize y(n) to obtain Where G is the ideal gain of the power amplifier PA; Among them I x [n] and Q x [n],I y [n] and Q y [n], are the in-phase and quadrature components of x(n) and y(n), respectively; S202. Perform sliding window processing on y′(n), where the window length is L, to obtain window data Y n =[y′(n),y′(n-1),…,y′(n-L+1)] T ; Preprocess y(n) to obtain the window data Y n =[y′(n),y′(n-1),…,y′(n-L+1)] T , as the window data corresponding to x(n); S203. For each set of digital signals in the sample set, steps S201 to S202 are repeated to complete the preprocessing of various digital signals in the sample set; S3. Construct a complex time-delay convolutional neural network containing a linear auxiliary branch and a nonlinear model as a digital predistortion model for the power amplifier; In the complex time-delay convolutional neural network constructed in step S3; The linear auxiliary branch consists of a fully connected layer with L input features and 1 output feature. It is used to perform a linear transformation on the complex input and generate the initial prediction value. The parameters in the linear auxiliary branch are trained using the gradient descent algorithm. The nonlinear model consists of multiple cascaded one-dimensional complex convolutional layers and fully connected layers, optimized using the Adam algorithm; The outputs of the linear auxiliary branch and the nonlinear model are combined through an adder as the final output of the complex time-delay convolutional neural network; S4. Use the pre-processed data from the sample set to train the digital pre-distortion model of the power amplifier to obtain the trained digital pre-distortion model parameters; S5. When performing actual digital pre-distortion of a power amplifier, the digital signal to be transmitted is first sent to a well-trained digital pre-distortion model for processing, and then transmitted to the power amplifier PA through a DAC and an up-converter. The power amplifier PA amplifies the received signal and then transmits it.

2. The method for optimizing digital predistortion of a power amplifier based on a complex convolutional neural network according to claim 1, wherein: In step S1, the acquisition process of any group of digital signals before and after passing through the power amplifier PA is as follows: The digital signal x(n) to be sent is transmitted to the power amplifier PA after passing through the DAC and up-converter. The power amplifier PA amplifies the received signal and then transmits it. The signal output by the power amplifier is down-converted by the down-converter and transmitted to the ADC for conversion to obtain the digital signal y(n). {x(n), y(n)} is a set of digital signals before and after passing through the power amplifier PA.

3. The method for optimizing power amplifier digital predistortion based on a complex convolutional neural network according to claim 1, wherein: The step S4 comprises: S401. Training linear auxiliary branch: A1. Window data Y obtained by preprocessing any set of digital signals in the sample set n =[y′(n),y′(n-1),…,y′(n-L+1)] T , input it into the linear auxiliary branch, and the linear auxiliary branch outputs the linearly transformed result y linear (n); A2. Using window data Y n The corresponding x(n) and y linear (n) Calculate the loss function loss1: loss1=|x(n)-y linear (n)| 2 A3. Based on the loss function loss1, the gradient descent algorithm is used to update and optimize the linear auxiliary branch. A4. Repeat steps A1 to A3 for different window data obtained by digital signal processing in the sample set until the linear auxiliary branch converges, that is, loss1 is less than a preset convergence threshold; S402. Fix the parameters of the linear auxiliary branch and train the nonlinear model: B1. Window data Y obtained by preprocessing any set of digital signals in the sample set n =[y′(n),y′(n-1),…,y′(n-L+1)] T , input them into the linear auxiliary branch and the nonlinear model respectively, and then the adder combines the output results of the linear auxiliary branch and the nonlinear model to obtain the output result y of the digital predistortion model DPD (n); B2. Using window data Y n The corresponding x(n) and y DPD (n) Calculate the loss function loss2: loss2=|x(n)-y DPD (n)| 2 B3. Based on the loss function loss2, the Adam algorithm is used to update and optimize the nonlinear model; B4. Repeat steps B1 to B3 for window data obtained by different digital signal processing in the sample set until the model converges, that is, loss2 is less than a preset convergence threshold.

Images