Adaptive power amplifier linearization method and device based on deep learning

Abstract

The invention discloses an adaptive power amplifier linearization method and device based on deep learning. The method comprises the following steps: acquiring a to-be-processed input signal; processing the input signal to be processed by adopting a trained hybrid network model to obtain a processed pre-distortion signal; wherein the trained hybrid network model takes data of a preset category as a training data set, samples in the training data set are selected, the selected samples and the unselected samples are endowed with different weights, the initial hybrid network model is trained, and in different stages of training, the selected samples and the unselected samples are subjected to different weights. Dynamically adjusting the weight of the samples in the training data set, wherein the data of the preset category comprises an input signal, an envelope related item, a lead item and a time delay item of the input signal, and a lead item and a time delay item of the envelope related item; and inputting the processed pre-distortion signal to a power amplifier for amplification to obtain a linearized output signal. The linearity of the power amplifier can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication, and particularly relates to an adaptive power amplifier linearization method and device based on deep learning. Background Art

[0002] As one of the key components in a wireless communication system, a power amplifier (PA) with good linearity is a prerequisite for ensuring the communication quality of the entire wireless communication system. However, currently transmitted signals generally have characteristics such as wide bandwidth and high peak-to-average ratio, and serious signal distortion will occur after they are transmitted through a power amplifier. Therefore, effective power amplifier linearization technology is required. Due to advantages such as moderate complexity and bandwidth, digital predistortion (DPD) technology has become the current mainstream power amplifier linearization technology. When device aging, environmental temperature changes, or the characteristics of the transmitted signal change, the characteristics of the power amplifier will also change accordingly, thereby affecting the linearization performance. The adaptive digital predistortion technology can update the parameters of the predistortion model in a timely manner according to the real-time collected data, which can well solve the above problems and is widely used in engineering. In recent years, many progresses have been made in the research of combining neural networks with digital predistortion technology. Some research has introduced a digital predistortion network model based on deep learning into an adaptive architecture to give full play to the superior performance of neural networks in fitting complex non-linear functions. However, a large amount of computing resources will be consumed during the training and inference processes of the neural network model, and the training efficiency also greatly affects the performance of adaptive digital predistortion.

[0003] In related technologies, digital predistortion based on polynomials is combined with a sample selection method. According to specific selection criteria, such as selecting samples based on the power characteristics of the signal or the memory effect, selection is performed from the signal data sampled from the feedback loop. The selected samples are used to construct a feature matrix, and the coefficients of the predistortion model are calculated; for digital predistortion based on deep learning, a neural network model based on CNN (Convolutional Neural Network) or RNN (Recurrent Neural Network) is constructed to fit the behavior characteristics of the power amplifier and then construct a predistortion model, and further apply it in an adaptive architecture. However, in the existing digital predistortion based on polynomials, due to the high correlation between polynomial basis functions, when the number of model parameters is approximately the same, compared with digital predistortion based on deep learning, its linearization performance is lower, and when the transmitted signal is relatively complex, the modeling accuracy is poor; the existing digital predistortion models based on deep learning focus on reducing the complexity of the model itself, but still require a large amount of training data to participate in model training, and there is a large amount of redundant data in the actually sampled data that contributes little to model establishment, which greatly increases the computing cost and training time.

[0004] Therefore, it is of great significance to provide a deep learning-based power amplifier linearization method that is more suitable for adaptive architectures. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention provides a deep learning-based adaptive power amplifier linearization method and device. The technical problems to be solved by the present invention are realized through the following technical solutions:

[0006] In a first aspect, the present invention provides a deep learning-based adaptive power amplifier linearization method, including:

[0007] An input signal to be processed;

[0008] Using a trained hybrid network model to process the input signal to be processed to obtain a processed predistortion signal; wherein, the trained hybrid network model uses data of a preset category as a training data set, selects samples in the training data set, assigns different weights to the selected samples and the unselected samples, trains the initial hybrid network model, and dynamically adjusts the weights of the samples in the training data set at different stages of training to obtain, the data of the preset category includes an input signal, an envelope-related term, an advanced term and a time-delay term of the input signal, and an advanced term and a time-delay term of the envelope-related term;

[0009] Inputting the processed predistortion signal into a power amplifier for amplification to obtain a linearized output signal.

[0010] In a second aspect, the present invention further provides a deep learning-based adaptive power amplifier linearization device, including a processor, a communication interface, a memory, and a communication bus, and the processor, the communication interface, and the memory complete communication with each other through the communication bus;

[0011] The memory is used to store a computer program;

[0012] The processor is used to implement the method mentioned above when executing the program stored on the memory.

[0013] Advantages of the present invention:

[0014] 1. For the deep learning-based adaptive power amplifier linearization method and device provided by the present invention, the predistortion network model with a CNN-RNN hybrid structure has better linearization performance than the polynomial-based digital predistortion and some predistortion network models under the condition of similar parameter quantities; the reparameterization method enables the network model in the present invention to utilize a multi-branch structure to obtain better training effects, and at the same time does not increase much inference time;

[0015] 2. The adaptive power amplifier linearization method and device based on deep learning provided by the present invention introduce a sample selection method into the training process of the neural network model. The present invention not only reduces the requirements of the neural network model for the amount of training data and the computational cost of the model, but also improves the training efficiency and linearization performance of the model, which is beneficial to the application of the neural network model in the adaptive predistortion technology.

[0016] The following will further elaborate on the present invention in conjunction with the accompanying drawings and embodiments. Description of the Drawings

[0017] Figure 1 is a flowchart of the adaptive power amplifier linearization method based on deep learning provided by the embodiment of the present invention;

[0018] Figure 2 is a schematic diagram of introducing future items into input data provided by the embodiment of the present invention;

[0019] Figure 3 is a schematic diagram of mapping one-dimensional data to two-dimensional data provided by the embodiment of the present invention;

[0020] Figure 4 is a schematic diagram of the sample selection process provided by the embodiment of the present invention;

[0021] Figure 5 is a schematic diagram of the CNN-RNN hybrid network model based on deep learning provided by the embodiment of the present invention;

[0022] Figure 6 is a schematic diagram of the CNN partial network structure and the reparameterization process provided by the embodiment of the present invention. Detailed Embodiments

[0023] The following further describes the present invention in detail with specific embodiments, but the embodiments of the present invention are not limited thereto.

[0024] The linearization technology of power amplifiers is an important means to improve the performance of wireless communication systems. Among them, digital predistortion (DPD) technology is a mainstream method. It inserts a predistorter that is inverse to the nonlinear characteristics of the power amplifier before the power amplifier, so as to obtain a linearly amplified output at the output end of the power amplifier. Given the powerful fitting ability of neural networks for complex nonlinear functions, combining neural networks with predistortion technology has become a new idea for researching and developing digital predistortion technology. However, in actual tests, the observation errors between the obtained model training samples are correlated. Therefore, in order to ensure the performance of the predistortion network model, a large amount of training data is required, which greatly increases the computational cost and the training time of the model. In addition, the relatively high complexity of the model itself will also increase the computational cost and the inference time. These problems are not conducive to the application of the predistortion network in the adaptive architecture.

[0025] In view of this, the present invention mainly proposes a low-complexity power amplifier linearization network model with a higher inference speed, which combines a sample selection method to reduce the total number of samples while covering key signal features, reduces the computational cost, assists in the training of the predistortion network model, and optimizes its training efficiency and linearization performance.

[0026] Please refer to Figure 1 , Figure 1 which is a flowchart of an adaptive power amplifier linearization method based on deep learning provided by an embodiment of the present invention. An adaptive power amplifier linearization method based on deep learning provided by the present invention includes:

[0027] S101. Obtain an input signal to be processed.

[0028] S102. Process the input signal to be processed using a trained hybrid network model to obtain a processed predistorted signal; wherein, the trained hybrid network model uses data of a preset category as a training data set, selects samples in the training data set, assigns different weights to the selected samples and the unselected samples, trains the initial hybrid network model, and dynamically adjusts the weights of the samples in the training data set at different stages of training. The data of the preset category includes an input signal, an envelope-related term, an advanced term and a delay term of the input signal, and an advanced term and a delay term of the envelope-related term.

[0029] Specifically, in this embodiment, the trained hybrid network model includes a trained pre-filtering layer, a trained time-series feature extraction layer, and a trained fully connected layer. The trained pre-filtering layer includes a trained convolutional layer. Processing the input signal to be processed using the trained hybrid network model to obtain a predistorted signal includes:

[0030] Perform a convolution operation on the input signal to be processed using the trained pre-filtering layer to obtain local spatial features;

[0031] Perform time-dimensional feature extraction on the local spatial features using the trained time-series feature extraction layer to obtain time-series features;

[0032] Process the time-series features using the trained fully connected layer to obtain the pre-distorted signal.

[0033] In this embodiment, the process of obtaining the trained hybrid network model includes:

[0034] Obtain data of multiple preset categories;

[0035] Preprocess the data of the preset categories, and use the preprocessed data of the preset categories as the training data set; the training data set includes multiple samples, and each sample includes the I and Q components of the input signal, the envelope correlation term, the leading term of the I and Q components of the input signal, the delay term of the I and Q components of the input signal, the leading term of the envelope correlation term, the delay term of the envelope correlation term, and the true test data of the I and Q components of the output signal;

[0036] Input part of the samples in the training data set into the p-th hybrid network model to be trained, and assign different weights to different samples to obtain the predicted data of the I and Q components of the output signal output during the p-th training process;

[0037] According to the difference between the predicted data of the I and Q components of the output signal output during the p-th training and the true test data of the I and Q components of the output signal in the samples used in the p-th training, calculate the fitting loss, and use it as the fitting loss during the p-th training process;

[0038] Perform backpropagation according to the fitting loss during the p-th training process to update the network parameters of the p-th trained hybrid network model to obtain the (p + 1)-th hybrid network model to be trained; iterate in this way until the number of training times or the convergence degree meets the preset conditions to obtain the trained hybrid network model.

[0039] In this embodiment, the preprocessing of the data of the preset categories includes:

[0040] Obtain the input and output signals of the power amplifier as the initial data of the preset category;

[0041] Use the method of cross-correlation to align the input and output signals in time series;

[0042] Convert the input and output signals after time series alignment into real signals of the I and Q components, and obtain the true test data of the I and Q components of the output signal;

[0043] Obtain the leading terms of the I and Q components of the input signal, as well as the delayed terms of the I and Q components of the input signal, the leading terms of the envelope correlation terms, and the delayed terms of the envelope correlation terms to construct a one-dimensional data set;

[0044] Convert the one-dimensional data set into a two-dimensional data set.

[0045] It can be understood that first, the input and output signals of the power amplifier measured by simulation or in reality are collected. Considering that there is a certain delay in the power amplifier itself, it is necessary to align the input and output data in time sequence by means of cross-correlation. It should be noted that the sampled data will be stored in the RAM, so the excitation at future times can be utilized.

[0046] Furthermore, considering that neural networks are more adept at processing real-valued data, it is necessary to process the input and output signals with time sequence alignment into real signals in two paths of I and Q; at the same time, due to the significant memory effect of the power amplifier, that is, the output of the power amplifier at each moment is related not only to the current input but also to the inputs at previous moments, it is necessary to consider the delayed terms of the input signal.

[0047] The behavior of a power amplifier with a strong memory effect can be strictly represented by a Volterra series in the form of a multi-dimensional impulse response, in the form of:

[0048]

[0049] where y(t) represents the response of the power amplifier at the current moment, x(t - τ i ) represents the excitation signal after a delay of τ i , i = 1, 2,..., ∞, h n (·) represents the nth-order Volterra kernel function;

[0050] Introduce the memory depth and truncate and approximate the above mathematical model as:

[0051]

[0052] where x(t - τ i ) represents the excitation signal after a delay of τ i , x(t + τ i ) represents the excitation signal advanced by τ i , T M represents the memory depth. From this, it can be noted that if the future excitation is available, it can be used to facilitate the modeling of the power amplifier and pre-distortion model.

[0053] Please refer to Figure 2 , Figure 2It is a schematic diagram of introducing input data for future items provided by an embodiment of the present invention. Let x k / 2-M / 2+1 ~x k / 2+M / 2 be used as inputs, and y be the corresponding model prediction outputs, k / 2 while is the corresponding true test data.

[0054] In addition, in order to introduce more abundant modeling features to reduce the training difficulty and increase the model accuracy, envelope-related terms of different orders are added to the input data.

[0055] In summary, the input data includes I and Q components, envelope-related terms, and leading and time-delay terms of both. Then, the formats of the input and output data are processed. The input data all consists of one-dimensional data that changes with time. Since the neural network model constructed in the present invention is a CNN-RNN hybrid structure, the data needs to be processed into a format suitable for convolution operations, as Figure 3 shown, Figure 3 is a schematic diagram of mapping one-dimensional data to two-dimensional data provided by an embodiment of the present invention. The one-dimensional data is mapped to two-dimensional data, where M represents the memory depth and k represents the non-linear order. Finally, the input data after format processing and the corresponding output data are encapsulated into a data set.

[0056] In this embodiment, different weights are assigned to different samples, including:

[0057] Dividing the upper and lower boundaries of different intervals of the histogram according to the amplitude of the input signal of the samples in the training data set to construct the histogram;

[0058] Using a preset genetic algorithm to calculate the optimal number of samples in different intervals of the histogram to optimize the histogram; among them, the difference between the output signal after linearizing the power amplifier and the ideal output signal is used as the value of the fitness function;

[0059] According to the optimal histogram, selecting the samples in each interval and obtaining the indexes of the selected samples;

[0060] During the training process, different weights are assigned to the selected samples and the unselected samples, and the weights are dynamically adjusted in different iteration times.

[0061] In this embodiment, the weights are dynamically adjusted in different iteration times, including:

[0062] Setting a mechanism for dynamically adjusting the weights, and its expression is:

[0063] ω=(1 - λ)·ω 1 +λ·ω 2 ;

[0064] Among them, λ represents the training progress ratio, that is, the current training times / total training times, ω 1 represents the weight of the selected sample, ω 2 represents the uniform weight, and ω represents the weight applied to the current training times.

[0065] It can be understood that, please refer to Figure 4 , Figure 4 is a schematic diagram of the sample selection process provided by the embodiment of the present invention. Based on the sample selection method of the genetic optimization histogram, the optimized histogram introduces the AM / AM characteristics of the power amplifier and the statistical distribution information of the transmitted signal into the sample selection process. Data with strong nonlinearity and high signal probability contribute more to the modeling. Therefore, more data are selected in these regions accordingly. First, the upper and lower boundaries of different regions (bins) of the histogram are divided according to the amplitude of the input signal, and then the optimal sample selection quantity in different bins is obtained through the genetic algorithm, so as to obtain the optimized histogram. The number of samples in different bins of the histogram represents the number of samples to be selected within the range of the signal amplitude. Finally, samples are selected according to the obtained optimal histogram, and the indexes of the selected samples are obtained. Specifically, the following markings are made for the division of the histogram: D j ={n i :θ j-1 <|x(n i )|<θ j}, where D j represents the set of the indexes of the selected samples in the j-th bin, θ j-1 and θ j respectively represent the upper limit and the lower limit of the j-th bin, n i represents the index of the selected sample. In addition, the total number of bins is denoted as J, and the number of samples in the j-th bin is denoted as d j . To ensure that data with different characteristics can be fully selected, J is set to 10. The d j of the histogram is generally set according to the probability distribution of the transmitted signal and the strong nonlinear region of the power amplifier, that is, more samples are selected in the strong nonlinear region and the high signal distribution probability region. The present invention selects to optimize d j through the genetic algorithm to minimize the normalized mean square error (NMSE) of the power amplifier output relative to the expected output. Specifically, first set d jThe upper and lower limits are [0, target sampling number]. Then, relevant parameters of the genetic algorithm and the fitness function are set. The relevant parameters are set as follows: population size 100, maximum number of generations 20, fraction of the next generation population generated by crossover 0.8, and mutation probability 1%. The fitness function guides the direction of algorithm optimization. Here, the error between the power amplifier output and the expected output is used as the measurement criterion. Based on the input-output data, a predistorter and a power amplifier model based on Generalized Memory Polynomial (GMP) are constructed. Considering that different predistortion learning structures have little influence on the result of sample selection, a relatively simple and efficient direct learning structure is adopted in the implementation of the fitness function. According to the samples selected from the histogram after each optimization, the predistorter parameters are iteratively updated through the direct learning structure, and the NMSE between the power amplifier output and the expected output is calculated. Each iteration runs 80 times, and the average value of the NMSE of the last 20 iterations is taken as the NMSE of this run. The average value of the NMSE obtained after 10 runs is taken again as the fitness of the current histogram. The smaller this value is, the higher the fitness, and the better the linearization performance of the predistorter calculated based on the samples selected by this histogram. Through multi-generation genetic optimization, the histogram with the highest fitness is obtained and used as the optimal histogram for the current signal type and power amplifier type. During the application of this optimal histogram, certain corrections may be required according to the actual signal distribution to make the total sampling number defined by the histogram strictly equal to the target sampling number. For example, when j the number of samples to be allocated is greater than the actual number of samples in this amplitude interval, random sampling can be performed from the adjacent j-1 to supplement the required samples. The above optimization process is pre-optimized once for a specific type of power amplifier and radiation signal and then applied to the adaptive architecture. Therefore, although the optimization calculation requires a certain amount of computational cost, its impact on the adaptive process can be ignored.

[0066] Furthermore, during the training process, different weights are assigned to the training set data, and the weights are dynamically adjusted at different training stages. Specifically, the weight of the selected samples is set to 5.0, and the weight of the unselected samples is set to 0.5, so as to pay more attention to important samples during the training process and improve the training efficiency and fitting accuracy of the model. In addition, a dynamic adjustment mechanism for weights is set: ω = (1 - λ)·ω 1 + λ·ω 2 , where λ represents the training progress ratio, calculated as the current training round / total training rounds, ω 1 represents the weight set according to sample selection, and ω 2represents uniform weights, while ω represents the weights finally applied to the training. This mechanism focuses on important samples in the early stage of training to accelerate convergence, and balances the attention to all samples in the late stage of training to improve the generalization ability.

[0067] In this embodiment, please refer to Figure 5 , Figure 5 which is a schematic diagram of a CNN-RNN hybrid network model based on deep learning provided by an embodiment of the present invention. The initial hybrid network model includes a pre-filtering layer, a temporal feature extraction layer, and a fully connected layer. The pre-filtering layer includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first batch normalization layer, a second batch normalization layer, and a third batch normalization layer; wherein, the sizes of the convolutional kernels of the first convolutional layer, the second convolutional layer, and the third convolutional layer are different; training the pre-filtering layer to obtain a trained pre-filtering layer, including:

[0068] Performing a convolution operation on part of the samples in the training dataset using the first convolutional layer to obtain first convolutional features, and processing the first convolutional features using the first batch normalization layer to obtain first features;

[0069] Performing a convolution operation on part of the samples in the training dataset using the second convolutional layer to obtain second convolutional features, and processing the second convolutional features using the second batch normalization layer to obtain second features;

[0070] Performing a convolution operation on part of the samples in the training dataset using the third convolutional layer to obtain third convolutional features, and processing the third convolutional features using the third batch normalization layer to obtain third features;

[0071] Adding the first feature, the second feature, and the third feature to obtain a fused feature;

[0072] Processing the fused feature using the temporal feature extraction layer to obtain temporal features;

[0073] Processing the temporal features using the fully connected layer to obtain the output signal of the training process;

[0074] When the number of training times or the degree of convergence meets the preset conditions, according to the output signal, determining the fitting loss of the training process, performing backpropagation on the fitting loss to determine the parameters of the first convolutional layer, the second convolutional layer, and the third convolutional layer, and determining the parameters of the first batch normalization layer, the second batch normalization layer, and the third batch normalization layer, and assigning the determined parameters to the first convolutional layer, the second convolutional layer, and the third convolutional layer, and merging them to obtain a trained convolutional layer to construct a trained pre-filtering layer.

[0075] In this embodiment, the expression of the fused feature is:

[0076] M (2) = bn(M(1) *W (5) , μ (5) , σ (5) , γ (5) , β (5) )

[0077] + bn(M (1) *W (3) , μ (3) , σ (3) , γ (3) , β (3) );

[0078] + bn(M (1) *W (1) , μ (1) , σ (1) , γ (1) , β (1) )

[0079] Among them, M (1) represents the sample in the training dataset input to the pre-filtering layer, M (2) represents the output of the pre-filtering layer, represents the weight of the first 3×3 convolutional layer with input channels C 1 , output channels C 2 , and dilation coefficient 2, represents the weight of the second 3×3 convolutional layer with input channels C 1 , output channels C 2 , represents the weight of the third 1×1 convolutional layer with input channels C 1 , output channels C 2 , μ (5) , σ (5) , γ (5) , β (5) represents the cumulative mean, standard deviation, learning ratio factor, and bias of the first batch normalization layer, μ (3) , σ (3) , γ (3) , β (3) represents the cumulative mean, standard deviation, learning ratio factor, and bias of the second batch normalization layer, μ (1) , σ (1) , γ (1) , β (1) represents the cumulative mean, standard deviation, learning ratio factor, and bias of the third batch normalization layer.

[0080] It can be understood that, please refer to Figure 6 , Figure 6It is a schematic diagram of the CNN partial network structure and the reparameterization process provided by the embodiments of the present invention. The pre-filtering layer is mainly used to extract the features contained in the data. It adopts a multi-branch structure during the training phase, which includes three branches in total, namely a dilated convolutional kernel with a size of 3*3 and a dilation coefficient of 2, a convolutional kernel with a size of 3*3, and a convolutional kernel branch with a size of 1*1. A BN (Batch Normalization) layer is connected after each of the three convolutional kernels. Among them, the convolutional kernel with a size of 3*3 and its dilated convolution are intended to extract the deep memory effect features of the power amplifier with a larger receptive field. The dilated convolution can effectively extract the local features between adjacent sampling points and the long-range features between distant sampling points. The convolutional branch with a size of 1*1 is used to avoid problems such as gradient disappearance through skip connection when the input and output dimensions of the proposed network block are inconsistent. During the training phase, the input data is respectively passed through the dilated convolutional layer with a convolutional kernel size of 3*3, the convolutional layer with a size of 3*3, and the convolutional layer with a size of 1*1. The output feature maps are added together to obtain the output data after feature fusion. The input channel numbers of the three convolutional kernels are all 1, the output channel numbers are all 3, and the strides are all 2, but the padding sizes are different. In order to make the output feature map sizes of each convolutional layer the same, the padding of the dilated convolutional layer is set to α dilation ×(k - 1) / / 2, where α dilation is the dilation coefficient, k represents the convolutional kernel size, / / represents rounding down. The 3*3 convolutional layer is the case when the dilation coefficient of the dilated convolutional layer is 1. The padding of the 1*1 convolutional layer is set to (padding of the 3*3 convolutional layer - k / / 2). Assuming the input feature map size is 5*5, the final output feature map size is 5*5.

[0081] During the inference phase, the branches are removed and integrated into a single convolutional kernel with a size of 5*5 through the reparameterization method. Regarding the structure reparameterization method, a shortcut branch similar to the residual network is constructed and modeled as y = g(x) + f(x) + h(x), where g(x) is implemented through a 1*1 convolutional kernel when the size of x does not match that of f(x), and f(x) and h(x) are respectively implemented by the convolution with a size of 3*3 and its dilated convolution. BN is used for each branch before summing up the three. Using and respectively represent the weights of the 3*3 convolution with an input channel of C 1 , an output channel of C 2 and its dilated convolution with a dilation coefficient of 2. represents the 1×1 convolution with an input channel of C 1 , an output channel of C 2 . Using μ (3) , σ (3) , γ (3) , β(3) and μ (5) , σ (5) , γ (5) , β (5) represent the cumulative mean, standard deviation, learning rate factor, and bias of the BN layer after 3*3 convolution and its dilated convolution respectively. μ (1) , σ (1) , γ (1) , β (1) are the relevant parameters of the BN layer of the identity branch. and represent the input and output of the network respectively. * represents the convolution operation. Therefore:

[0082] M (2) = bn(M (1) * W (5) , μ (5) , σ (5) , γ (5) , β (5) )

[0083] + bn(M (1) * W (3) , μ (3) , σ (3) , γ (3) , β (3) );

[0084] + bn(M (1) * W (1) , μ (1) , σ (1) , γ (1) , β (1) )

[0085] According to the definition of the BN layer:

[0086]

[0087] Therefore, each BN layer and its previous convolutional layer can be converted into a convolutional layer with a bias vector:

[0088]

[0089] bn(M * W, μ, σ, γ, β) :,i,:,: = (M * W') :,i,:,: + b' i ;

[0090] After the above transformations, a 5*5 kernel, a 3*3 kernel, and a 1*1 kernel, as well as three bias vectors, can be obtained respectively. The final bias is obtained by adding the three bias vectors. The 3*3 and 1*1 kernels are zero-padded to obtain two degenerate 5*5 kernels, and the three kernels are added together to obtain the final 5*5 kernel. The multi-branch structure can extract multi-scale features through multiple branches during the training phase, thereby improving the network's performance on complex tasks. At the same time, it can improve the gradient flow to efficiently train the network. In the inference phase, the structure of simple convolution can effectively improve the inference speed of the model.

[0091] To reduce the number of model parameters and avoid problems such as gradient vanishing or gradient explosion during the training process of traditional RNN networks, the temporal feature extraction layer adopts a structure based on the gated recurrent unit (GRU). The output form of the pre-filtering convolutional layer is (batch size, number of channels, image height, image width), which can extract the local spatial features of the data, that is, the high-dimensional features of the data at the corresponding time step. First, the dimension order is adjusted to change the output data form to (batch size, image height, number of channels, image width), and then a flattening operation is performed, and the data form is adjusted to (batch size, image height, image width × number of channels). At this time, the data form is equivalent to (batch size, number of time steps, number of input features). The features extracted on different channels are integrated into one channel to increase the number of features included in the same time step. The adjusted data is input into the GRU unit to capture the relationship of the data in the time dimension, and the last time step is intercepted as the output of the GRU unit. The number of hidden layer units of the GRU unit is set to 10, the number of hidden layers is 1, and the input size corresponds to the number of input features. For example, if the output size of the convolutional layer is 3*5*5, then the time step length is 5, and the number of input features is 5*5. Therefore, the input size is set to 25.

[0092] Finally, a fully connected layer is used to integrate the output of the temporal feature extraction layer. The number of input units of the first fully connected layer is equal to the number of hidden layer units, the number of output units is 5, and the activation function is set to the tanh function; the number of input units of the second fully connected layer is 5, and the number of output units is 2, which is used to output the I / Q components of the output data of the final prediction.

[0093] S103. Input the processed predistorted signal into a power amplifier for amplification to obtain a linearized output signal.

[0094] Specifically, in this embodiment, the trained predistortion network model is cascaded with a power amplifier to implement the power amplifier linearization function. The input data is first passed through the predistortion network, and the output predistorted signal is input into the power amplifier after operations such as upconversion to obtain a better linearity output result.

[0095] In summary, for the adaptive power amplifier linearization method based on deep learning provided by the present invention, when constructing the data set, on the basis of different characteristic delay terms, the future terms of the features are introduced, which is beneficial to the establishment of the power amplifier and predistortion models; a new predistortion network model with a CNN-RNN hybrid structure is proposed. The CNN part adopts a multi-branch structure during the training stage to optimize the training effect, and is transformed into a simple convolution structure through the reparameterization method during the inference stage to improve the inference speed. The RNN part is composed of GRU units, and the overall complexity of the model is relatively low; the sample selection method based on the histogram is used to assist the training of the neural network model. By paying more attention to the selected key samples, the training efficiency and model performance of the neural network model are improved, which is beneficial to the application of the neural network model in the adaptive predistortion technology.

[0096] Based on the same inventive concept, the present invention also provides an adaptive power amplifier linearization device based on deep learning, which is used to implement the adaptive power amplifier linearization method based on deep learning provided in the above embodiments of the present invention. For the embodiments of the method, please refer to the above, and details will not be repeated here; the device includes a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory complete communication with each other through the communication bus;

[0097] The memory is used to store computer programs;

[0098] The processor is used to implement the method of the above embodiments when executing the programs stored on the memory.

[0099] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant is intended to cover non-exclusive inclusion, so that an article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the article or device including the element. "Connection" or "connected" and other similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The orientation or positional relationship indicated by "upper", "lower", "left", "right", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be construed as a limitation of the present invention.

[0100] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification.

[0101] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. An adaptive power amplifier linearization method based on deep learning, characterized in that: include: Get the input signal to be processed; The trained hybrid network model is used to process the input signal to be processed to obtain a processed predistortion signal; wherein the trained hybrid network model uses preset category data as a training data set, selects samples in the training data set, assigns different weights to the selected samples and the unselected samples, trains the initial hybrid network model, and dynamically adjusts the weights of the samples in the training data set at different stages of the training, wherein the preset category data includes input signals, envelope-related items, leading items and time-delay items of the input signals, and leading items and time-delay items of the envelope-related items; The processed predistortion signal is input to a power amplifier for amplification to obtain a linearized output signal.

2. The method for adaptive power amplifier linearization based on deep learning according to claim 1, characterized in that: The trained hybrid network model includes a trained pre-filtering layer, a trained temporal feature extraction layer and a trained fully connected layer, and the trained pre-filtering layer includes a trained convolutional layer; The method of using the trained hybrid network model to process the input signal to be processed to obtain a predistortion signal includes: Using the trained pre-filter layer to perform a convolution operation on the input signal to be processed to obtain local spatial features; Using the trained time series feature extraction layer to extract time dimension features from the local spatial features to obtain time series features; The time series features are processed using a trained fully connected layer to obtain a pre-distorted signal.

3. The method for adaptive power amplifier linearization based on deep learning according to claim 1, characterized in that: The process of obtaining the trained hybrid network model includes: Acquire data of a plurality of preset categories; Preprocessing the data of the preset category, and using the preprocessed data of the preset category as a training data set; the training data set includes a plurality of samples, each sample including an I component and a Q component of an input signal, an envelope-related item, an advance item of the I component and the Q component of the input signal, a delay item of the I component and the Q component of the input signal, an advance item of the envelope-related item, a delay item of the envelope-related item, and real test data of the I component and the Q component of the output signal; Inputting some samples in the training data set into the hybrid network model to be trained for the pth time for training, assigning different weights to different samples, and obtaining the prediction data of the I component and the Q component of the output signal output during the pth training process; Calculate the fitting loss based on the difference between the predicted data of the I component and the Q component of the output signal output by the p-th training and the actual test data of the I component and the Q component of the output signal in the sample used in the p-th training, and use it as the fitting loss of the p-th training process; Back propagation is performed according to the fitting loss of the p-th training process to update the network parameters of the hybrid network model trained for the p-th time, and the hybrid network model to be trained for the p+1th time is obtained; this is iterated until the number of training times or the degree of convergence meets the preset conditions, and the trained hybrid network model is obtained.

4. The method for adaptive power amplifier linearization based on deep learning according to claim 3, characterized in that: The preprocessing of the preset category of data includes: Acquire input and output signals of the power amplifier as initial data of a preset category; Using a cross-correlation method, the input and output signals are aligned in time sequence; Convert the input and output signals after timing alignment into real signals of I component and Q component, and obtain the real test data of I component and Q component of the output signal; Acquire the leading term of the I component and the Q component of the input signal, the time delay term of the I component and the Q component of the input signal, the leading term of the envelope-related term, and the time delay term of the envelope-related term to construct a one-dimensional data set; The one-dimensional data set is converted into a two-dimensional data set.

5. The method for adaptive power amplifier linearization based on deep learning according to claim 3, characterized in that: The method of assigning different weights to different samples includes: Dividing upper and lower boundaries of different intervals of the histogram according to the amplitude of the input signal of the samples in the training data set to construct a histogram; A preset genetic algorithm is used to calculate the optimal number of samples in different intervals in the histogram to optimize the histogram; wherein the difference between the linearized output signal of the power amplifier and the ideal output signal is used as the value of the fitness function; According to the optimal histogram, select the samples in each interval and get the index of the selected samples; During the training process, different weights are assigned to selected samples and unselected samples, and the weights are dynamically adjusted in different iterations.

6. The method for adaptive power amplifier linearization based on deep learning according to claim 5, characterized in that: The dynamically adjusting weights in different iteration times includes: Set the mechanism for dynamically adjusting weights, the expression is: ω=(1-λ)·ω1+λ·ω2; Among them, λ represents the training progress ratio, that is, the current training times / total training times, ω1 represents the weight of the selected sample, ω2 represents the uniform weight, and ω represents the weight applied to the current training times.

7. The method for adaptive power amplifier linearization based on deep learning according to claim 3, characterized in that: The initial hybrid network model includes a pre-filtering layer, a temporal feature extraction layer and a fully connected layer, wherein the pre-filtering layer includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a first batch normalization layer, a second batch normalization layer and a third batch normalization layer; wherein the convolution kernels of the first convolutional layer, the second convolutional layer and the third convolutional layer have different sizes; and the pre-filtering layer is trained to obtain a trained pre-filtering layer, comprising: Using the first convolutional layer to perform a convolution operation on some samples in the training data set to obtain a first convolutional feature, and using the first batch normalization layer to process the first convolutional feature to obtain a first feature; Using the second convolutional layer to perform a convolution operation on some samples in the training data set to obtain a second convolutional feature, and using the second batch normalization layer to process the second convolutional feature to obtain a second feature; Using the third convolutional layer to perform a convolution operation on some samples in the training data set to obtain a third convolutional feature, and using the third batch normalization layer to process the third convolutional feature to obtain a third feature; Adding the first feature, the second feature and the third feature to obtain a fusion feature; The fusion feature is processed by the time series feature extraction layer to obtain a time series feature; Processing the time series features using the fully connected layer to obtain an output signal of a training process; When the number of training times or the degree of convergence meets the preset conditions, the fitting loss of the training process is determined according to the output signal, the fitting loss is back-propagated, the parameters of the first convolutional layer, the second convolutional layer and the third convolutional layer are determined, and the parameters of the first batch normalization layer, the second batch normalization layer and the third batch normalization layer are determined, the determined parameters are assigned to the first convolutional layer, the second convolutional layer and the third convolutional layer, and the layers are merged to obtain the trained convolutional layer to construct the trained pre-filtering layer.

8. The method for adaptive power amplifier linearization based on deep learning according to claim 7, characterized in that: The expression of the fusion feature is: Among them, M (1) represents the samples in the training data set input to the pre-filter layer, M (2) represents the output of the pre-filtering layer, It represents the weight of the first convolutional layer of 3*3 with input channel C1, output channel C2 and expansion factor 2. Represents the weight of the second convolutional layer of 3*3 with input channel C1 and output channel C2. Represents the weight of the third convolutional layer with 1*1 input channel C1 and output channel C2, μ (5) ,σ (5) ,γ (5) ,β (5) represents the cumulative mean, standard deviation, learning scale factor, and bias of the first batch normalization layer, μ (3) ,σ (3) ,γ (3) ,β (3) represents the cumulative mean, standard deviation, learning scale factor, and bias of the second batch normalization layer, μ (1) ,σ (1) ,γ (1) ,β (1) Represents the cumulative mean, standard deviation, learned scale factor, and bias of the third batch normalization layer.

9. An adaptive power amplifier linearization device based on deep learning, comprising a processor, a communication interface, a memory and a communication bus, characterized in that: The processor, the communication interface and the memory communicate with each other via the communication bus; The memory is used to store computer programs; The processor is used to implement the method according to any one of claims 1 to 8 when executing the program stored in the memory.

Images