Broadband digital pre-distortion method and system based on external attention mechanism

Abstract

The invention discloses a broadband digital pre-distortion method and system based on an external attention mechanism, and the method comprises the steps: constructing an external attention intensive gated loop network model which comprises a feature extraction layer, an input layer, a GRU layer, a hidden full-connection layer, an external attention layer and an output full-connection layer; based on the baseband IQ signal after feature extraction and preprocessing and an IQ signal output by a power amplifier, constructing input data of an external attention intensive gated loop network model; training an external attention intensive gating loop network model; and inputting data to the trained external attention intensive gated loop network model to obtain a predicted value output by the network model to the power amplifier. The invention provides a novel broadband digital pre-distortion architecture of an external attention intensive gated loop network. An external attention mechanism is combined with an intensive connection layer, so that the network better compensates the nonlinear characteristic of a power amplifier.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication, and particularly relates to a broadband digital predistortion method and system based on an external attention mechanism. Background Art

[0002] In the field of wireless communication, with the limited nature of communication spectrum resources and the continuous increase in the demand for data transmission rate, the linearization technology of power amplifiers has become one of the research hotspots. To ensure the transmission quality of signals, power amplifiers usually need to work at high power efficiency. However, the non-linear characteristics of power amplifiers will introduce signal distortion, thereby affecting the performance of communication systems. Digital Predistortion (DPD) technology is one of the main means to solve this problem currently. The core idea of digital predistortion technology is to pre-distort the signal in the digital domain before the signal enters the power amplifier, so that after the signal passes through the power amplifier, the non-linear distortion effect is cancelled, thereby achieving a linearized output.

[0003] Existing digital predistortion technologies, such as predistortion methods based on polynomial or memory polynomial models, although they can effectively compensate for the non-linear distortion of power amplifiers, have insufficient compensation accuracy when facing complex non-linearities and memory effects, especially in the high-frequency band and for broadband signals; and with the increase in signal complexity in modern communication systems, the number of parameters of traditional models increases, resulting in a significant increase in computational complexity and making it difficult to run in real time in practical applications. Especially in scenarios that require high-precision compensation, complex calculations will further affect the efficiency and delay performance of the system. Secondly, existing predistortion models have limited generalization ability and poor applicability when facing different types of power amplifiers and different working conditions. In addition, traditional digital predistortion methods respond slowly when facing dynamically changing signals and cannot quickly adapt to the non-linear changes of power amplifiers. Especially in a mobile communication environment, this problem is particularly prominent. Summary of the Invention

[0004] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, and provide a broadband digital predistortion method and system based on an external attention mechanism, and propose a broadband digital predistortion architecture of a novel external attention dense gated recurrent network, which enables the network to better compensate for the non-linear characteristics of power amplifiers through the combination of an external attention mechanism and a dense connection layer.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A broadband digital predistortion method based on an external attention mechanism, comprising the following steps:

[0007] Construct an external attention gated recurrent network model, including a feature extraction layer, an input layer, a GRU layer, a hidden fully connected layer, an external attention layer, and an output fully connected layer;

[0008] Based on the baseband IQ signal after feature extraction and preprocessing and the IQ signal output by the power amplifier, construct the input data of the external attention gated recurrent network model;

[0009] Train the external attention gated recurrent network model;

[0010] Input the data into the trained external attention gated recurrent network model to obtain the predicted value of the power amplifier output by the network model.

[0011] The present invention also includes a broadband digital predistortion system based on an external attention mechanism. The system adopts the broadband digital predistortion method provided by the present invention, and the system includes a signal preprocessing module and an external attention gated recurrent network module;

[0012] The signal preprocessing module is used to perform feature extraction and preprocessing on the input signal, and based on the baseband IQ signal after feature extraction and preprocessing and the IQ signal output by the power amplifier, construct and output the input data of the external attention gated recurrent network module;

[0013] The external attention gated recurrent network module is used to construct and train an external attention gated recurrent network model, process the output of the signal preprocessing module, and obtain the predicted value of the power amplifier output.

[0014] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0015] 1. The present invention constructs a neural network model that combines a gated recurrent unit (GRU) with an external attention mechanism; the GRU can capture the long-term dependence of signals and effectively model complex memory effects, while the external attention mechanism can focus on the key features in the input signal and consider the potential dependence between different samples. Through this structure, the present invention has higher compensation accuracy in the high-frequency band and broadband signals, overcoming the defect of insufficient compensation accuracy of the prior art in a complex non-linear environment.

[0016] 2. After introducing the external attention mechanism, the neural network structure of the present invention can reduce the redundant processing of high-dimensional data and effectively control the computational complexity of the model. Compared with traditional predistortion models that require a large number of parameters for complex calculations, the network of the present invention reduces the computational overhead through weight sharing and feature extraction, thereby improving the real-time performance of the system and enabling the model to adapt to the high data rate and complex signal environment in modern communication systems.

[0017] 3. The combination of GRU and the external attention mechanism adopted by the present invention enhances the generalization ability of the model. During the training process of the predistortion model, the neural network can learn the general laws of the nonlinear distortion of the power amplifier, so that it can still maintain a good compensation effect in different amplifier environments, solving the defect of poor generalization ability of the existing technology.

[0018] 4. Due to the efficient modeling and reasoning capabilities of the GRU layer and the external attention mechanism in the present invention, the pre-trained model can perform rapid compensation without the need for retraining every time there is a new input or environmental change. In this way, the model can maintain high-efficiency reasoning performance in the face of changing signals, ensuring real-time performance in complex wireless communication environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flowchart of the method of the present invention;

[0020] Figure 2 is a schematic structural diagram of the external attention dense gated recurrent network model;

[0021] Figure 3 is the training process of the model;

[0022] Figure 4 is a comparison diagram of TEST_LOSS and VAL_LOSS of different models;

[0023] Figure 5 is a schematic comparison diagram of the number of samples processed per second by three models in the embodiment;

[0024] Figure 6 is a schematic comparison diagram of NMSE, ACPR, and EVM of the test set of the model in this embodiment and the prior art model under 100 training iterations;

[0025] Figure 7 is the power spectral density diagram of the external attention dense gated recurrent network model;

[0026] Figure 8 is the 64-QAM constellation diagram of the external attention dense gated recurrent network model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] The present invention will be further described in detail below in conjunction with the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0028] Description of related terms:

[0029] DPD: Digital Predistortion, digital predistortion.

[0030] GRU: Gated Recurrent Unit.

[0031] LSTM: Long Short-Term Memory network.

[0032] VDLSTM: Vector Decomposed Long Short-Term Memory network.

[0033] RVTDCNN: Real-Valued Tapped Delay Convolutional Network.

[0034] GMP: Generalized Memory Polynomial.

[0035] DGRU: Dense Gated Recurrent Unit.

[0036] EA-DGRU, DGRUEXTERNAL: External Attention Dense Gated Recurrent Unit.

[0037] PA: Power Amplifier, power amplifier.

[0038] NMSE: Normalized Mean Square Error.

[0039] EVM: Error Vector Magnitude.

[0040] ACLR: Adjacent Channel Leakage Ratio.

[0041] ACPR: Adjacent Channel Power Ratio.

[0042] VAL_LOSS: Loss on the validation set.

[0043] VAL_NMSE: Normalized Mean Square Error on the validation set.

[0044] VAL_EVM: Error Vector Magnitude on the validation set.

[0045] VAL_ACLR_AVG: Average Adjacent Channel Leakage Ratio on the validation set.

[0046] TEST_LOSS: Loss on the test set.

[0047] TEST_NMSE: Normalized Mean Square Error on the test set.

[0048] TEST_EVM: Error Vector Magnitude on the test set.

[0049] TEST_ACLR_AVG: Average Adjacent Channel Leakage Ratio on the test set.

[0050] Embodiment

[0051] As Figure 1 and Figure 2 shown, the present invention, a broadband digital predistortion method based on an external attention mechanism, includes the following steps:

[0052] S1. Construct an external attention dense gated recurrent network model, including a feature extraction layer, an input layer, a GRU layer, a hidden fully connected layer, an external attention layer, and an output fully connected layer;

[0053] S2. Based on the baseband IQ signal x(n) after feature extraction and preprocessing and the IQ signal y(n) output by the power amplifier, construct the input data Xin(n) and Yout(n) of the external attention dense gated recurrent network model. Let Iin(n) and Qin(n) represent the real and imaginary parts of x(n), and Iout(n) and Qout(n) represent the real and imaginary parts of y(n).

[0054] After constructing the input data of the external attention dense gated recurrent network model, divide the input data into a training set, a validation set, and a test set according to a ratio of 6:2:2.

[0055] S3. Train the external attention dense gated recurrent network model;

[0056] S4. Input data into the trained external attention dense gated recurrent network model to obtain the predicted value of the output of the power amplifier by the network model.

[0057] In this embodiment, as Figure 2 shown, after inputting data into the external attention dense gated recurrent network model, first the feature extraction layer extracts the feature values of the input data, and extracts the input data into the real part Iin(n), the imaginary part Qin(n), the signal amplitude A(n), the cube of the signal amplitude A 3 (n), the cosine component cos(θ), and the sine component sin(θ).

[0058] The feature values processed by the feature extraction layer are input into the GRU layer for processing in a size of 64*50*6. The number of hidden layers of this layer is 8; the GRU layer controls the flow of data through an update gate and a reset gate, specifically including:

[0059] Update gate. The GRU layer first calculates the hidden state h t-1 at the previous time step and the current input x t , and controls the fusion of the current state and the past state through the update gate to ensure that long-term dependencies are maintained; the update gate formula is:

[0060] Z t =σ(W z ·[h t-1, x t )

[0061] where x t is the current input, h t-1 is the hidden state at the previous time step, W z is the weight matrix, σ is the activation function, and z t is the update gate value;

[0062] The reset gate is used to determine whether to reset the hidden state, that is, whether to ignore the previous hidden information when calculating the hidden state at the current time step. The formula is:

[0063] r t = σ(W r · [h t-1 , x t )

[0064] where r t is the reset gate value;

[0065] The candidate hidden state is calculated based on the reset hidden state and the current input signal to obtain the candidate hidden state at the current time step. The formula is:

[0066]

[0067] Hidden state update: Through the update gate, the hidden state at the previous time step and the candidate hidden state are weighted and averaged to output the current hidden state h t . The formula is:

[0068]

[0069] After being processed by the GRU layer, the hidden state H is output and passed to the external attention layer.

[0070] The external attention layer is based on the external attention mechanism. The external attention mechanism works independently of individual samples through two linear layers with shared parameters and can capture the potential connections between different samples in the dataset. The specific process is as follows:

[0071] The input signal is linearly transformed to generate the query vector Q, and then the similarity is calculated with the key matrix M k stored externally to generate the attention weight matrix A:

[0072]

[0073] where Q ∈ R n×d represents the feature matrix of the input signal, and M k ∈ R s×dis an external key memory matrix, s is the number of memory units, and d is the feature dimension; Norm represents the normalization operation to ensure the numerical stability of the output attention weight matrix;

[0074] Using the generated attention weight matrix A, combined with the external value matrix M v , generate the updated feature representation F out :

[0075] F out = AM v

[0076] where M v ∈R s×d is the external value memory matrix;

[0077] Through this process, the external attention mechanism dynamically adjusts the important features in the input signal and dynamically adjusts the contributions of different features, thereby optimizing the global dependence of the feature representation.

[0078] After passing through the external attention layer, the features output by the external attention layer are non-linearly mapped through the ReLU activation function. The definition of the ReLU activation function is:

[0079] f(x) = max(0, x)

[0080] Finally, the output of the feature extraction layer is connected in parallel with the features after passing through the external attention layer and the ReLU activation function to the input end of the output fully connected layer. The calculation formula of the output fully connected layer is:

[0081] y = W fc ·H′ + b fc

[0082] where W fc is the weight matrix of the output fully connected layer, b fc is the bias term, and y is the predicted value of the external attention dense gated recurrent network model for the output of the power amplifier.

[0083] In this embodiment, as Figure 3 shown, training the external attention dense gated recurrent network model specifically includes:

[0084] Use the 200MHz dataset to perform behavioral modeling on the PA. Its training objective is to minimize L PA , where the 200MHz dataset uses 200MHz, 10 channels × 20MHz, I / Q modulated orthogonal frequency division multiplexing signals, and each channel uses 64-QAM and 64 subcarriers; L PA The calculation formula is as follows:

[0085]

[0086] Among them, B is the batch size in mini-batch stochastic gradient descent, and y′[t] is the target output. is the predicted output of the PA model. To solve the vanishing gradient problem and strengthen training, the feature and label sequences are segmented into shorter frames, each of size T and with a step size of S.

[0087] The trained PA model is cascaded with the DPD model, and let L CAS be minimized, and its formula is:

[0088]

[0089] Among them, is the predicted output of the cascaded model, x[t] is the actual input, and G is the target gain of the cascaded DPD-PA system; the PA model selects the model with the minimum NMSE, and the DPD model selects the model with the minimum ACPR. The formulas for NMSE, ACPR, and EVM are as follows:

[0090]

[0091] Among them, I y [n] and Q y [n] are the in-phase and quadrature signal components derived from the FFT. and represent the output of the analog PA.

[0092]

[0093] Among them, P main and P adj are the main and adjacent channel powers respectively.

[0094]

[0095] Among them, is the frequency-domain representation of the predicted signal. X[f] is the frequency-domain representation of the target signal, and X[f] = FFT(x); N is the number of sub-channels, indicating that the main channel is divided into N equal-width intervals. represents the summation of the frequency points of the c-th sub-channel.

[0096] For this embodiment, in actual implementation, the GRU layer of the external attention dense gated recurrent network model can be replaced by an LSTM or other recurrent neural network structures.

[0097] The generalization verification of the external attention dense gated recurrent network model of this embodiment is shown in Table 1 below, which shows the test performance indicators of the external attention dense gated recurrent network model of this embodiment and other existing models on the validation set and the test set.

[0098]

[0099] Table 1

[0100] As can be seen from Table 1, the external attention dense gated recurrent network model performs best in terms of loss and all other metrics on the validation set and the test set, and its test performance is close to its validation performance, showing stable performance. Compared with the existing traditional GMP model, all metrics have been significantly improved.

[0101] As Figure 4 shown, it is a comparison graph of TEST_LOSS and VAL_LOSS for different models.

[0102] Perform real-time verification on the external attention dense gated recurrent network model of this embodiment:

[0103] Add timing to the code running and process 7,680 test set samples.

[0104] The total time spent by the DGRUEXTERNAL model (i.e., the external attention dense gated recurrent network model) is 0.134456 s, and it processes 57,122.33 samples per second on average; the total time spent by the RVTDCNN model is 0.322712 s, and it processes 23,798.30 samples per second on average; the total time spent by the GMP model is 1.956374 s, and it processes 3,925.63 samples per second on average. As Figure 5 shown, it is a comparison schematic diagram of the number of samples processed per second by the three models. It can be seen that the real-time performance of DGRUEXTERNAL is significantly higher than that of the GMP and RVTDCNN models.

[0105] As Figure 6 shown, it is a comparison schematic diagram of NMSE, ACPR, and EVM of the test set between the model of this embodiment and the existing technology model under 100 training epochs. As shown in Table 2 below, it is the performance of each model on the test set.

[0106]

[0107] Table 2

[0108] As Figure 7 shown, it is the power spectral density graph of the external attention dense gated recurrent network model; as Figure 8 shown, it is the 64-QAM constellation graph of the external attention dense gated recurrent network model.

[0109] In another embodiment, a broadband digital predistortion system based on an external attention mechanism is further provided. The system adopts the broadband digital predistortion method of the above embodiment. The system includes a signal preprocessing module and an external attention dense gated recurrent network module;

[0110] The signal preprocessing module is used to extract and preprocess the features of the input signal, and based on the baseband IQ signal after feature extraction and preprocessing and the IQ signal output by the power amplifier, construct and output the input data of the external attention dense gated recurrent network module;

[0111] The external attention dense gated recurrent network module is used to construct and train an external attention dense gated recurrent network model, process the output of the signal preprocessing module, and obtain a predicted value of the output of the power amplifier.

[0112] It should also be noted that in this specification, terms such as "including", "comprising" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or device including the said element.

[0113] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A broadband digital predistortion method based on an external attention mechanism, characterized in that It includes the following steps: Construct an external attention dense gated recurrent network model, including a feature extraction layer, an input layer, a GRU layer, a hidden fully connected layer, an external attention layer, and an output fully connected layer; Based on the baseband IQ signal after feature extraction and preprocessing and the IQ signal output by the power amplifier, construct the input data of the external attention dense gated recurrent network model; Train the external attention dense gated recurrent network model; Input data into the trained external attention dense gated recurrent network model to obtain the predicted value of the power amplifier output by the network model.

2. The broadband digital predistortion method based on an external attention mechanism according to claim 1, characterized in that Construct the input data of the external attention dense gated recurrent network model, specifically: Based on the baseband IQ signal x(n) after feature extraction and preprocessing and the IQ signal y(n) output by the power amplifier, construct the input data Xin(n) and Yout(n) of the external attention dense gated recurrent network model. Let Iin(n) and Qin(n) represent the real and imaginary parts of x(n), and Iout(n) and Qout(n) represent the real and imaginary parts of y(n).

3. The broadband digital predistortion method based on an external attention mechanism according to claim 2, wherein After constructing the input data of the external attention dense gated recurrent network model, it also includes dividing the input data into a training set, a validation set, and a test set according to a ratio of 6:2:

2.

4. The broadband digital predistortion method based on an external attention mechanism according to claim 1, characterized in that Input the data into the external attention gated recurrent network model. The feature extraction layer extracts the eigenvalue of the input data and extracts the input data into the real part Iin(n), the imaginary part Qin(n), the signal amplitude A(n), the cube of the signal amplitude A 3 (n), the cosine component cos(θ), and the sine component sin(θ).

5. The broadband digital predistortion method based on an external attention mechanism according to claim 4, wherein The feature values processed by the feature extraction layer are input into the GRU layer for processing in a size of 64*50*6. The number of hidden layers of this layer is 8; the GRU layer controls the flow of data through an update gate and a reset gate, specifically including: Update gate, the GRU layer first calculates the hidden state h of the previous time step t-1 and the current input x t to calculate, and control the fusion of the current state and the past state through the update gate to ensure that long-term dependencies are maintained; the update gate formula is: z t = σ(W z · [h t-1 , x t ) where x t is the current input, h t-1 is the hidden state at the previous moment, W z is the weight matrix, σ is the activation function, and z t is the update gate value; Reset gate, used to determine whether to reset the hidden state, that is, whether to ignore the previous hidden information when calculating the hidden state at the current time step. The formula is: r t = σ(W r · [h t-1 , x t ) where r t is the reset threshold value; Candidate hidden state, based on the reset hidden state and the current input signal, calculate the candidate hidden state at the current moment. The formula is: Hidden state update, through the update gate, weighted average the hidden state of the previous moment and the candidate hidden state, and output the current hidden state h t , the formula is: After being processed by the GRU layer, the hidden state H is output and passed to the external attention layer.

6. The broadband digital predistortion method based on an external attention mechanism according to claim 5, wherein The external attention layer is based on the external attention mechanism. The external attention mechanism works independently of individual samples through two linear layers with shared parameters and can capture the potential connections between different samples in the dataset. The specific process is: The input signal undergoes a linear transformation to generate a query vector Q, which is then compared with the key matrix M stored externally k for similarity calculation to generate an attention weight matrix A: where \(Q\in\mathbb{R}\) n×d represents the feature matrix of the input signal, \(M\) k \(\in\mathbb{R}\) s×d is the external key memory matrix, \(s\) is the number of memory units, \(d\) is the feature dimension; Norm represents the normalization operation to ensure the numerical stability of the output attention weight matrix; Using the generated attention weight matrix A and combining it with the external value matrix M v to generate the updated feature representation F out : F out = AM v where M v ∈R s×d is an external value memory matrix; Through this process, the external attention mechanism realizes dynamic adjustment of the important features in the input signal, dynamically adjusts the contributions of different features, and thus optimizes the global dependence of the feature representation.

7. The broadband digital predistortion method based on an external attention mechanism according to claim 6, characterized in that After passing through the external attention layer, the features output by the external attention layer are non-linearly mapped through the ReLU activation function. The definition of the ReLU activation function is: f(x) = max(0, x) Finally, the output of the feature extraction layer is connected in parallel with the features after passing through the external attention layer and the ReLU activation function to the input end of the output fully connected layer. The calculation formula of the output fully connected layer is: y = W fc ·H′ + b fc Among them, W fc is the weight matrix of the output fully connected layer, and b fc is the bias term, and y is the predicted value of the output of the power amplifier by the external attention dense gated recurrent network model.

8. The broadband digital predistortion method based on an external attention mechanism according to claim 1, characterized in that Train the external attention dense gated recurrent network model, specifically including: The behavior of the PA is modeled using a 200 MHz data set, and the training objective is to minimize L PA where the 200 MHz data set uses 200 MHz, 10 channels × 20 MHz, I / Q modulated orthogonal frequency division multiplexing signals, and each channel uses 64-QAM and 64 subcarriers; L PA The calculation formula is as follows: where B is the batch size in mini-batch stochastic gradient descent, and y′[t] is the target output, is the predicted output of the PA model. To solve the vanishing gradient problem and enhance training, the feature and label sequences are segmented into shorter frames, each of size T with a stride of S; Cascade the trained PA model with the DPD model and minimize L CAS The formula is as follows: Among them, is the predicted output of the cascaded model, x[t] is the actual input, and G is the target gain of the cascaded DPD-PA system; the PA model selects the model with the minimum NMSE, and the DPD model selects the model with the minimum ACPR. The formulas for NMSE, ACPR, and EVM are as follows: Among them, I y [n] and Q y [n] are in-phase and quadrature signal components derived from FFT, and represent the output of the analog PA; Among them, P main and P adj are the main and adjacent channel powers respectively; Among them, is the frequency-domain representation of the prediction signal, X[f] is the frequency-domain representation of the target signal, X[f] = FFT(x); N is the number of sub-channels, indicating that the main channel is divided into N equally wide intervals; denotes the summation of the frequency points of the c-th sub-channel.

9. The broadband digital predistortion method based on an external attention mechanism according to claim 1, wherein The GRU layer of the external attention dense gated recurrent network model can be replaced by an LSTM or other recurrent neural network structures.

10. A broadband digital predistortion system based on an external attention mechanism, characterized in that, The system adopts the broadband digital predistortion method described in any one of claims 1-9. The system includes a signal preprocessing module and an external attention dense gated recurrent network module; A signal preprocessing module, which is used to extract features and preprocess the input signal, and based on the baseband IQ signal after feature extraction and preprocessing and the IQ signal output by the power amplifier, construct and output the input data of the external attention dense gated recurrent network module; An external attention dense gated recurrent network module, which is used to construct and train an external attention dense gated recurrent network model, process the output of the signal preprocessing module, and obtain the predicted value of the power amplifier output.

Images