Down-sampling DPD method and device based on two-stage attention neural network

Abstract

The invention discloses a downsampling DPD method and device based on a two-stage attention neural network. The method comprises the following steps: determining coarse estimation DPD parameters by adopting a memory polynomial method; constructing a sample set based on the plurality of input sample signals and the coarse estimation DPD signal corresponding to each input sample signal; wherein the coarse estimation DPD signal corresponding to each input sample signal is determined based on the input sample signal and the coarse estimation DPD parameter; training a DPD neural network model by using the sample set, and realizing down-sampling DPD signal estimation according to the trained model; the step of training the model by using the sample set comprises the following steps: inputting an input sample signal into the model, so that the model outputs an output DPD signal corresponding to the input sample signal; incremental training is carried out on the DPD neural network model according to the difference between the output DPD signal corresponding to the input sample signal and the coarse estimation DPD signal, the calculation cost is reduced, and the nonlinear correction efficiency of the power amplifier is improved.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a downsampling DPD (Digital Pre-distortion) method and device based on a dual-stage attention neural network. Background Art

[0002] Digital pre-distortion technology was proposed as early as the 1980s to correct the nonlinear characteristics of power amplifiers. The core idea is to artificially introduce distortion to offset the distortion of the power amplifier, thereby reducing the nonlinear effect of the power amplifier on the signal. However, since the nonlinear characteristics of the power amplifier will cause the signal bandwidth to expand, the ADC (Analog-to-Digital Converter) used for feedback in DPD usually needs to sample the output signal of the power amplifier at a rate of at least five times the signal bandwidth. This means that in broadband wireless communication systems, DPD technology not only needs to solve the problem of nonlinear distortion, but also needs to cope with the challenges of increased hardware costs and increased system energy consumption brought about by high sampling rate ADCs.

[0003] A patent application with application number CN2020107870485.6 proposes an undersampling digital predistortion method based on the Landweber iterative algorithm. The method first restores the power amplifier output signal under undersampling conditions to obtain the restored power amplifier output signal, and establishes the forward model of the power amplifier based on the restored power amplifier output signal; secondly, a DAC (Digital-to-Analog Converter) identical to the transmitting channel is added to form an auxiliary channel, and the newly added auxiliary channel and the original feedback channel are combined to form a USR (Under-Sampling Restoration) iterative hardware loop, and then the power amplifier forward model and predistorter are corrected by the iterative algorithm. This method mainly performs digital predistortion technology on the basis of undersampling technology, and then further introduces the Landweber iterative algorithm to propose a digital predistortion structure based on Landweber's USR. Finally, data closer to the real power amplifier output signal can be restored at the same sampling rate, achieving a better predistortion effect.

[0004] The undersampling digital predistortion method based on the Landweber iterative algorithm has a certain improvement in performance compared to the traditional USR digital predistortion method. However, this method still has some shortcomings, such as the need for multiple iterative calculations, high computational cost, long calculation time and low accuracy of the calculation results. Summary of the invention

[0005] To solve the above problems existing in the prior art, the present invention provides a downsampling DPD method and apparatus based on a two-stage attention neural network.

[0006] The technical problems to be solved by the present invention are achieved through the following technical solutions:

[0007] In a first aspect, the present invention provides a downsampling DPD method based on a two-stage attention neural network, and the downsampling DPD method includes:

[0008] Using the memory polynomial method to determine the coarse estimate DPD parameters;

[0009] Based on a plurality of input sample signals and the corresponding coarse estimate DPD signals of each input sample signal, constructing a sample set; wherein, the corresponding coarse estimate DPD signal of each input sample signal is determined based on the input sample signal and the coarse estimate DPD parameters;

[0010] Using the sample set to train a pre-constructed DPD neural network model, and realizing downsampling DPD signal estimation according to the trained DPD neural network model;

[0011] The step of using the sample set to train a pre-constructed DPD neural network model includes:

[0012] Inputting the input sample signal into the DPD neural network model, so that the DPD neural network model outputs the corresponding output DPD signal of the input sample signal; incrementally training the DPD neural network model according to the difference between the output DPD signal corresponding to the input sample signal and the coarse estimate DPD signal.

[0013] Optionally, using the memory polynomial method to determine the coarse estimate DPD parameters includes:

[0014] Obtaining the downsampled baseband pilot signal and the downsampled feedback signal;

[0015] Using the downsampled baseband pilot signal and the downsampled feedback signal to construct a memory polynomial model, and calculating the model coefficients of the memory polynomial model;

[0016] Obtaining the reconstructed feedback signal according to the model coefficients and the baseband pilot signal;

[0017] Using the reconstructed feedback signal and the baseband pilot signal to construct an inverse memory polynomial model;

[0018] Using the least squares method to solve the inverse memory polynomial model to obtain the coarse estimate DPD parameters.

[0019] Optionally, the corresponding coarse estimate DPD signal of each input sample signal is:

[0020]

[0021] Among them, represents the rough estimated DPD signal corresponding to the nth input sample signal; n = 0, 1,..., N; N represents the total number of input sample signals; a kq represents the rough estimated DPD parameters obtained when the non-linear order is k and the memory depth is q; k = 1, 2,..., K; K represents the maximum non-linear order; q = 0, 1,..., Q; Q represents the maximum memory depth; u(n - q) represents the signal value at the previous q moments of the nth moment.

[0022] Optionally, incrementally training the DPD neural network model according to the difference between the output DPD signal and the rough estimated DPD signal corresponding to the input sample signal includes:

[0023] Using a preset loss function to calculate the loss according to the difference between the output DPD signal and the rough estimated DPD signal corresponding to the input sample signal;

[0024] Iteratively updating the function value in the direction of reducing the loss until the loss function converges or the number of iterations reaches a preset number to complete the training.

[0025] Optionally, the loss function is:

[0026]

[0027] Among them, Loss represents the loss; n = 0, 1,..., N 0 -1; N 0 represents the total number of input sample signals used to calculate the loss function; represents the ith signal value in the output DPD signal corresponding to the nth input sample signal, where i is the index of the signal value; represents the ith signal value in the rough estimated DPD signal corresponding to the nth input sample signal.

[0028] In a second aspect, the present invention provides a downsampling DPD device based on a two-stage attention neural network. The downsampling DPD device includes:

[0029] A determination module for determining rough estimated DPD parameters by using the memory polynomial method;

[0030] A construction module for constructing a sample set based on a plurality of input sample signals and the rough estimated DPD signal corresponding to each input sample signal; among them, the rough estimated DPD signal corresponding to each input sample signal is determined based on the input sample signal and the rough estimated DPD parameters;

[0031] A training module, configured to train a pre - constructed DPD neural network model using the sample set, and estimate the down - sampled DPD signal according to the trained DPD neural network model; the training of the pre - constructed DPD neural network model using the sample set includes: inputting an input sample signal into the DPD neural network model, so that the DPD neural network model outputs an output DPD signal corresponding to the input sample signal; and incrementally training the DPD neural network model according to the difference between the output DPD signal corresponding to the input sample signal and the rough - estimated DPD signal.

[0032] Optionally, the determining module is specifically configured to:

[0033] Obtain the down - sampled base - band pilot signal and the down - sampled feedback signal; construct a memory polynomial model using the down - sampled base - band pilot signal and the down - sampled feedback signal, and calculate the model coefficients of the memory polynomial model; obtain a reconstructed feedback signal according to the model coefficients and the base - band pilot signal; construct an inverse memory polynomial model using the reconstructed feedback signal and the base - band pilot signal; and solve the inverse memory polynomial model using the least - squares method to obtain the rough - estimated DPD parameters.

[0034] Optionally, the rough - estimated DPD signal corresponding to each input sample signal is:

[0035]

[0036] where represents the rough - estimated DPD signal corresponding to the nth input sample signal; n = 0, 1, …, N; N represents the total number of input sample signals; a kq represents the rough - estimated DPD parameter obtained when the non - linear order is k and the memory depth is q; k = 1, 2, …, K; K represents the maximum non - linear order; q = 0, 1, …, Q; Q represents the maximum memory depth; u(n - q) represents the signal value at the qth moment before the nth moment.

[0037] Optionally, the training module incrementally trains the DPD neural network model according to the difference between the output DPD signal corresponding to the input sample signal and the rough - estimated DPD signal, including:

[0038] Calculating a loss according to the difference between the output DPD signal corresponding to the input sample signal and the rough - estimated DPD signal using a preset loss function;

[0039] Iteratively updating the function value in the direction of reducing the loss until the loss function converges or the number of iterations reaches a preset number to complete the training.

[0040] Optionally, the loss function is:

[0041]

[0042] where Loss represents the loss; n = 0, 1, ..., N 0 -1; N 0 represents the total number of input sample signals used to calculate the loss function; represents the i-th signal value in the output DPD signal corresponding to the n-th input sample signal, where i is the index of the signal value; represents the i-th signal value in the rough estimate DPD signal corresponding to the n-th input sample signal.

[0043] In a third aspect, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete mutual communication through the communication bus;

[0044] The memory is used to store a computer program;

[0045] The processor, when executing the computer program stored on the memory, implements the method steps of any one of the above-mentioned downsampling DPD methods based on a two-stage attention neural network.

[0046] In a fourth aspect, the present invention provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the method steps of any one of the above-mentioned downsampling DPD methods based on a two-stage attention neural network.

[0047] For the downsampling DPD method based on a two-stage attention neural network provided by the present invention, first, the memory polynomial method is used to determine the rough estimate DPD parameters. The rough estimate DPD parameters only need to be calculated once before the start of training of the DPD neural network model and do not require iteration. In addition, the DPD neural network model only needs to be trained once using the constructed sample set before the first use. When in use, the downsampling DPD signal estimation can be directly realized according to the trained DPD neural network model, overcoming the problems of large computational cost caused by multiple internal and external loop iterations in the existing downsampling DPD methods and data transmission delay caused by re-iteration every time a new signal is sent, reducing the computational cost and improving the nonlinear correction efficiency of the power amplifier.

[0048] The following will further elaborate on the present invention in conjunction with the drawings. Description of the Drawings

[0049] Figure 1It is a schematic flow chart of a downsampling DPD method based on a two-stage attention neural network provided by an embodiment of the present invention;

[0050] Figure 2 It is a schematic flow chart of determining the rough estimate DPD parameters provided by an embodiment of the present invention;

[0051] Figure 3 It is a schematic structural diagram of a DPD neural network model provided by an embodiment of the present invention;

[0052] Figure 4 It is a schematic structural diagram of a digital predistortion system applying the downsampling DPD method based on a two-stage attention neural network provided by an embodiment of the present invention;

[0053] Figure 5 It is a power spectral density diagram of the output signal of a power amplifier without DPD;

[0054] Figure 6 It is a power spectral density diagram of the output signal of a power amplifier corrected by using the downsampling DPD method provided by an embodiment of the present invention;

[0055] Figure 7 It is a schematic diagram of signal constellation comparison before and after correction;

[0056] Figure 8 It is a schematic structural diagram of a downsampling DPD device based on a two-stage attention neural network provided by an embodiment of the present invention;

[0057] Figure 9 It is a schematic structural diagram of an electronic device. Specific embodiments

[0058] The following further describes the present invention in detail with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0059] To solve the problems of large computational cost caused by the need for multiple inner and outer loop iterations in the existing downsampling DPD method and data transmission delay caused by the need to re-iterate every time a new signal is sent, an embodiment of the present invention provides a downsampling DPD method based on a two-stage attention neural network. Refer to Figure 1 , Figure 1 It is a schematic flow chart of a downsampling DPD method based on a two-stage attention neural network provided by an embodiment of the present invention, specifically including the following steps:

[0060] Step S101, determining the rough estimate DPD parameters by using the memory polynomial method.

[0061] In the embodiment of the present invention, determining the rough estimate DPD parameters by using the memory polynomial method includes:

[0062] Obtain the downsampled baseband pilot signal and the downsampled feedback signal;

[0063] Construct a memory polynomial model using the downsampled baseband pilot signal and the downsampled feedback signal, and calculate the model coefficients of the memory polynomial model;

[0064] Obtain the reconstructed feedback signal based on the model coefficients and the baseband pilot signal;

[0065] Construct an inverse memory polynomial model using the reconstructed feedback signal and the baseband pilot signal;

[0066] Solve the inverse memory polynomial model using the least squares method to obtain the rough estimated DPD parameters.

[0067] In the embodiments of the present invention, refer to Figure 2 , Figure 2 is a schematic flow chart for determining the rough estimated DPD parameters provided by the embodiments of the present invention. First, generate a 16QAM baseband original signal with a length of 1000, upsample it by 5 times, and then obtain the baseband pilot signal through a raised cosine filter with a roll-off factor of 0.25. Convert the baseband pilot signal into an analog pilot signal through a DAC (Digital-to-Analog Converter), and then input it into a broadband power amplifier after upconversion. The behavior model of the broadband power amplifier can be modeled by a memory polynomial model, and the broadband power amplifier model is expressed as:

[0068]

[0069] Among them, y(n) represents the output signal of the broadband power amplifier at the nth moment; h kq represents the complex coefficient of the memory polynomial model obtained when the non-linear order is k and the memory depth is q; k = 1, 2,..., K; K represents the maximum non-linear order; q = 0, 1,..., Q; Q represents the maximum memory depth; x(n - q) represents the signal value at the qth moment before the nth moment.

[0070] Specifically, let K = 5 and Q = 2, and the specific values are set as:

[0071]

[0072] Among them, j represents the imaginary unit; x(n) is the input signal of the broadband power amplifier at the nth moment. The output signal of the broadband power amplifier needs to be fed back to the front end through the feedback channel to implement the calculation of DPD. In actual situations, Gaussian white noise will be introduced when the signal passes through the feedback channel. The signal after passing through the feedback channel is the superposition of the output signal of the broadband power amplifier and the Gaussian white noise signal. Therefore, the finally obtained feedback signal can be expressed as:

[0073] r(n) = y(n) + w(n);

[0074] Where r(n) is the feedback signal; y(n) is the output signal of the broadband power amplifier; w(n) is the Gaussian white noise signal.

[0075] Downsample the feedback signal, for example, set the sampling rate of the ADC to 1 / 4 of the DAC rate. When the feedback signal passes through the low-rate ADC, the downsampled feedback signal r(m) will be obtained. Since the downsampling ratio is 4 times, r(m) = r(4n). Sample the baseband pilot signal at the same downsampling rate to obtain the downsampled baseband pilot signal x(m). Use the downsampled pilot baseband signal and the downsampled feedback signal to construct the memory polynomial model of the broadband power amplifier:

[0076] r = X D g;

[0077] Where r represents the M-dimensional vector composed of the downsampled feedback signal, M is the length of the downsampled feedback signal, r = [r(0), r(1), …, r(M-1)] T ; g represents the model coefficients of the memory polynomial model, that is, the power amplifier coefficients, expressed as g = [g 10 , …, g 1Q , g 30 , …, g 3Q , …, g k0 , …, g KQ T ; X D represents the M×9-dimensional matrix composed of the downsampled baseband pilot signal, defined as is the non-linear memory polynomial with non-linear order k and memory depth q at the m-th moment, x(m-q) represents the signal value at the q moments before the m-th moment, so it can be expressed as:

[0078]

[0079] Where the superscript T represents the transpose operation.

[0080] Use the least squares method to calculate the model coefficients of the memory polynomial model:

[0081]

[0082] Where the superscript H represents the conjugate transpose operation.

[0083] In the embodiments of the present invention, the reconstructed feedback signal is obtained according to the model coefficients and the baseband pilot signal:

[0084] ​

[0085] where r rec (n) represents the reconstruction feedback signal at the n-th moment, and this reconstruction feedback signal is a high-sampling-rate feedback signal; g kq represents the model coefficient of the memory polynomial model when the non-linear order is k and the memory depth is q; x(n - q) represents the baseband pilot signal without downsampling at the previous q moments of the n-th moment.

[0086] Construct an inverse memory polynomial model using the above reconstruction feedback signal and baseband pilot signal:

[0087] x = R rec a;

[0088] where x represents the N-dimensional vector composed of the baseband pilot signal without downsampling, x = [x(0), x(1), …, x(N - 1)] T ; R rec has a format similar to X D and is an N×9 matrix composed of the reconstruction feedback signal. N is the length of the pilot signal, and the is defined as is the non-linear memory polynomial at the n-th moment when the non-linear order is k and the memory depth is q; r rec (n - q) represents the signal value at the previous q moments of the n-th moment, and R rec is expressed as:

[0089]

[0090] where a is the complex coefficient vector of the inverse memory polynomial model, that is, the roughly estimated DPD parameter, a = [a 10 , …, a 1Q , a 30 , … a 3Q , … a KQ .

[0091] Use the least squares method to solve and obtain the roughly estimated DPD parameter a:

[0092]

[0093] The superscript H represents the operation of finding the transpose conjugate.

[0094] Step S102, construct a sample set based on multiple input sample signals and the roughly estimated DPD signals corresponding to each input sample signal; where the roughly estimated DPD signal corresponding to each input sample signal is determined based on this input sample signal and the roughly estimated DPD parameter.

[0095] Specifically, the sample set includes the real part, imaginary part, and modulus of each input sample signal, as well as the real part and imaginary part of the roughly estimated DPD signal corresponding to the input sample signal. Among them, the input sample signal includes the current input sample signal and the historical input sample signal. The real part, imaginary part, and modulus of each input sample signal can be obtained from the input complex sample signal, and the input complex signal value is known at the transmitting end. The calculation formula for the modulus is where x I (n) is the real part of the signal, and x Q (n) is the imaginary part of the signal. The real part and imaginary part of the roughly estimated DPD signal corresponding to each input sample signal can be determined based on the input sample signal and the roughly estimated DPD parameters:

[0096]

[0097] where represents the roughly estimated DPD signal corresponding to the input sample signal at the nth moment; n = 0, 1,..., N; N represents the total number of input sample signals; a kq represents the roughly estimated DPD parameters obtained when the non - linear order is k and the memory depth is q; k = 1, 2,..., K; K represents the maximum non - linear order; q = 0, 1,..., Q; Q represents the maximum memory depth; u(n - q) represents the signal value at the qth moment before the nth moment.

[0098] In the embodiment of the present invention, the sample set may include a training sample set and a test sample set. Let K = 5, Q = 2, and u(n) be the input sample signal of the DPD neural network model at the nth moment. Therefore, the sample set can be expressed as represents the baseband pilot signal, which is a newly generated signal during model testing, u R (n) represents the real part of u(n), and u I (n) represents the imaginary part of u(n), |u(n)| represents the modulus of u(n), and Q represents the memory depth, Q = 2. represents the roughly estimated DPD signal, expressed as where is the real part of , is the imaginary part of . When training the model, the output data set is There is no output data set during testing. The number of the training sample set is 1000, and the number of the test sample set is 100000.

[0099] Step S103: Train the pre-constructed DPD neural network model using the sample set, and implement downsampled DPD signal estimation based on the trained DPD neural network model. Training the pre-constructed DPD neural network model using the sample set includes: inputting the input sample signal into the DPD neural network model so that the DPD neural network model outputs the output DPD signal corresponding to the input sample signal; performing incremental training on the DPD neural network model according to the difference between the output DPD signal corresponding to the input sample signal and the coarsely estimated DPD signal.

[0100] In the embodiment of the present invention, a DPD neural network model including an attention mechanism module and a fully connected module is constructed. Refer to Figure 3 , Figure 3 which is a schematic structural diagram of the DPD neural network model provided by the embodiment of the present invention. Among them, the first layer of the attention mechanism module is a non-linear layer, and its activation function is tanh. The number of nodes in this layer is 9. The output of the i-th in the first layer can be expressed as:

[0101]

[0102] where w i (n) represents the relationship between the i-th input and the pre-estimated DPD output, represents the sample vector of the n-th input, {α i , β i , b i} are all parameters of the attention module and will be adjusted according to the loss function during model training. The second layer is a weight calculation layer, which is used to calculate the weight of each input neuron, reduce the importance of redundant estimated values, improve the utilization efficiency of input data, and improve the performance of calibration.

[0103] The output of the second layer can be expressed as:

[0104]

[0105] where, ξ i (n) is the weight of the i-th data in the n-th input sample signal; exp() is an exponential operation function; w i (n) is the output value of the i-th data in the n-th input sample signal of the first layer; i is the index value of the input sample signal vector. Multiply the calculated weight by the input sample signal, and the obtained weighted signal is This step is represented as Figure 3 in Since the weights of input sample signals with high correlation are larger, and the weights of input sample signals with low correlation are smaller, the weights of redundant information can be effectively reduced, and the accuracy of calibration can be improved.

[0106] In the embodiments of the present invention, considering the situation that there is noise in the feedback channel, a DPD attention mechanism neural network model module is used to reduce the weight of redundant information, overcoming the problem that the existing system can only be implemented in an ideal feedback channel. The embodiments of the present invention can be unaffected by the noise in the feedback channel and still have a good calibration effect in the case of more noise, improving the stability.

[0107] The attention mechanism module can obtain the weights of each part of the input sample signal, make full use of limited data to mine the desired information, overcome the defects of low data utilization efficiency and large amount of required data in the prior art, and enable the embodiments of the present invention to still obtain good nonlinear calibration results with less data.

[0108] The fully connected module is located behind the attention mechanism module, multiplies the weight calculated by the attention module by the input sample signal, and the obtained weighted signal χ i (n) is the input of the fully connected module. The fully connected module contains two hidden layers. To simplify the model as much as possible, the number of nodes in the first hidden layer is set to 5, and the number of nodes in the second hidden layer is set to 10. All hidden layers use the tanh activation function. The output of the fully connected layer can be expressed as:

[0109]

[0110] Figure 3 where f 1 represents tanh(C 1 χ(n)+b 1 ), and f 2 represents tanh(C 2 f 1 +b 2 )

[0111] Among them, represents the output DPD signal; C 1 , C 2 , b 1 , b 2 all represent the coefficients of the hidden layer.

[0112] Among them, C 1 is expressed as:

[0113]

[0114] Among them, b 1 is expressed as L 1 is the number of nodes in the first hidden layer.

[0115] In the embodiments of the present invention, C 2 , b 2The expression is similar, only the number of nodes is different. C 2 is expressed as:

[0116]

[0117] where b 2 is expressed as L 2 is the number of nodes in the second hidden layer. χ(n) is a vector composed of weighted input sample signals, and χ(n) is expressed as χ(n) = [χ 1 (n), χ 2 (n), …, χ 3Q+3 (n)].

[0118] In the embodiment of the present invention, the DPD neural network model is incrementally trained according to the difference between the output DPD signal corresponding to the input sample signal and the rough estimated DPD signal, including:

[0119] Using a preset loss function, calculate the loss according to the difference between the output DPD signal corresponding to the input sample signal and the rough estimated DPD signal;

[0120] Iteratively update the function value in the direction of reducing the loss until the loss function converges or the number of iterations reaches the preset number of times to complete the training.

[0121] In the embodiment of the present invention, when the model is trained, the input sample signal in the sample set is input into the DPD neural network model, and the output of the DPD neural network model is the real part and the imaginary part of the output DPD signal. The Adam optimization algorithm and the mean square error loss (MSE) are used to adjust the model parameters. The Adam optimization algorithm is an adaptive learning rate optimization algorithm used in deep learning. The loss function can be expressed as:

[0122]

[0123] where Loss represents the loss; n = 0, 1,..., N 0 -1; N 0 represents the total number of input sample signals used to calculate the loss function; is the true value, that is, the i-th signal value in the output DPD signal corresponding to the n-th input sample signal. i is the index of the signal value. When i = 1, this value is the real part of the signal, and when i = 2, this value is the imaginary part of the signal; is the predicted value, that is, the i-th signal value in the rough estimated DPD signal corresponding to the n-th input sample signal; The learning rate can be set to 0.085.

[0124] The downsampling DPD method based on the two-stage attention neural network provided by the embodiment of the present invention can be applied to the digital predistortion system. SeeFigure 4 , Figure 4 is a schematic structural diagram of a digital predistortion system applying the downsampling DPD method based on a two-stage attention neural network provided by an embodiment of the present invention. The system includes a baseband signal generation module, a baseband signal processing module, and a radio frequency module. Among them, the baseband signal generation module includes a binary signal generator and a baseband signal modulator, and the baseband signal generation module is used to generate a baseband transmission signal. Among them, the binary signal generator is used to generate a binary bit stream, and the baseband signal modulator is used to modulate the binary signal into a QAM baseband signal. When the system starts to work, a pilot signal will be generated first, and when it works officially, the baseband transmission signal to be transmitted will be generated.

[0125] The baseband signal processing module includes DPD, DAC, and a low-speed ADC. DPD includes a DPD pre-estimation module and an attention mechanism neural network model module. The DPD pre-estimation module is used to preliminarily estimate the parameters of DPD, including the reconstruction of the feedback signal, the construction of the power amplifier inverse model, and the extraction of the coarse-estimated DPD parameters. This part only needs to calculate a small amount of pilot signals and only runs once when the system starts to work.

[0126] The DPD neural network model is used to estimate the DPD signal. According to the coarsely estimated DPD parameters and the input sample signal to be transmitted, a DPD signal can be initially obtained. This signal can reflect the correlation with the transmission signal, but it is not accurate and is vulnerable to the influence of noise in the feedback channel. According to the correlation, the weights of the real part, imaginary part, and modulus of the current and historical transmission signals can be obtained. Further, the DPD behavior is learned through the DPD neural network model to obtain an accurate DPD signal. When the system starts to work, the DPD neural network model needs to be trained. When the system works officially, the transmission signal can be directly input into the model, and the output of the model is the accurately estimated DPD signal. The DAC is used to convert the digital baseband signal into an analog baseband signal, and the low-speed ADC is used to downsample the feedback signal.

[0127] The radio frequency module includes a signal upconverter, a broadband power amplifier PA, a feedback channel, an attenuator, a signal downconverter, and an anti-aliasing filter, and is used to modulate the baseband signal onto a high-frequency carrier and amplify the signal power through the power amplifier. The feedback channel is used to feedback the power amplifier output signal to the front end of the system for DPD calculation. Before DPD calculation, the signal needs to pass through an attenuator to remove the influence of the signal power amplification caused by the power amplifier, and pass through a downconverter and an anti-aliasing filter to restore the signal to a baseband signal.

[0128] When the system applying the downsampling DPD method based on the dual-stage attention neural network provided by the embodiment of the present invention is in the formal working stage, the system structure is simple and it works through the DPD neural network model, which overcomes the problem that a large number of DPD calculation modules are still required to work in the working stage of the existing system, reduces the complexity of the system, and improves the working efficiency of the system.

[0129] In the embodiment of the present invention, the rough estimation of DPD parameters only needs to be calculated once before the DPD neural network model starts training and no iteration is required. In addition, the DPD neural network model only needs to be trained once using the constructed sample set before the first use. When in use, the downsampling DPD signal estimation can be directly realized according to the trained DPD neural network model, which overcomes the problems of large computational cost caused by multiple internal and external loop iterations and data transmission delay caused by re-iteration every time a new signal is sent in the existing downsampling DPD method, reduces the computational cost, and improves the power amplifier nonlinear correction efficiency.

[0130] The simulation experiment of the downsampling DPD method based on the dual-stage attention neural network provided by the embodiment of the present invention is as follows:

[0131] Draw the power spectral density diagrams of the power amplifier output signals before and after correction. By observing the severity of the extension of the power spectral density, the severity of signal nonlinear distortion can be seen. The more severe the spectrum extension, the more severe the distortion. See Figure 5 , Figure 5 is the power spectral density diagram of the power amplifier output signal without DPD. When there is no DPD correction, the power spectral density extension phenomenon of the power amplifier output signal is obvious, and the lower the signal-to-noise ratio, the more severe the distortion phenomenon. See Figure 6 , Figure 6 is the power spectral density diagram of the power amplifier output signal corrected by using the downsampling DPD method provided by the embodiment of the present invention. It can be seen that the extension phenomenon of the power spectral density should be significantly improved, and the power spectral density of the signal is already very close to that of the original signal. And as the noise increases, the correction effect can still be maintained.

[0132] See Figure 7 , Figure 7It is a schematic diagram of the signal constellation before and after calibration. A schematic diagram of the signal constellation before and after calibration is drawn when SNR = 15 dB. SNR represents the signal-to-noise ratio. It can be seen that without calibration, the constellation diagram of the signal shows diffusion and distortion, which will affect the subsequent decoding and recognition of the signal, resulting in an increase in the bit error rate and seriously affecting the communication quality. After adopting the downsampling DPD method based on the two-stage attention neural network provided by the embodiment of the present invention, the distortion of the signal constellation diagram is significantly improved, which can effectively reduce the increase in the bit error rate caused by power amplifier distortion at the receiving end and improve the communication quality.

[0133] Based on the same inventive concept, the embodiment of the present invention also provides a downsampling DPD device based on a two-stage attention neural network. Refer to Figure 8 , Figure 8 It is a schematic structural diagram of a downsampling DPD device based on a two-stage attention neural network provided by the embodiment of the present invention. The downsampling DPD device includes:

[0134] A determination module 801, configured to determine the coarse estimate DPD parameters by using the memory polynomial method;

[0135] A construction module 802, configured to construct a sample set based on multiple input sample signals and the corresponding coarse estimate DPD signals of each input sample signal; wherein, the corresponding coarse estimate DPD signal of each input sample signal is determined based on the input sample signal and the coarse estimate DPD parameters;

[0136] A training module 803, configured to train a pre-constructed DPD neural network model by using the sample set, and implement downsampling DPD signal estimation according to the trained DPD neural network model; the training of the pre-constructed DPD neural network model by using the sample set includes: inputting the input sample signal into the DPD neural network model, so that the DPD neural network model outputs the corresponding output DPD signal of the input sample signal; incrementally training the DPD neural network model according to the difference between the output DPD signal corresponding to the input sample signal and the coarse estimate DPD signal.

[0137] In the embodiment of the present invention, the coarse estimate DPD parameters only need to be calculated once before the DPD neural network model starts training, without iteration. In addition, the DPD neural network model only needs to be trained once by using the constructed sample set before the first use, and can directly implement downsampling DPD signal estimation according to the trained DPD neural network model during use, overcoming the problems of large computational cost caused by multiple internal and external loop iterations in the existing downsampling DPD methods and data transmission delay caused by re-iteration every time a new signal is sent, reducing the computational cost and improving the efficiency of power amplifier nonlinearity correction.

[0138] Optionally, the determining module is specifically configured to:

[0139] Obtain the downsampled baseband pilot signal and the downsampled feedback signal; construct a memory polynomial model by using the downsampled baseband pilot signal and the downsampled feedback signal, and calculate the model coefficients of the memory polynomial model; obtain a reconstructed feedback signal according to the model coefficients and the baseband pilot signal; construct an inverse memory polynomial model by using the reconstructed feedback signal and the baseband pilot signal; and solve the inverse memory polynomial model by using the least squares method to obtain the rough estimated DPD parameters.

[0140] Optionally, the rough estimated DPD signal corresponding to each input sample signal is:

[0141]

[0142] where represents the rough estimated DPD signal corresponding to the nth input sample signal; n = 0, 1,..., N; N represents the total number of input sample signals; a kq represents the rough estimated DPD parameters obtained when the non-linear order is k and the memory depth is q; k = 1, 2,..., K; K represents the maximum non-linear order; q = 0, 1,..., Q; Q represents the maximum memory depth; u(n - q) represents the signal value at the qth moment before the nth moment.

[0143] Optionally, the training module incrementally trains the DPD neural network model according to the difference between the output DPD signal corresponding to the input sample signal and the rough estimated DPD signal, including:

[0144] Calculate the loss by using a preset loss function according to the difference between the output DPD signal corresponding to the input sample signal and the rough estimated DPD signal;

[0145] Iteratively update the function value in the direction of reducing the loss until the loss function converges or the number of iterations reaches a preset number of times to complete the training.

[0146] Optionally, the loss function is:

[0147]

[0148] where Loss represents the loss; n = 0, 1,..., N 0 -1; N 0 represents the total number of input sample signals used to calculate the loss function; represents the ith signal value in the output DPD signal corresponding to the nth input sample signal, and i is the index of the signal value; It represents the i-th signal value in the roughly estimated DPD signal corresponding to the n-th input sample signal.

[0149] An embodiment of the present invention also provides an electronic device, as Figure 9 shown, Figure 9 is a schematic structural diagram of an electronic device, including a processor 901, a communication interface 902, a memory 903, and a communication bus 904. Among them, the processor 901, the communication interface 902, and the memory 903 complete mutual communication through the communication bus 904.

[0150] The memory 903 is used to store computer programs.

[0151] When the processor 901 is used to execute the program stored on the memory 903, it implements the method steps of any of the above-mentioned downsampling DPD methods based on the dual-stage attention neural network.

[0152] The communication bus mentioned in the above electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, only a thick line is used to represent it in the figure, but it does not mean that there is only one bus or one type of bus.

[0153] The communication interface is used for communication between the above electronic device and other devices.

[0154] The memory may include a Random Access Memory (RAM), or may also include a Non-Volatile Memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0155] The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0156] The present invention also provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium. When the computer program is executed by a processor, the method steps of any of the above-mentioned downsampling DPD methods based on a two-stage attention neural network are implemented.

[0157] Optionally, the computer-readable storage medium may be a non-volatile memory (NVM), such as at least one disk memory.

[0158] Optionally, the above-mentioned computer-readable storage medium may also be at least one storage device located far from the aforementioned processor.

[0159] In another embodiment of the present invention, a computer program product containing instructions is also provided. When it runs on a computer, the computer is caused to execute the method steps of any of the above-mentioned downsampling DPD methods based on a two-stage attention neural network.

[0160] It should be noted that the terms "first", "second", etc. are used to distinguish similar objects and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present invention.

[0161] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. mean 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.

[0162] Although the present invention has been described in connection with various embodiments herein, however, in the process of implementing the claimed invention, those skilled in the art can understand and achieve other variations of the disclosed embodiments by viewing the accompanying drawings and the disclosure. In the description of the present invention, the term "including" does not exclude other components or steps, the term "a" or "one" does not exclude a plurality of cases, and the meaning of "a plurality" is two or more unless otherwise specifically defined. In addition, certain measures are described in different embodiments, but this does not mean that these measures cannot be combined to produce good effects.

[0163] The method provided by the embodiments of the present invention can be applied to an electronic device. Specifically, the electronic device can be: a desktop computer, a portable computer, a smart mobile terminal, a server, etc. There is no limitation here, and any electronic device that can implement the present invention belongs to the protection scope of the present invention.

[0164] For the device / electronic device / storage medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and for the relevant parts, reference can be made to the partial description of the method embodiments.

[0165] It should be noted that the device, electronic device, and storage medium of the embodiments of the present invention are respectively the device, electronic device, and storage medium applying the above-mentioned downsampling DPD method based on a two-stage attention neural network. Then all the embodiments of the above-mentioned downsampling DPD method based on a two-stage attention neural network are applicable to the device, electronic device, and storage medium, and can achieve the same or similar beneficial effects.

[0166] 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. A downsampling DPD method based on a two-stage attention neural network, characterized in that: The downsampling DPD method comprises: The memory polynomial method is used to determine the rough estimated DPD parameters; Constructing a sample set based on multiple input sample signals and a coarse estimated DPD signal corresponding to each input sample signal; wherein the coarse estimated DPD signal corresponding to each input sample signal is determined based on the input sample signal and the coarse estimated DPD parameter; Using the sample set to train a pre-built DPD neural network model, and implementing down-sampled DPD signal estimation based on the trained DPD neural network model; The method of using the sample set to train the pre-built DPD neural network model includes: The input sample signal is input into the DPD neural network model so that the DPD neural network model outputs an output DPD signal corresponding to the input sample signal; and the DPD neural network model is incrementally trained according to the difference between the output DPD signal corresponding to the input sample signal and the coarse estimated DPD signal.

2. The downsampling DPD method according to claim 1, characterized in that: The memory polynomial method is used to determine the rough estimated DPD parameters, including: Acquire a downsampled baseband pilot signal and a downsampled feedback signal; constructing a memory polynomial model using the downsampled baseband pilot signal and the downsampled feedback signal, and calculating model coefficients of the memory polynomial model; Obtaining a reconstructed feedback signal according to the model coefficients and the baseband pilot signal; constructing an inverse memory polynomial model using the reconstructed feedback signal and the baseband pilot signal; The inverse memory polynomial model is solved using the least squares method to obtain a rough estimate of the DPD parameters.

3. The downsampling DPD method according to claim 1, characterized in that: The coarse estimated DPD signal corresponding to each input sample signal is: in, represents the rough estimated DPD signal corresponding to the nth input sample signal; n=0,1,...,N; N represents the total number of input sample signals; a kq It represents the coarse estimated DPD parameters when the nonlinear order is k and the memory depth is q; k = 1, 2, ..., K; K represents the maximum nonlinear order; q = 0, 1, ..., Q; Q represents the maximum memory depth; u(nq) represents the signal value at the q moments before the nth moment.

4. The method according to claim 1, characterized in that: Incrementally training the DPD neural network model according to the difference between the output DPD signal corresponding to the input sample signal and the coarse estimated DPD signal, comprising: Calculating the loss according to the difference between the output DPD signal corresponding to the input sample signal and the coarse estimated DPD signal using a preset loss function; The function value is iteratively updated in the direction of reducing the loss until the loss function converges or the number of iterations reaches a preset number, and the training is completed.

5. The downsampling DPD method according to claim 4, characterized in that: The loss function is: Wherein, Loss represents the loss; n=0, 1, ..., N0-1; N0 represents the total number of input sample signals used to calculate the loss function; represents the i-th signal value in the output DPD signal corresponding to the n-th input sample signal, where i is the index of the signal value; Represents the i-th signal value in the coarse estimated DPD signal corresponding to the n-th input sample signal.

6. A downsampling DPD device based on a two-stage attention neural network, characterized in that: The down-sampling DPD device comprises: A determination module, used for determining a rough estimated DPD parameter by using a memory polynomial method; A construction module, configured to construct a sample set based on a plurality of input sample signals and a coarse estimated DPD signal corresponding to each input sample signal; wherein the coarse estimated DPD signal corresponding to each input sample signal is determined based on the input sample signal and the coarse estimated DPD parameter; A training module is used to train a pre-constructed DPD neural network model using the sample set, and realize down-sampled DPD signal estimation based on the trained DPD neural network model; the training of the pre-constructed DPD neural network model using the sample set includes: inputting an input sample signal into the DPD neural network model so that the DPD neural network model outputs an output DPD signal corresponding to the input sample signal; and incrementally training the DPD neural network model according to the difference between the output DPD signal corresponding to the input sample signal and the coarse estimated DPD signal.

7. The down-sampling DPD device according to claim 6, characterized in that: The determination module is specifically used for: A downsampled baseband pilot signal and a downsampled feedback signal are obtained; a memory polynomial model is constructed using the downsampled baseband pilot signal and the downsampled feedback signal, and model coefficients of the memory polynomial model are calculated; a reconstructed feedback signal is obtained according to the model coefficients and the baseband pilot signal; an inverse memory polynomial model is constructed using the reconstructed feedback signal and the baseband pilot signal; and the inverse memory polynomial model is solved using a least squares method to obtain a coarse estimated DPD parameter.

8. The down-sampling DPD device according to claim 6, characterized in that: The coarse estimated DPD signal corresponding to each input sample signal is: in, represents the rough estimated DPD signal corresponding to the nth input sample signal; n=0,1,...,N; N represents the total number of input sample signals; a kq It represents the coarse estimated DPD parameters when the nonlinear order is k and the memory depth is q; k = 1, 2, ..., K; K represents the maximum nonlinear order; q = 0, 1, ..., Q; Q represents the maximum memory depth; u(nq) represents the signal value at the q moments before the nth moment.

9. An electronic device, characterized in that: It includes a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, used to implement the downsampling DPD method according to any one of claims 1 to 5 when executing a computer program stored in a memory.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the downsampling DPD method according to any one of claims 1 to 5 is implemented.

Images