Down-sampling dpd method and apparatus based on two-stage attention neural network

Abstract

The application discloses a kind of based on two-stage attention neural network's down-sampling DPD method and device, method includes: using memory polynomial method determines rough estimate DPD parameter;Based on a plurality of input sample signals and each input sample signal corresponding rough estimate DPD signal constructs sample set;Wherein, each input sample signal corresponding rough estimate DPD signal is determined based on the input sample signal and rough estimate DPD parameter;Utilize sample set training DPD neural network model, and according to the model after training realizes down-sampling DPD signal estimation;Utilize sample set training model includes: input input sample signal into model, to make model output the input sample signal corresponding output DPD signal;According to the difference between the output DPD signal and rough estimate DPD signal corresponding to input sample signal, incremental training is carried out to DPD neural network model, reduce the calculation cost, improve the power amplifier nonlinear correction efficiency.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of communication, and particularly relates to a down-sampling DPD (Digital Pre-distortion) method and device based on a two-stage attention neural network. BACKGROUND

[0002] The digital pre-distortion technology was proposed as early as the 1980s to correct the nonlinear characteristics of a power amplifier. The core idea is to artificially introduce distortion to offset the distortion of the power amplifier, thereby reducing the nonlinear influence of the power amplifier on the signal. However, since the nonlinear characteristics of the power amplifier can cause signal bandwidth expansion, the ADC (Analog-to-Digital Converter) used for feedback in the 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 a wideband wireless communication system, the DPD technology not only needs to solve the problem of nonlinear distortion, but also needs to cope with the challenges of increased hardware cost and rising system energy consumption brought by high sampling rate ADC.

[0003] The patent application with the application number CN2020107870485.6 proposes an under-sampling digital pre-distortion method based on the Landweber iterative algorithm. The method first restores the output signal of the power amplifier under the condition of under-sampling to obtain the restored output signal of the power amplifier, and establishes a forward model of the power amplifier based on the restored output signal of the power amplifier. Secondly, an auxiliary channel composed of the same DAC (Digital-to-Analog Converter) as the transmission channel is added, and the new auxiliary channel and the original feedback channel jointly form a hardware loop of USR (Under-Sampling Restoration) iteration, and then the power amplifier forward model and the pre-distortion device are corrected through the iterative algorithm. This method mainly performs digital pre-distortion technology on the basis of under-sampling technology, and then further introduces the Landweber iterative algorithm to propose a digital pre-distortion structure based on the USR of Landweber. Finally, it can restore data closer to the real output signal of the power amplifier at the same sampling rate, achieving better pre-distortion effect.

[0004] The under-sampling digital pre-distortion method based on the Landweber iterative algorithm has certain performance improvement compared with the traditional USR digital pre-distortion method. However, this method still has some deficiencies, such as the need for multiple iterations, high calculation cost, long calculation time and low accuracy of calculation results. SUMMARY

[0005] In order to solve the above problems existing in the prior art, the application provides a down-sampling DPD method and device based on a two-stage attention neural network.

[0006] The technical problem to be solved by the application is solved by the following technical scheme:

[0007] In a first aspect, the application provides a down-sampling DPD method based on a two-stage attention neural network, which comprises the following steps.

[0008] The rough estimation DPD parameter is determined by using a memory polynomial method;

[0009] A sample set is constructed based on a plurality of input sample signals and a rough estimation DPD signal corresponding to each input sample signal, wherein the rough estimation DPD signal corresponding to each input sample signal is determined based on the input sample signal and the rough estimation DPD parameter;

[0010] The pre-constructed DPD neural network model is trained by using the sample set, and the down-sampling DPD signal estimation is realized according to the trained DPD neural network model;

[0011] The pre-constructed DPD neural network model is trained by using the sample set, and the down-sampling DPD signal estimation is realized according to the trained DPD neural network model;

[0012] 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 rough estimation DPD signal.

[0013] Optionally, the rough estimation DPD parameter is determined by using a memory polynomial method, which comprises the following steps.

[0014] The down-sampling baseband pilot signal and the down-sampling feedback signal are obtained;

[0015] A memory polynomial model is constructed by using the down-sampling baseband pilot signal and the down-sampling feedback signal, and model coefficients of the memory polynomial model are calculated;

[0016] The reconstructed feedback signal is obtained according to the model coefficients and the baseband pilot signal;

[0017] An inverse memory polynomial model is constructed by using the reconstructed feedback signal and the baseband pilot signal;

[0018] The inverse memory polynomial model is solved by using a least square method, and the rough estimation DPD parameter is obtained.

[0019] Optionally, the rough estimation DPD signal corresponding to each input sample signal is as follows:

[0020]

[0021] wherein, denotes the coarse estimation DPD signal corresponding to the nth input sample signal; n=0,1,...,N; N denotes the total number of input sample signals; a kq denotes the coarse estimation DPD parameter obtained when the non-linear order is k and the memory depth is q; k=1,2,...,K; K denotes the maximum non-linear order; q=0,1,...,Q; Q denotes the maximum memory depth; u(n-q) denotes the signal value of the previous q time points at the nth time point.

[0022] Optionally, 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 estimation DPD signal, comprising:

[0023] a loss is calculated according to the difference between the output DPD signal corresponding to the input sample signal and the coarse estimation DPD signal by using a preset loss function;

[0024] the function value is iteratively updated in the direction of reducing the loss until the loss function converges or the iteration number reaches a preset number of times, and the training is completed.

[0025] Optionally, the loss function is:

[0026]

[0027] wherein, Loss denotes the loss; n=0,1,...,N0-1; N0 denotes the total number of input sample signals used for calculating the loss function; denotes 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; denotes the ith signal value in the coarse estimation DPD signal corresponding to the nth input sample signal.

[0028] In a second aspect, the present application provides a down-sampling DPD device based on a two-stage attention neural network, comprising:

[0029] A determination module is configured to determine coarse estimation DPD parameters by using a memory polynomial method;

[0030] A construction module is configured to construct a sample set based on a 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 parameters;

[0031] The training module is configured to train a pre-constructed DPD neural network model by using the sample set, and to implement down-sampling 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 comprises: 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 performing incremental training on the DPD neural network model according to a difference between the output DPD signal corresponding to the input sample signal and a coarsely estimated DPD signal.

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

[0033] The down-sampled baseband pilot signal and the down-sampled feedback signal are obtained; a memory polynomial model is constructed by using the down-sampled baseband pilot signal and the down-sampled 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 by using the reconstructed feedback signal and the baseband pilot signal; and the inverse memory polynomial model is solved by using a least square method, so that coarsely estimated DPD parameters are obtained.

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

[0035]

[0036] wherein, indicates the coarsely estimated DPD signal corresponding to the nth input sample signal; n = 0, 1, …, N; N indicates the total number of input sample signals; a kq indicates the coarsely estimated DPD parameter obtained when the non-linear order is k and the memory depth is q; k = 1, 2, …, K; K indicates the maximum non-linear order; q = 0, 1, …, Q; Q indicates the maximum memory depth; u(n-q) indicates the signal value of the previous q time points at the nth time point.

[0037] Optionally, the training module performs 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, and the incremental training comprises:

[0038] a loss is calculated according to the difference between the output DPD signal corresponding to the input sample signal and the coarsely estimated DPD signal by using a preset loss function;

[0039] a function value is iteratively updated in a direction of reducing the loss until the loss function converges or the number of iterations reaches a preset number of times, and the training is completed.

[0040] Optionally, the loss function is:

[0041]

[0042] wherein, Loss represents the loss; n = 0, 1, …, N0-1; N0 represents the total number of input sample signals for calculating the loss function; represents 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; represents the i-th signal value in the rough estimation DPD signal corresponding to the n-th input sample signal.

[0043] In a third aspect, the present application provides an electronic device, comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory complete the communication among each other through the communication bus;

[0044] The memory is used for storing a computer program.

[0045] The processor is used for executing the computer program stored on the memory, and realizes the method steps of any one of the above-mentioned DPD methods based on the two-stage attention neural network.

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

[0047] The DPD method based on the two-stage attention neural network provided by the present application firstly determines the rough estimation DPD parameter by using the memory polynomial method, and the rough estimation DPD parameter only needs 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 at the beginning using the constructed sample set, and the trained DPD neural network model can be directly used to estimate the down-sampling DPD signal, which overcomes the problems of large calculation cost caused by multiple inner and outer loop iterations and data transmission delay caused by re-iteration of each new signal transmission, reduces the calculation cost, and improves the power amplifier nonlinear correction efficiency.

[0048] The present application will be further described in detail below with reference to the accompanying drawings and the present application. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 is a flowchart of a DPD method based on a two-stage attention neural network provided by an embodiment of the present application;

[0050] Figure 2 is a flowchart of determining a rough estimate DPD parameter provided by an embodiment of the present application;

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

[0052] Figure 4 is a structural diagram of a digital pre-distortion system provided by an embodiment of the present application based on a two-stage attention neural network down-sampling DPD method;

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

[0054] Figure 6 is a power spectral density diagram of a power amplifier output signal after correction by the down-sampling DPD method provided by an embodiment of the present application;

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

[0056] Figure 8 is a structural diagram of a down-sampling DPD device based on a two-stage attention neural network provided by an embodiment of the present application;

[0057] Figure 9 is a structural diagram of an electronic device. DETAILED DESCRIPTION

[0058] The present application will be further described in detail below in combination with specific embodiments, but the embodiments of the present application are not limited thereto.

[0059] In order to solve the problems of large calculation cost caused by multiple inner and outer loop iterations and data transmission delay caused by re-iteration of each new signal transmission in the existing down-sampling DPD method, an embodiment of the present application provides a down-sampling DPD method based on a two-stage attention neural network, as shown in Figure 1 , Figure 1 is a flowchart of a down-sampling DPD method based on a two-stage attention neural network provided by an embodiment of the present application, specifically comprising the following steps:

[0060] In step S101, a memory polynomial method is used to determine a rough estimate DPD parameter.

[0061] In an embodiment of the present application, the memory polynomial method is used to determine a rough estimate DPD parameter, comprising:

[0062] A down-sampled baseband pilot signal and a down-sampled feedback signal are obtained;

[0063] constructing a memory polynomial model by using the down-sampled baseband pilot signal and the down-sampled feedback signal, and calculating model coefficients of the memory polynomial model;

[0064] obtaining a reconstructed feedback signal according to the model coefficients and the baseband pilot signal;

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

[0066] solving the inverse memory polynomial model by using a least square method to obtain a coarse estimation DPD parameter.

[0067] In the embodiment of the present application, referring to Figure 2 , Figure 2 is a flowchart for determining a coarse estimation DPD parameter provided by the embodiment of the present application. First, a 16QAM baseband original signal with a length of 1000 is generated, up-sampled by 5 times, and then a baseband pilot signal is obtained by using a raised cosine filter with a roll-off coefficient of 0.25. The baseband pilot signal is converted into an analog pilot signal by a DAC (Digital-to-Analog Converter), and then input into a wideband power amplifier through an up-converter. The behavior model of the wideband power amplifier can be modeled by using a memory polynomial model, and the wideband power amplifier model is represented as:

[0068]

[0069] wherein y(n) represents an output signal of the wideband power amplifier at the n th moment; h kq represents a 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 a signal value at the n th moment and the previous q moments.

[0070] Specifically, K = 5 and Q = 2, and the specific numerical settings are as follows:

[0071]

[0072] wherein j represents an imaginary unit; x(n) is an input signal of the wideband power amplifier at the n th moment. The output signal of the wideband power amplifier needs to be fed back to the front end through a feedback channel to realize the calculation of DPD. In the actual situation, the signal will introduce Gaussian white noise through the feedback channel, and the signal after the feedback channel is the superposition of the output signal of the wideband power amplifier and the Gaussian white noise signal, and therefore the final obtained feedback signal can be represented as:

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

[0074] Wherein, r(n) is feedback signal, y(n) is output signal of wideband power amplifier, w(n) is Gaussian white noise signal.

[0075] The feedback signal is down-sampled, for example, the sampling rate of ADC is set to 1 / 4 of the rate of DAC. After the feedback signal passes through the ADC with low rate, the down-sampled feedback signal r(m) is obtained, because the sampling rate is reduced by 4 times, so r(m)=r(4n). The baseband pilot signal is sampled at the same down-sampling rate, and the down-sampled baseband pilot signal x(m) is obtained. The memory polynomial model of the wideband power amplifier is constructed by using the down-sampled pilot baseband signal and the down-sampled feedback signal:

[0076] r=X D g;

[0077] Wherein, r represents the M-dimensional vector composed of the down-sampled feedback signal, M is the length of the down-sampled feedback signal, r=[r(0), r(1), …, r(M-1)] T ; g represents the model coefficient of the memory polynomial model, that is, the power amplifier coefficient, which is expressed as g=[g 10 ,…,g 1Q ,g 30 ,…,g 3Q ,…,g k0 ,…,g KQ ] T ; X D represents the Mx9-dimensional matrix composed of the down-sampled baseband pilot signal, which is defined as is the nonlinear memory polynomial with the nonlinear order k and the memory depth q at the mth moment, x(m-q) represents the signal value at the previous q moments at the mth moment, so it can be expressed as:

[0078]

[0079] Wherein, the superscript T represents the transpose operation.

[0080] The model coefficient of the memory polynomial model is calculated by using the least square method:

[0081]

[0082] Wherein, the superscript H represents the transpose conjugate operation.

[0083] In the embodiment of the application, the reconstructed feedback signal is obtained according to the model coefficient and the baseband pilot signal:

[0084]

[0085] Wherein, r rec(n) represents a reconstructed feedback signal at the n th moment, the reconstructed feedback signal being a feedback signal with a high sampling rate; g kq represents a model coefficient of a memory polynomial model with a nonlinear order of k and a memory depth of q; x(n-q) represents a baseband pilot signal without down-sampling at the n th moment.

[0086] An inverse memory polynomial model is constructed using the reconstructed feedback signal and the baseband pilot signal:

[0087] x = R rec a;

[0088] wherein x represents an N-dimensional vector composed of the baseband pilot signal without down-sampling, x = [x(0), x(1), …, x(N-1)] T ; R rec is similar to X D , which is an N x 9 matrix composed of the reconstructed feedback signal, N being the length of the pilot signal, wherein is defined as is a nonlinear memory polynomial with a nonlinear order of k and a memory depth of q at the n th moment; r rec (n-q) represents a signal value at the n th moment and the previous q moments, R rec is represented as:

[0089]

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

[0091] The coarse estimation of the DPD parameter a is obtained by using the least square method:

[0092]

[0093] The superscript H represents a transpose conjugate operation.

[0094] In step S102, a sample set is constructed based on a plurality of input sample signals and a coarse estimation of a DPD signal corresponding to each input sample signal; wherein the coarse estimation of the DPD signal corresponding to each input sample signal is determined based on the input sample signal and the coarse estimation of the DPD parameter.

[0095] Specifically, the sample set includes the real part, the imaginary part and the modulus value of each input sample signal, and the real part and the imaginary part of the coarse estimation DPD signal corresponding to the input sample signal. The input sample signal includes the current input sample signal and the historical input sample signal. The real part, the imaginary part and the modulus value of each input sample signal can be obtained according to the input complex sample signal, and the input complex signal value is known at the sending end. The calculation formula of the modulus value is Wherein, x I (n) is the real part of the signal, x Q (n) is the imaginary part of the signal. The real part and the imaginary part of the coarse estimation DPD signal corresponding to each input sample signal can be determined based on the input sample signal and the coarse estimation DPD parameter:

[0096]

[0097] Wherein, represents the coarse estimation DPD signal corresponding to the input sample signal at the n time; n = 0, 1,..., N; N represents the total number of input sample signals; a kq represents the coarse estimation 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 of the previous q time at the n time.

[0098] In the embodiment of the application, the sample set can 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 n time. Therefore, the sample set can be represented as represents the baseband pilot signal, which is a newly generated signal when the model is tested, u R (n) represents the real part of u(n), u I (n) represents the imaginary part of u(n), and |u(n)| represents the modulus value of u(n). Q represents the memory depth, and Q = 2. represents the coarse estimation DPD signal, and represents Wherein is the real part of , and is the imaginary part of . The output data set when training the model is There is no output data set when testing. The number of the training sample set is 1000, and the number of the test sample set is 100000.

[0099] Step S103: Train a pre-built DPD neural network model using a sample set, and estimate the downsampled DPD signal based on the trained DPD neural network model; training the pre-built DPD neural network model using a 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; and incrementally training the DPD neural network model based on the difference between the output DPD signal corresponding to the input sample signal and the coarsely estimated DPD signal.

[0100] In this embodiment of the invention, a DPD neural network model including an attention mechanism module and a fully connected module is constructed. See [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the DPD neural network model provided in an embodiment of the present invention. The first layer of the attention mechanism module is a non-linear layer with an activation function of tanh. This layer has 9 nodes, and the i-th output of the first layer can be expressed as:

[0101]

[0102] Where w i (n) represents the relationship between the i-th input and the estimated DPD output. Let {α} represent the sample vector of the nth input. i ,β i ,b i These are all parameters of the attention module, which are adjusted according to the loss function during model training. The second layer is the weight calculation layer, used to calculate the weight of each input neuron, reduce the importance of redundant estimates, improve the efficiency of using input data, and improve the performance of correction.

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

[0104]

[0105] Where, ξ i (n) represents the weight of the i-th data in the n-th input sample signal; exp() is the exponential operation function; w i (n) represents the output value of the i-th data point in the n-th input sample signal of the first layer; i is the index value of the input sample signal vector. Multiplying the calculated weights by the input sample signal yields the weighted signal. This step is in Figure 3 The Chinese character is represented as Since the input sample signal with high correlation has a larger weight, and the input sample signal with low correlation has a smaller weight, the weight of redundant information can be effectively reduced, thereby improving the accuracy of the correction.

[0106] In the embodiment of the present application, the noise in the feedback channel is considered, and the 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 embodiment of the present application can be not affected by the noise in the feedback channel, and still has good correction effect in the case of more noise, improving the stability.

[0107] The attention mechanism module can obtain the weight of each part of the input sample signal, fully utilize limited data to mine the desired information, overcome the defects of low data utilization efficiency and large required data in the prior art, so that the embodiment of the present application can still obtain good nonlinear correction results with less data.

[0108] The full connection module is located behind the attention mechanism module, and the weight calculated by the attention module is multiplied by the input sample signal to obtain a weighted signal χ i (n), which is the input of the full connection module. The full connection module includes two hidden layers. In order to simplify the model as much as possible, the number of nodes of the first hidden layer is set to 5, and the number of nodes of the second hidden layer is set to 10. All hidden layers use tanh activation function. The output of the full connection layer can be expressed as:

[0109]

[0110] Figure 3 wherein f1 represents tanh(C1χ(n)+b1), and f2 represents tanh(C2f1+b2)

[0111] wherein, represents the output DPD signal; C1, C2, b1 and b2 represent the coefficients of the hidden layer.

[0112] wherein C1 is expressed as:

[0113]

[0114] wherein b1 is expressed as L1 is the number of nodes of the first hidden layer.

[0115] In the embodiment of the present application, the expressions of C2 and b2 are similar, and only the number of nodes is different. C2 is expressed as:

[0116]

[0117] wherein b2 is expressed as L2 is the number of nodes of the second hidden layer. χ(n) is a vector composed of the weighted input sample signal, and χ(n) is expressed as χ(n)=[χ1(n),χ2(n),…,χ 3Q+3 (n)].

[0118] In the embodiment of the present application, 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 coarsely estimated DPD signal, comprising:

[0119] The loss is calculated according to the difference between the output DPD signal corresponding to the input sample signal and the coarsely estimated DPD signal by using a preset loss function.

[0120] 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 of times.

[0121] In the embodiment of the present application, the input sample signal in the sample set is input to the DPD neural network model during model training, 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, and the Adam optimization algorithm is an adaptive learning rate optimization algorithm used in deep learning. The loss function can be expressed as:

[0122]

[0123] 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; 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, i = 1 is the real part of the signal, and i = 2 is the imaginary part of the signal; is the predicted value, that is, the i-th signal value in the coarsely estimated DPD signal corresponding to the n-th input sample signal; the learning rate can be set to 0.085.

[0124] The down-sampling DPD method based on the two-stage attention neural network provided in the embodiment of the present application can be applied to a digital pre-distortion system, as shown in Figure 4 , Figure 4 is a digital pre-distortion system structure diagram applying the down-sampling DPD method based on the two-stage attention neural network provided in the embodiment of the present application, the system includes a baseband signal generation module, a baseband signal processing module and a radio frequency module; wherein 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 formally, a baseband transmission signal that needs to be transmitted will be generated.

[0125] The baseband signal processing module comprises a DPD, a DAC and a low-speed ADC, the DPD comprises a DPD pre-estimation module and an attention mechanism neural network model module.

[0126] The DPD neural network model is used for DPD signal estimation. According to the rough estimation of the DPD parameter and the input sample signal to be transmitted, a DPD signal can be obtained initially, which can reflect the correlation with the transmission signal, but is not accurate and is easily affected by the noise in the feedback channel. According to the correlation, the current and historical transmission signal real part, imaginary part and modulus weight can be obtained. Further, the DPD neural network model is used to learn the DPD behavior to obtain an accurate DPD signal. The DPD neural network model needs to be trained when the system starts to work, and the transmission signal can be directly input into the model when the system formally works, and the output of the model is the accurate 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 down-sample the feedback signal.

[0127] The radio frequency module comprises a signal up-converter, a wideband power amplifier PA, a feedback channel, an attenuator, a signal down-converter and an anti-aliasing filter, which is used for modulating the baseband signal to a high-frequency carrier and amplifying the signal power by the power amplifier. The feedback channel is used for feeding back the power amplifier output signal to the system front end to calculate the DPD. Before the DPD calculation, the signal needs to pass through the attenuator to remove the influence of the signal power amplification caused by the power amplifier, and the down-converter and the anti-aliasing filter are used to restore the signal to the baseband signal.

[0128] When the system provided by the application example based on the two-stage attention neural network down-sampling DPD method is formally worked, the system structure is simple, the DPD neural network model is worked, the existing system in the working stage still needs a large number of DPD calculation modules to work, the complexity of the system is reduced, and the working efficiency of the system is improved.

[0129] In the embodiment of the present application, the rough estimation of the DPD parameter only needs 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 at the beginning using the constructed sample set, and during use, the trained DPD neural network model can be directly used to realize the down-sampling DPD signal estimation, overcoming the problems of the existing down-sampling DPD method, i.e. the large calculation cost caused by multiple inner and outer loop iterations and the data transmission delay caused by the need to re-iterate each time a new signal is transmitted, thereby reducing the calculation cost and improving the efficiency of the power amplifier nonlinear correction.

[0130] The simulation experiment of the down-sampling DPD method based on the two-stage attention neural network provided by the embodiment of the present application is as follows:

[0131] The power spectrum density graph of the power amplifier output signal before and after correction is drawn, and the severity of the power spectrum density extension can be observed to determine the severity of the signal nonlinear distortion. The more serious the spectrum expansion, the more serious the distortion. See Figure 5 , Figure 5 is the power spectrum density graph of the power amplifier output signal without DPD. The power spectrum density of the power amplifier output signal without DPD correction is significantly expanded, and the lower the signal-to-noise ratio, the more serious the distortion. See Figure 6 , Figure 6 is the power spectrum density graph of the power amplifier output signal after correction using the down-sampling DPD method provided by the embodiment of the present application. It can be seen that the power spectrum density extension should be significantly improved, and the power spectrum density of the signal is very close to the power spectrum density of the original signal, and the correction effect can still be maintained as the noise increases,

[0132] See Figure 7 , Figure 7 is a comparison diagram of the signal constellation before and after correction. The constellation comparison diagram of the signal before and after correction when SNR is 15dB is drawn. SNR represents the signal-to-noise ratio. It can be seen that without correction, the constellation diagram of the signal is diffused and deformed, which will affect the decoding and recognition of the subsequent signal, leading to an increase in the bit error rate, thereby seriously affecting the communication quality. After using the down-sampling DPD method based on the two-stage attention neural network provided by the embodiment of the present application, the distortion of the signal constellation diagram is significantly improved, which can effectively reduce the increase in the bit error rate caused by the distortion of the power amplifier at the receiving end, thereby improving the communication quality.

[0133] Based on the same inventive concept, the embodiment of the present application also provides a down-sampling DPD device based on a two-stage attention neural network, see Figure 8 , Figure 8is a structure schematic diagram of a down-sampling DPD device based on a two-stage attention neural network provided by an embodiment of the present application, and the down-sampling DPD device comprises:

[0134] The determination module 801 is configured to determine the coarse DPD parameter by using a memory polynomial method.

[0135] The construction module 802 is configured to construct a sample set based on a plurality of input sample signals and a coarse DPD signal corresponding to each input sample signal, wherein the coarse DPD signal corresponding to each input sample signal is determined based on the input sample signal and the coarse DPD parameter.

[0136] The training module 803 is configured to train a pre-constructed DPD neural network model by using the sample set, and to realize down-sampling 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 comprises: 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 performing incremental training on the DPD neural network model according to a difference between the output DPD signal corresponding to the input sample signal and the coarse DPD signal.

[0137] In the embodiment of the present application, the coarse DPD parameter only needs 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 at the beginning by using the constructed sample set, and down-sampling DPD signal estimation can be realized directly according to the trained DPD neural network model in use, thereby overcoming the problems of large calculation cost caused by multiple inner and outer loop iterations and data transmission delay caused by the need to reiterate each time a new signal is transmitted in the existing down-sampling DPD method, reducing the calculation cost and improving the power amplifier nonlinear correction efficiency.

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

[0139] acquire a down-sampled baseband pilot signal and a down-sampled feedback signal; construct a memory polynomial model by using the down-sampled baseband pilot signal and the down-sampled feedback signal, and calculate 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 obtain the coarse DPD parameter by solving the inverse memory polynomial model by using a least square method.

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

[0141]

[0142] wherein, denotes the coarse estimation DPD signal corresponding to the nth input sample signal; n = 0, 1, …, N; N denotes the total number of input sample signals; a kq denotes the coarse estimation DPD parameter obtained when the non-linear order is k and the memory depth is q; k = 1, 2, …, K; K denotes the maximum non-linear order; q = 0, 1, …, Q; Q denotes the maximum memory depth; u(n-q) denotes the signal value of the previous q time points at the nth time point.

[0143] Optionally, the training module performs 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 coarse estimation DPD signal, including:

[0144] using a preset loss function, calculating a loss according to the difference between the output DPD signal corresponding to the input sample signal and the coarse estimation DPD signal;

[0145] 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 of times to complete the training.

[0146] Optionally, the loss function is:

[0147]

[0148] wherein, Loss denotes the loss; n = 0, 1, …, N0-1; N0 denotes the total number of input sample signals used to calculate the loss function; denotes 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; denotes the ith signal value in the coarse estimation DPD signal corresponding to the nth input sample signal.

[0149] The embodiment of the present application also provides an electronic device, such as Figure 9 as shown, Figure 9 is a structural schematic diagram of an electronic device, which comprises a processor 901, a communication interface 902, a memory 903 and a communication bus 904, wherein 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 a computer program.

[0151] The processor 901 is used to execute the program stored on the memory 903, and realizes the method steps of any one of the above-mentioned two-stage attention neural network-based down-sampling DPD methods.

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

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

[0154] The memory can include a Random Access Memory (RAM), and can also include a Non-Volatile Memory (NVM), for example, at least one disk memory. Optionally, the memory can also be at least one storage device located away from the above-mentioned processor.

[0155] The above-mentioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; can also be a Digital Signal Processing (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component.

[0156] The application further provides a computer readable storage medium. The computer readable storage medium stores a computer program. When the computer program is executed by a processor, the method steps of any one of the above-mentioned DPD methods based on a two-stage attention neural network are implemented.

[0157] Optionally, the computer readable storage medium can be a Non-Volatile Memory (NVM), for example, at least one disk memory.

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

[0159] In yet another embodiment of the present application, there is also provided a computer program product comprising instructions which, when executed on a computer, cause the computer to carry out the method steps of any of the above described methods of down-sampling DPD based on a two-stage attention neural network.

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

[0161] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", 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 application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in the present specification.

[0162] Although the present application is described herein in connection with various embodiments, those skilled in the art will appreciate and understand that other variations of the disclosed embodiments can be understood and implemented in the process of implementing the claimed application by viewing the drawings and the disclosure. In the description of the present application, the word "comprising" does not exclude other components or steps, "one" or "an" does not exclude a plurality, and "plurality" means two or more, unless otherwise explicitly specified. In addition, some measures are described in different embodiments, but this does not mean that these measures cannot be combined to produce good results.

[0163] The method provided by the embodiments of the present application can be applied to electronic devices. Specifically, the electronic device can be: a desktop computer, a portable computer, a smart mobile terminal, a server, etc. Herein, any electronic device that can implement the present application is within the protection scope of the present application.

[0164] For device / electronic device / storage medium embodiments, because they are basically similar to method embodiments, the description is relatively simple, and the relevant part can be referred to the part of the method embodiment.

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

[0166] The above is a further detailed description of the present application in combination with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For ordinary skilled persons in the technical field to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be considered as falling within the protection scope of the present application.

Claims

1. A down-sampling DPD method based on a two-stage attention neural network, characterized in that, The downsampling DPD method comprises: The rough estimation DPD parameter is determined by using a memory polynomial method; A sample set is constructed based on a plurality of input sample signals and a rough estimation DPD signal corresponding to each input sample signal, wherein the rough estimation DPD signal corresponding to each input sample signal is determined based on the input sample signal and the rough estimation DPD parameter; The pre-constructed DPD neural network model is trained by using the sample set, and the downsampling DPD signal estimation is realized according to the trained DPD neural network model; The pre-constructed DPD neural network model is trained by using the sample set, and the downsampling DPD signal estimation is realized according to the trained DPD neural network model; 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 rough estimation DPD signal; The rough estimation DPD parameter is determined by using a memory polynomial method, which comprises: The down-sampled baseband pilot signal and the down-sampled feedback signal are obtained; A memory polynomial model is constructed by using the down-sampled baseband pilot signal and the down-sampled 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 by using the reconstructed feedback signal and the baseband pilot signal; The inverse memory polynomial model is solved by using a least square method, and the rough estimation DPD parameter is obtained; The rough estimation DPD signal corresponding to each input sample signal is: ； in, Indicates the first A coarse estimate of the DPD signal corresponding to each input sample signal; ; This indicates the total number of input sample signals; This indicates that in nonlinear order of... Memory depth is At that time, the coarse estimate of DPD parameters was obtained; ; Indicates the maximum nonlinear order; ; Indicates the maximum memory depth; Indicates the first Before the moment q The signal value at that moment; The DPD neural network model comprises an attention mechanism module and a full connection module; The first layer of the attention mechanism module is a nonlinear layer; the output of the nonlinear layer is represented as: ​ ； wherein, represents a relationship between the i th input and the estimated DPD output, represents a sample vector of the n th input, are parameters of the attention mechanism module; tanh represents an activation function; superscript represents a transpose operation; The second layer of the attention mechanism module is a weight calculation layer; the first output of the weight calculation layer is represented as: ​ ； wherein, is the weight of the n th data in the i th input sample signal; is an exponential operation function; is the output value of the n th data in the i th input sample signal of the first layer; i is an index value of the input sample signal vector.

2. The method of claim 1, wherein, 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 estimation DPD signal, which comprises: A loss is calculated by using a preset loss function according to the difference between the output DPD signal corresponding to the input sample signal and the rough estimation DPD signal; The function value is iteratively updated in the direction of reducing the loss until the loss function converges or the iteration number reaches a preset number.

3. The down-sampling DPD method of claim 2, wherein, The loss function is: ； wherein represents the loss; ; represents the total number of input sample signals used to calculate the loss function; represents the th signal value in the output DPD signal corresponding to the th input sample signal, i is an index of the signal value; represents the th signal value in the coarse estimate DPD signal corresponding to the th input sample signal.

4. A down-sampling DPD apparatus based on a two-stage attention neural network, characterized in that, The downsampling DPD device comprises: A determination module configured to determine a rough estimation 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 rough estimation DPD signal corresponding to each input sample signal, wherein the rough estimation DPD signal corresponding to each input sample signal is determined based on the input sample signal and the rough estimation DPD parameter; The training module is configured to train a pre-constructed DPD neural network model by using the sample set and to implement down-sampling DPD signal estimation according to the trained DPD neural network model. The determining module is specifically configured to: The down-sampling baseband pilot signal and the down-sampling feedback signal are obtained, and a memory polynomial model is constructed by using the down-sampling baseband pilot signal and the down-sampling 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 by using the reconstructed feedback signal and the baseband pilot signal; and the inverse memory polynomial model is solved by using a least square method to obtain the coarse estimation DPD parameter. The coarse estimation DPD signal corresponding to each input sample signal is: ； wherein, denotes a coarse estimate DPD signal corresponding to the input sample signal; ; denotes the total number of input sample signals; denotes the coarse estimate DPD parameters obtained for a nonlinearity order of and a memory depth of ; ; denotes the maximum nonlinearity order; ; denotes the maximum memory depth; denotes the signal value of the preceding q time instants. The DPD neural network model comprises an attention mechanism module and a full connection module. The first layer of the attention mechanism module is a nonlinear layer; the output of the nonlinear layer is represented as: ​ ； wherein, represents a relationship between the i input and the estimated DPD output, represents a sample vector of the n input, are parameters of the attention mechanism module; tanh represents an activation function; superscript represents a transpose operation; The second layer of the attention mechanism module is a weight calculation layer; the first output of the weight calculation layer is represented as: ​ ； wherein, is the weight of the n th data in the i th input sample signal; is an exponential operation function; is the output value of the n th data in the i th input sample signal of the first layer; i is an index value of the input sample signal vector.

5. An electronic device, comprising: The device comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory are in communication with each other through the communication bus. The memory is configured to store a computer program. The processor is configured to execute the computer program stored in the memory to implement the down-sampling DPD method of any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer program stored in the computer readable storage medium is executed by the processor to implement the down-sampling DPD method of any one of claims 1-3.

Images