Design Method of Neural Network-Based Digital Predistortion Model

Abstract

The present invention belongs to the technical field of digital predistortion in satellite communication transmission. Aiming at the technical problems of the lack of utilization of effective information and phase offset in existing digital predistortion algorithms, a "design method of a digital predistortion model based on neural network" is proposed. Combining the amplitude / phase and isolation and interaction characteristics of the quadrature / in-phase demultiplexing of communication data, a digital predistortion model integrating vector decomposition and multi-expert decision is designed based on the LSTM model. Considering the in-phase / quadrature components and amplitude / phase information comprehensively, the original distorted signal and the phase recovery signal are spliced across layers through a vector splicing module to learn the isolation and interaction effects between different features.

Description

Technical Field

[0001] The present invention belongs to the technical field of digital predistortion in satellite communication transmission. Background Art

[0002] Satellite communication has many advantages such as wide coverage, long transmission distance, large communication capacity, good transmission quality, flexible and rapid networking, and high confidentiality, and has become a highly competitive communication means today. As a key component in a satellite communication system, the linearization technology of a radio frequency power amplifier is crucial for improving the quality of satellite communication; and DPD (Digital Predistortion Technology) is an important means of the linearization technology of a radio frequency power amplifier, which compensates for the nonlinear distortion at the output end of the radio frequency power amplifier by introducing opposite distortion at the signal input end, avoiding problems such as a decline in signal transmission quality, a reduction in spectral efficiency, and adjacent channel interference.

[0003] For example, the predistortion algorithm based on a polynomial model can reduce nonlinear distortion to a certain extent, but it has limitations in dealing with complex nonlinear characteristics. Especially in a communication environment with multiple frequency bands and multiple channels, it is difficult to meet the high-precision requirements of modern communication systems. Due to its flexibility and self-adaptability, a neural network can better capture and compensate for complex nonlinear distortion, and has been widely used in the field of digital predistortion in recent years. Existing neural network digital predistortion methods include:

[0004] Combined with an RNN (Recurrent Neural Network), the memory effect of the RNN is used for modeling, which improves the predistortion effect in a dynamic environment. However, due to the problem of gradient disappearance or gradient explosion during the process of calculating long sequence features, it is difficult to capture long-distance time-dependent relationships, and the performance shows a weak memory ability for historical inputs. Therefore, a variant LSTM model of the RNN is proposed. By adding cell memory units and a gating mechanism to select and remember or forget important feature information, it can capture the memory effect in a longer time domain range in digital predistortion technology. However, the LSTM model lacks the utilization of signal amplitude and phase information.

[0005] Furthermore, by constructing a VDLSTM (Vector Decomposition Long Short-Term Memory Network) to perform nonlinear calculation on the signal amplitude, and then restoring the phase in a linear weighted manner, the technical problem that the above LSTM model lacks the utilization of signal amplitude and phase information is solved. However, the VDLSTM completely focuses on amplitude and phase information during calculation, lacking the utilization of the isolation and interaction between signal quadrature and in-phase branches, which affects the final result of nonlinear compensation; at the same time, the phase is restored by a linear weighted summation method, which ignores the nonlinearity and memory of the phase, resulting in phase offset. Summary of the Invention

[0006] To improve the limitations of existing digital predistortion algorithms, which lack the utilization of effective information and have technical problems with phase offset, the present invention proposes a "method for designing a digital predistortion model based on a neural network".

[0007] The method for designing a digital predistortion model based on a neural network, as Figure 1 shown, the predistortion model includes an amplitude learning module, a phase calculation module, a phase recovery module, a vector splicing module, a multi-expert network learning module, and a fusion decision module;

[0008] After the input data is obtained, the distorted signal is decomposed into amplitude and phase, and input into the amplitude learning module for amplitude learning to output an amplitude learning signal and input into the phase calculation module for processing by using sine and cosine functions to respectively generate phase information signals for the in-phase path and the quadrature path; the amplitude learning signal and the phase information signal are input into the phase recovery module for linear transformation and multiplied by the in-phase path and quadrature path phase signals to output a phase recovery signal, the phase recovery signal and the distorted signal are input into the vector splicing module for splicing in the feature dimension to output a spliced signal, the spliced signal is input into the multi-expert network learning module, and a multi-expert decision signal is output through calculation, and the multi-expert decision signal is input into the fusion decision module to output a final decision signal;

[0009] The amplitude learning module includes a gating mechanism and a cell memory unit. The decomposed distorted signal updates the cell memory unit and the hidden layer state of the decomposed signal itself through the gating mechanism and cyclic calculation. When the cyclic calculation ends, an amplitude learning signal is output; the gating mechanism includes an input gate, a forget gate, and an output gate, and the specific calculation formulas are as follows:

[0010] ,

[0011] : The activation value of the forget gate, where 1 is defined as completely retaining and 0 is defined as completely discarding;

[0012] : The Sigmoid activation function;

[0013] : The weights and biases of the forget gate;

[0014] : The hidden state at the previous moment;

[0015] : The input at the current moment;

[0016] ,

[0017] : The activation value of the input gate;

[0018] : Weights and biases of the input gate;

[0019] ,

[0020] : Activation value of the output gate;

[0021] : Weights and biases of the output gate;

[0022] ,

[0023] : State of the cell memory unit at the current time;

[0024] : State of the cell memory unit at the previous time;

[0025] : Candidate memory at the current time;

[0026] ,

[0027] () is the hyperbolic tangent function;

[0028] 、 : Weights and biases of the cell memory unit;

[0029] ,

[0030] : Hidden state at the current time;

[0031] The multi-expert network learning module receives the splicing signal, and respectively focuses on learning the quadrature branch, the feature interaction branch, and the in-phase branch through each channel expert network, and correspondingly outputs the quadrature branch signal, the feature interaction branch signal, and the in-phase branch signal;

[0032] The fusion decision module receives the quadrature branch signal, the in-phase branch signal, and the feature interaction branch signal, performs feature fusion in a weighted fusion manner, and outputs the final decision signal; After building the predistorter model and performing iterative training, the final VDMOE (Vector Decomposed Mixture Of Experts) model is obtained.

[0033] Technical effects:

[0034] Combining the isolation and interaction characteristics of the amplitude / phase and quadrature / in-phase splitting of communication data, a digital predistorter model integrating vector decomposition and multi-expert decision-making is designed based on the LSTM model.

[0035] 1. Comprehensively consider the in-phase / quadrature components and amplitude / phase information to avoid the problem of loss of effective feature information in the input.

[0036] 2. Cross-layer splice the original distorted signal and the phase recovery signal through the vector splicing module to avoid forgetting the initial features as the learning progresses in the network learning process, solve the technical problem that the existing methods ignore the non-linearity and memory of the phase, and at the same time avoid problems such as excessive compression or stretching of features and vanishing gradients or exploding gradients due to non-linear operations in a certain layer, and accelerate the convergence speed.

[0037] 3. Perform subsequent non-linear time series learning through the multi-expert network learning module to learn the isolation and interaction between different features, further enhance the adaptability of the algorithm, and finally obtain the final decision signal by weighted summation of the discrimination results of each expert channel. The non-linear learning of the phase also solves the technical problem of phase offset.

[0038] To illustrate the excellent performance of the method used in the present invention, the VDMOE model of the present invention is used to predict the test data, and at the same time, its performance is evaluated with the existing LSTM model and VDLSTM model through various metrics. The specific evaluation metrics include the power spectral density diagram, the mean square error value (MSE), and the AM-AM and AM-PM characteristic curves. As Figure 3 shown, the power spectral density diagram shows the spectral characteristics of the signal after predistortion processing. The distortion of the signal is improved after adding the predistorter. By comparing the differences in the corresponding curves of the LSTM model and the VDLSTM model, it can be found that the present invention (green line) is closer to the original signal.

[0039] The mean square error (MSE), as a commonly used error measurement index, can quantify the difference between the output signal of the predistorter and the ideal signal, and reflect the accuracy and reliability of the predistorter in signal recovery. The following table shows that the model built by the present invention has the smallest MSE value and the predicted value is closest to the ideal signal.

[0040] DPD MSE (dB) Without adding DPD -19.52 LSTM -57.13 VDLSTM -54.66 VDMOE -59.22

[0041] The AM-AM and AM-PM characteristic curves can intuitively show the amplitude and phase changes of the signal under different input powers. As Figure 4As shown in the figure, by comparing the AM-AM diagrams of the signals before and after adding the VDMOE model, it can be found that there is obvious nonlinear distortion without passing through the VDMOE model. In contrast, the VDMOE model performs timing calculations on the signal amplitude through the vector decomposition module and learns the isolation and interaction between multi-dimensional features through the multi-expert network, making its nonlinear distortion performance better on the AM-AM. As Figure 5 shown in the figure, by comparing the AM-PM diagrams of the signals before and after different DPD models, it can be found that the phase distortion is relatively serious when the DPD is the VDLSTM model. In contrast, considering the phase timing and memory characteristics in the present invention and performing subsequent nonlinear learning through the multi-expert network learning module, the phase distortion is smaller. Brief Description of the Drawings

[0042] Figure 1 This is the overall structural block diagram of the predistorter model of the present invention.

[0043] Figure 2 This is the schematic flow block diagram of the training mechanism process of the embodiment of the present invention.

[0044] Figure 3 This is the comparison diagram of the power spectral density of different DPD models in the technical effects.

[0045] Figure 4 This is the comparison diagram of the AM-AM characteristic curves of the signals before and after the present invention.

[0046] Figure 5 This is the comparison diagram of the AM-PM characteristic curves of the signals before and after different DPD models. Detailed Embodiment

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Using the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

[0048] In this embodiment, data acquisition is first performed. The signal source is modulated by Binary Phase Shift Keying (BPSK) to obtain an initial signal. After preprocessing the initial signal, a two-dimensional distorted signal in the time dimension and the feature dimension is obtained. The initial signal and the distorted signal are aligned and collected, and the distorted signal is input into the built predistorter model;

[0049] Further, the specific preprocessing of the initial signal is as follows: after up-conversion operation, the distorted signal is obtained through a radio frequency power amplifier and attenuation, and the radio frequency power amplifier operates in a non-linear working range.

[0050] The gating mechanism described in the amplitude learning module includes an input gate, a forget gate, and an output gate. By introducing the gating mechanism to regulate the flow of information, the problems of gradient disappearance and explosion are effectively alleviated. The specific calculation formulas are as follows:

[0051] , ①

[0052] , ②

[0053] , ③

[0054] , ④

[0055] , ⑤

[0056] , ⑥

[0057] Among them, t represents the current moment, and t - 1 represents the previous moment. is the input at the current moment, f represents the forget gate, i represents the input gate, o represents the output gate, C represents the memory cell state. is the candidate memory at the current moment, h represents the hidden state. and are the weight and bias parameters. represents the activation function Sigmoid(). () represents the hyperbolic tangent function.

[0058] Furthermore, the expression of the activation function Sigmoid() is as follows:

[0059] ,

[0060] The expression of the hyperbolic tangent function tanh() is as follows:

[0061] ,

[0062] Among them, is a constant. is the input variable.

[0063] The vector concatenation module performs horizontal concatenation on features of different dimensions to expand them into the form of high-dimensional features. The specific calculation formulas are as follows:

[0064] ,

[0065] ,

[0066] ,

[0067] Among them, and respectively represent the in-phase and quadrature components of the distorted signal, and respectively represent the in-phase and quadrature components of the in-phase branch of the phase recovery signal, , , are the spliced signals input to the multi-expert network learning module. Specifically, is the splicing vector input to the expert network of the quadrature branch channel, and outputs the quadrature branch signal; is the splicing vector input to the expert network of the feature interaction branch channel, and outputs the feature interaction branch signal; is the splicing vector input to the expert network of the in-phase branch channel, and outputs the in-phase branch signal. Each channel expert network is calculated separately, and the structure is an LSTM neural network and a fully connected layer. The calculation formulas are the same as Formulas ①~⑥, which will not be elaborated here.

[0068] In this embodiment, the pre-distorter model is iteratively trained in a way of pre-training plus fine-tuning.

[0069] As Figure 2 shown, on the left is the pre-training link. First, a post-inverse model is trained, and its weights are used as the initial weights of the pre-distorter model. The input of the post-inverse model is the attenuated distorted signal, and the ideal target is the signal directly output by the signal source. The post-inverse model is equivalent to being complementary to the power amplifier to form distortion compensation. Then, through the indirect learning structure, it is provided to the loop fine-tuning training on the right to further fine-tune the pre-distorter model. First, an ideal gain signal is defined. Assuming the signal output by the signal source is , after linear amplification, the ideal signal without distortion is . In the experiment, we refer to the amplification gains of the power amplifier and the attenuator, and set so as to correspond to the training environment of the post-distortion model.

[0070] Furthermore, the AdamW backpropagation algorithm is used for iterative training. To further improve the calculation efficiency and the robustness of the model, batch training is performed on the network and the batch size is set to 256. The initial learning rate of the network is set to 0.0005; the total number of training iterations is 50 times, and the validation set score is calculated once for each iteration; finally, the model with the highest validation set score is selected and migrated to the test set for calculation. The overall training process of the model uses a GPU (Graphics Processing Unit) server, and the GPU chip is NVIDIA GeForce RTX 4090 with a memory capacity of 24GB.

[0071] The content not described in detail in this specification belongs to the prior art well-known to those skilled in the art. At the same time, for those of ordinary skill in the art, there will be changes in the specific implementation and application scope according to the idea of the present invention. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for designing a digital predistorter model based on a neural network, characterized in that: The predistorter model includes an amplitude learning module, a phase calculation module, a phase recovery module, a vector splicing module, a multi-expert network learning module and a fusion decision module; After the input data is acquired, the distorted signal is decomposed into amplitude and phase, input into the amplitude learning module for amplitude learning to output the amplitude learning signal, and input into the phase calculation module for processing by using sine and cosine functions to generate phase information signals of the in-phase path and the orthogonal path respectively; the amplitude learning signal and the phase information signal are input into the phase recovery module through linear transformation and multiplied with the in-phase path and the orthogonal path phase signals to output the phase recovery signal; the phase recovery signal and the distorted signal are input into the vector splicing module, spliced ​​in the feature dimension to output the spliced ​​signal; the spliced ​​signal is input into the multi-expert network learning module, and the multi-expert decision signal is output through calculation; the multi-expert decision signal is input into the fusion decision module to output the final decision signal; The amplitude learning module includes a gating mechanism and a cell memory unit. The decomposed distorted signal is updated through the gating mechanism and cyclic calculation to update the hidden layer state of the cell memory unit and the decomposed signal itself. When the cyclic calculation is completed, the amplitude learning signal is output; the gating mechanism includes a forget gate f, an input gate i and an output gate o. The specific calculation formula is as follows: ， : The activation value of the forget gate, 1 means completely retained, and 0 means completely discarded; : Sigmoid activation function; : Weight and bias of the forget gate; : The hidden state of the previous moment; : Input at the current moment; ， : Activation value of input gate; : weight and bias of input gate; ， : The activation value of the output gate; : Weight and bias of the output gate; ， : The current state of the cell memory unit; : The cell memory unit state at the last moment; : Candidate memory at the current moment; ， () is the hyperbolic tangent function; , : Weights and biases of cell memory units; ， : The hidden state at the current moment; The multi-expert network learning module receives the spliced ​​signal, and respectively focuses on learning the orthogonal branch, the feature interaction branch and the in-phase branch through each channel expert network, and correspondingly outputs the orthogonal branch signal, the feature interaction branch signal and the in-phase branch signal; The fusion decision module receives the orthogonal branch signal, the in-phase branch signal and the feature interaction branch signal, performs feature fusion in a weighted fusion manner, and outputs a final decision signal; after building a pre-distorter model and performing iterative training, the final VDMOE model is obtained.

2. The method for designing a digital predistorter model based on a neural network according to claim 1, characterized in that: The input data acquisition is specifically as follows: the signal source is modulated by two-phase shift keying to obtain an initial signal, the initial signal is preprocessed to obtain a two-dimensional distorted signal in the time dimension and the feature dimension, the initial signal and the distorted signal are aligned and collected, and the distorted signal is input into the constructed predistorter model.

3. The method for designing a digital predistorter model based on a neural network according to claim 2, characterized in that: The initial signal preprocessing is specifically as follows: after the frequency up-conversion operation, the distorted signal is obtained after passing through a radio frequency power amplifier and attenuation, and the radio frequency power amplifier operates in a nonlinear working range.

4. The method for designing a digital predistorter model based on a neural network according to claim 1, characterized in that: The activation function in the amplitude learning module is specifically: , the hyperbolic tangent function is specifically , in, is a constant, is the input variable.

5. The method for designing a digital predistorter model based on a neural network according to claim 1, characterized in that: The specific calculation formula of the vector splicing module is as follows: ， ， ， in, and represent the in-phase and quadrature components of the distorted signal, respectively. and Represent the in-phase and quadrature components of the phase recovery signal, respectively. is the concatenation vector of the input orthogonal branch channel expert network, is the concatenated vector of the input feature interactive branch channel expert network, is the concatenation vector of the input in-phase branch channel expert network.

6. The method for designing a digital predistorter model based on a neural network according to claim 1, characterized in that: After the predistorter model is built, iterative training is performed as follows: first, a post-inverse model is trained, and its weight is used as the initial weight of the predistorter model. The input of the post-inverse model is the distorted signal after attenuation processing. The ideal goal is that the signal source directly outputs the signal. The post-inverse model and the power amplifier complement each other to form distortion compensation. After that, the predistorter model is further fine-tuned through the indirect learning structure, and the signal source output signal is , the ideal gain signal after linear amplification without distortion is , update the error of the predistorter model and set .

7. The method for designing a digital predistorter model based on a neural network according to claim 6, characterized in that: The back propagation algorithm used in iterative training is AdamW, the batch size is set to 256, the initial learning rate of the network is set to 0.0005, and the total number of training iterations is 50.

Images