Digital pre-distortion method based on KAN and novel neural network
Through the combination of KAN and new neural networks, the problems of large amount of computation and overfitting of traditional algorithms are solved, and non-linear fitting of low-computing and high-precision power amplifiers are achieved, which is suitable for low-computing power platforms.
Patent Information
- Application Number
- CN202510348961.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-24
- Publication Date
- 2025-07-25
AI Technical Summary
The amount of parameters of the existing traditional digital predistortion algorithm increases with memory depth and order, resulting in large amounts of calculations, making it difficult to deploy on low-computing platforms, and the existing deep network models are easily overfitted, making it difficult to effectively fit the nonlinear distortion of the power amplifier.
The Kolmogorov-Arnold network (KAN) is used to extract the delayed 5G signals, and then splice it with the original delayed 5G signals and input it into a new neural network (NNN). Through multiple rounds of signal propagation and full-connection layer processing, non-linear fit is achieved, and the calculation complexity is reduced in combination with logarithmic operations.
It realizes nonlinear fitting of power amplifiers with high precision and low computing volume, is suitable for low computing power platforms, and has strong versatility and efficient nonlinear signal processing capabilities.
Smart Images
Figure CN120377835A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of digital signal processing, and particularly relates to a digital predistortion method based on KAN and a new neural network. Background Art
[0002] As an important component in a wireless communication system, a power amplifier provides sufficient power for transmitting signals to ensure that the signals can maintain sufficient strength during transmission, so as to be able to cover a farther distance or pass through obstacles. However, when the power amplifier operates near the saturation state, nonlinear distortion will occur, resulting in nonlinear distortion of the signals and thus affecting the performance of the communication system.
[0003] However, the number of parameters of existing traditional digital predistortion algorithms increases exponentially with the memory depth and order, so it is difficult to utilize the features of high dimensions or long time slices, resulting in difficulty in improving the nonlinear fitting ability. And existing digital predistortion algorithms based on deep networks have a large computational amount and a large model scale, are not suitable for online parameter solving, require a large amount of training data, are prone to overfitting, and are difficult to be deployed on low-computing-power platforms. Therefore, it is necessary to seek a digital predistortion neural network model with advanced performance and low computational amount to identify and regress the nonlinear signal processing behavior. Summary of the Invention
[0004] The main purpose of the present invention is to overcome the disadvantages and deficiencies of the existing technology, and provide a digital predistortion method based on KAN and a new neural network, which can achieve efficient nonlinear characterization of radio frequency signals and reach an ideal model accuracy under the premise of low computational amount.
[0005] To achieve the above purpose, the present invention adopts the following technical solutions: A digital predistortion method based on KAN and a new neural network, the digital predistortion method comprising the following steps: S1. Perform a logarithmic operation on the time-delay 5G signal in complex form and then restore it to real form and input it into the Kolmogorov-Arnold network, hereinafter referred to as KAN for short; S2. Concatenate the output of KAN and the original time-delay 5G signal and input them into the new neural network in real form, and perform signal propagation for a certain number of rounds in the new neural network according to parameter settings, hereinafter referred to as NNN for short; S3. Extract the output of NNN and input it into a fully connected layer to generate a two-dimensional output.
[0006] Further, in step S1, the time-delayed 5G signal is first logarithmically operated in complex form and then restored to real form and input into KAN. This step converts the multiplication between different moments of the time-delayed 5G signal into addition under logarithm, reducing the computational complexity and enabling the subsequent network to fit the memory distortion between data.
[0007] Further, in step S2, the features extracted by the neural network KAN and the original time-delayed 5G signal are concatenated and input into the new neural network in real form. The features extracted by KAN are basis functions for fitting memory distortion. When combined with the original signal, they can accurately fit the non-linear distortion of the signal in the subsequent network.
[0008] Further, the new neural network has two types of nodes: input nodes and central nodes. The input nodes are used to receive the input from the upper layer network, and the central nodes can further improve the non-linear fitting ability of the new neural network by increasing the number of nodes.
[0009] Further, each node in the new neural network has three state variables, and the state variables of all nodes are combined together to form a memory pool , a candidate pool and an output pool . The memory pool is the medium for multi-round signal propagation. Multiple rounds of signals are superimposed on the memory pool and finally generate an output. The candidate pool resets the corresponding node state variables that reach the threshold in each round of signal propagation, and the output pool outputs the corresponding node state variables that reach the threshold in each round of signal propagation to the memory pool.
[0010] Further, the signal propagation method in the new neural network is
[0011]
[0012]
[0013] ; where represents the adjacency matrix of all nodes randomly initialized in the new neural network, represents a zero vector of dimension, n is the dimension of the input node, is the number of expandable nodes that can be set, represents the intermediate variable of this round of propagation, min(, ) represents the minimum operation, The expression of In this step, by simulating the EEG propagation mode in the brain, when the cumulative potential in the candidate pool reaches the activation threshold, the output pool generates an output, and the memory pool accumulates the output pool information in multiple rounds of propagation. The signal propagation mode of the class residual connection of the three pools enables step-by-step processing of the signal, which is consistent with the physical characteristic of signal attenuation over time. The above network design enables the network to well handle the nonlinearity and memory effect of the power amplifier.
[0014] Furthermore, the output of the new neural network is the memory pool. The memory pool fully fits the nonlinearity of the power amplifier in multiple rounds of propagation.
[0015] Furthermore, in step S3, the input dimension of the fully connected layer is the dimension of the memory pool of the new neural network, and the output is a two-dimensional output, denoted as This step regresses the features obtained by the upper-layer network fitting to obtain the actual digital predistortion result.
[0016] Furthermore, in step S3, an ideal digital predistortion data set generated by iterative learning control (ILC) with the original distorted data is output as the label, and the KAN and the new neural network are trained with the mean square error (MSE) as the loss function. The data set obtained by ILC can well provide training materials for our model, and the network trained with the MSE error has strong generality and can be directly applied to corresponding devices.
[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects: A digital predistortion method based on KAN and a new neural network provided by the present application extracts features of the time-delay 5G signal through the neural network KAN, and uses the features extracted by the neural network KAN and the original time-delay 5G signal as inputs, and uses the new neural network to perform nonlinear fitting on the output of the power amplifier. Finally, a fully connected layer is used to obtain the output after linear compensation. The logarithmic operation in this method can convert the multiplication between elements into a linear combination between logarithms. This operation helps to extract the features of the signal, enabling the subsequent network to more efficiently learn the features of the signal. The multi-round propagation process of the signal is consistent with the physical characteristic of signal attenuation over time, enabling the network to well handle the nonlinearity and memory effect of the power amplifier. The data set acquisition method and the model training method are simple and easy to use, and have strong generality and can be directly applied in practice. In conclusion, this method realizes high-precision, low computational complexity, and strong generality power amplifier nonlinear fitting. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can be obtained based on these drawings.
[0019] Figure 1 is a structural block diagram of a digital predistortion method based on KAN and a new neural network provided by an embodiment of the present application; Figure 2 is a Power Spectral Density diagram of the original data and the data processed by this method, where the solid line origin curve is the original data and the dashed line DPD curve is the data processed by this method. Detailed implementation manners
[0020] To enable those skilled in the art to better understand the solutions of the present application, the following will clearly and completely describe the technical solutions in the embodiments of the present application in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.
[0021] The mention of "embodiment" in the present application means that the specific features, structures, or characteristics described in conjunction with the embodiment may be included in at least one embodiment of the present application. The phrase appears in various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described in the present application can be combined with other embodiments.
[0022] Embodiment 1 This embodiment discloses a digital predistortion method based on KAN and a new neural network. The digital predistortion method includes the following steps: S1. Perform logarithmic operation on the time-delay 5G signal in complex form and then restore it to real form and input it into the Kolmogorov - Arnold network, hereinafter referred to as KAN for short.
[0023] Assume an input time-delay 5G signal is , where is a settable KAN memory depth parameter, represents the real part of the time-delay 5G signal at time t - m, represents the imaginary part of the time-delay 5G signal at time t - m, then the mathematical expression of this step is:
[0024] Among them, is the imaginary unit, is the result of performing a logarithmic operation on the input time-delay 5G signal, represents the real part of represents the imaginary part of represents the neural network KAN. In this step, the input signal undergoes a complex logarithmic operation. Through the preset memory depth parameter setting, a real-form signal is generated for subsequent processing by the neural network. The logarithmic operation in this operation can convert the multiplication between elements into a linear combination between logarithms, which helps to extract the features of the signal and enables the subsequent network to learn the features of the signal more efficiently.
[0025] S2. After splicing the output of KAN and the original time-delay 5G signal, input them into the new neural network in real form, and perform signal propagation for a certain number of rounds in the new neural network according to the parameter settings. The new neural network is a network with two types of nodes: input nodes and extended nodes. The number of input nodes is consistent with the dimension of the input data and the number of extended nodes is determined by the parameter . Each node has three state variables, and the state variables of all nodes are combined together to form a memory pool , a candidate pool and an output pool . The input nodes and extended nodes are numbered starting from 0 in sequence, that is, the nodes numbered from 0 to number are input nodes, and the nodes numbered to number are extended nodes. There are connections between each node, and the edge weights are randomly initialized, so there is an adjacency matrix , is the edge weight between node i and node j, . The mathematical expression of the operation mechanism of the new neural network is as follows:
[0026]
[0027]
[0028]
[0029] Among them, p is a settable propagation round parameter, is the output after passing through the new neural network for the p-th time, It is a new type of neural network with three pools: a memory pool , a candidate pool and an output pool . Before signal propagation, the three pools receive inputs according to the following formula :
[0030] where is a -dimensional zero vector, and is the number of expandable nodes that can be set. In each subsequent round of propagation, executes the following formula:
[0031]
[0032]
[0033]
[0034] where is the adjacency matrix of all nodes randomly initialized in the new type of neural network, is an intermediate variable for this round of propagation, and . Such a fully connected graph connection method ensures that each node can interact with other nodes for information; the design of the three pools is inspired by the way of electroencephalogram propagation in the human brain. The candidate pool accumulates potential until the activation threshold, and then the output pool generates an output. The memory pool accumulates the information of the output pool in multiple rounds of propagation; the signal propagation method of the class residual connection of the three pools enables step-by-step processing of signals, which is in line with the physical characteristics of signal attenuation over time; the above network design enables the network to well handle the nonlinearity and memory effect of the power amplifier.
[0035] S3. Extract the output of the new type of neural network and input it into the fully connected layer to generate a two-dimensional output, using the output of the ideal digital predistorter as a label and using MSE (mean square error) as the loss function for training. The mathematical expression is as follows:
[0036]
[0037] where is the fully connected layer with an output dimension of 2, and represents the signal after digital predistortion. The MSE loss function is used during training to optimize the network parameters, ultimately minimizing the error between the output signal and the ideal predistorted signal.
[0038] Example 2 Referring to steps S1 to S3 in a digital predistortion method based on KAN and a new neural network disclosed in Embodiment 1, this embodiment further discloses a digital predistortion method based on KAN and a new neural network, and the process is as follows: First, generate an ideal digital predistortion data set through DPD cycling or Iterative Learning Control (ILC), which contains the original input data X and the ideal DPD output data Y, both of which are one-dimensional complex data with a length of . Next, sample the data set according to a time window, and sample N - a - b - 1 sample pairs in the form of ([X t-b+1 , …, X t , X t+1 , …, X t+a , [Y t ), where a represents the size of the look-ahead time window, b represents the size of the look-behind time window, and after normalization, the training set and the test set are divided according to a certain ratio (6:4). In this embodiment, .
[0039] The input dimension and output dimension of the neural network KAN are set to 32, the grid scale is set to 5, and the spline order is set to 3; the number of input nodes of the new neural network is set to 64, the number of extended nodes is set to 100, and the number of propagation rounds is set to 3; the input dimension of the fully connected layer is set to 164, and the output dimension is set to 2.
[0040] Batch_Size is set to 1000, the learning rate is set to 0.01, the loss function uses MSE, the optimizer uses the Adam optimizer, and the adaptive learning rate policy is ReduceLROnPlateau, where factor = 0.5, threshold = 1e - 4, patience = 10, min_lr = 1e - 5, eps = 1e - 4, and the learning rate is adjusted every ten rounds with the training loss as the index. The number of training rounds is set to 10000.
[0041] Table 1. Comparison experiment results table with mainstream models
[0042] It can be seen from the comparison experiment results with mainstream models in Table 1 above that both VDTDNN and BOTDNN use second-order amplitude terms for enhancement. VDTDNN has 32 hidden neurons, BOTDNN has 50 linear filters, and the hidden state dimension of LSTM is 51 and there is only one layer. Therefore, the digital predistortion method based on KAN and a new neural network disclosed in this embodiment exceeds the performance of VDTDNN, BOTDNN, and LSTM, and realizes high-precision power amplifier nonlinear fitting.
[0043] Figure 2 It is the Power Spectral Density graph of the original data and the data processed by this method. Among them, the solid line origin curve is the original data, and the dotted line DPD curve is the data processed by this method. It can be seen from the figure that the out-of-band leakage of the data processed by this method has been significantly improved.
[0044] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that the present invention is not limited by the described action sequence, because according to the present invention, certain steps can be performed in other sequences or simultaneously.
[0045] The above embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.
Claims
1. A digital predistortion method based on KAN and a new neural network, characterized in that, The digital predistortion method includes the following steps: S1. Take the time-delay 5G signal in complex form, perform logarithmic operation, and then restore it to real form and input it into the Kolmogorov-Arnold network, hereinafter referred to as KAN for the Kolmogorov-Arnold network; S2. Concatenate the output of KAN and the original time-delay 5G signal, input it into the new neural network in real form, and perform signal propagation for a certain number of rounds in the new neural network according to parameter settings, hereinafter referred to as NNN for the new neural network; S3. Extract the output of NNN and input it into the fully connected layer to generate a two-dimensional output.
2. The digital predistortion method based on KAN and a new neural network according to claim 1, wherein In step S1, first take the time-delay 5G signal in complex form for logarithmic operation, and then restore it to real form and input it into KAN.
3. A digital predistortion method based on KAN and a new neural network according to claim 1, characterized in that, In step S2, concatenate the features extracted by the neural network KAN and the original time-delay 5G signal, and input it into the new neural network in real form.
4. A digital predistortion method based on KAN and a new neural network according to claim 1, characterized in that The new neural network has two types of nodes: input nodes and central nodes.
5. A digital predistortion method based on KAN and a new neural network according to claim 4, characterized in that, Each node in the new neural network has three state variables, and the state variables of all nodes are combined together to form a memory pool , a candidate pool and an output pool .
6. A digital predistortion method based on KAN and a new neural network according to claim 5, characterized in that The signal propagation method in the new neural network is ； Among them, represents the adjacency matrix of all nodes randomly initialized in the new neural network, represents a zero vector of dimension n, where n is the input node dimension, is the number of expandable nodes that can be set, represents the intermediate variable of this round of propagation, and min(, ) represents the operation of finding the minimum value, The expression of is as follows:
7. A digital predistortion method based on KAN and a new neural network according to claim 1, characterized in that, The output of the new neural network is the memory pool .
8. A digital predistortion method based on KAN and a new neural network according to claim 1, characterized in that In the step S3, the input dimension of the fully connected layer is the dimension of the memory pool of the new neural network, and the output is a two-dimensional output, expressed as , which are the real part and the imaginary part of the two-dimensional output of the fully connected layer, respectively.
9. A digital predistortion method based on KAN and a new neural network according to claim 1, characterized in that In the step S3, using the output of the ideal digital predistorter as a label, which are the real part and the imaginary part of the output of the ideal digital predistorter respectively, train the KAN and the new neural network with the mean square error MSE as the loss function.