Power amplifier predistortion method based on KAN neural network

Abstract

The invention relates to a power amplifier predistortion method based on a KAN neural network, and belongs to the technical field of radio frequency power amplifier linearization. The problems that a traditional polynomial model is insufficient in precision and a traditional neural network is large in parameter quantity are solved. The technical scheme comprises: collecting and processing input and output signals of a power amplifier; constructing a KAN neural network or a KAN model improved based on a memory polynomial; training the model; and pre-distortion signals are generated and input into the power amplifier. The method has the technical effects that the linearization precision of the power amplifier is remarkably improved, the model complexity is greatly reduced, the training time of the improved model is shortened, the industrial deployment universality is enhanced, and the green communication development is promoted.

Description

Technical Field

[0001] The present invention belongs to the technical field of radio frequency power amplifier linearization, and relates to a power amplifier predistortion method based on a KAN neural network. Background Art

[0002] As a core component of RF systems, power amplifiers (PAs) account for the vast majority of the total power consumption of the RF front-end. When processing high peak-to-average ratio signals, PAs generate significant nonlinear distortion, severely limiting communication system performance. To mitigate this distortion, pre-distortion technology has become a mainstream industry solution.

[0003] Existing predistortion models fall into two main categories. One type is polynomial models derived from Volterra series theory, including memory polynomial models, Hammerstein models, and Wiener models. These models have simple structures and are easy to implement in hardware. However, when the power amplifier's nonlinear characteristics are complex, their modeling accuracy is significantly insufficient, making it difficult to meet the requirements of high-linearity applications. The other type is neural network models, such as BP neural networks and RBF neural networks. While these models can achieve high fitting accuracy, they suffer from complex network structures, large number of parameters, and high computational resource consumption, which limits their real-time processing capabilities.

[0004] The core contradiction facing the current technology lies between the insufficient accuracy of polynomial models and the excessive complexity of neural network models. With the increasing demands for power amplifier linearity and energy efficiency in 5G and future communication systems, a new predistortion model with both high accuracy and low complexity is urgently needed. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a power amplifier predistortion method based on the KAN neural network to solve the problem that the modeling accuracy of the polynomial model is not high enough and the traditional neural network model is large in scale and has many parameters. By utilizing the advantages of the KAN neural network architecture, higher accuracy is provided while maintaining a smaller network scale, and the B-spline curve that constitutes the activation function in the KAN model can be replaced by a combination of basic terms of the memory polynomial to save the training time required for the fitting process.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A power amplifier predistortion method based on a KAN neural network comprises the following steps:

[0008] S1: Collect the input signal and output signal of the power amplifier as training signals;

[0009] S2: performing time domain processing and normalization processing on the training signal to generate a model input matrix and an output matrix;

[0010] S3: Construct a pre-distortion model, where the pre-distortion model is any one of the following:

[0011] (a) KAN neural network model;

[0012] (b) KAN neural network model improved based on memory polynomial;

[0013] S4: training the predistortion model using the processed training signal;

[0014] S5: The signal to be predistorted is input into the trained predistortion model after the same processing as S2, and the model output signal is denormalized to generate a predistorted signal;

[0015] S6: Input the predistortion signal into a power amplifier, so that the power amplifier outputs a linearized signal.

[0016] Further, the S2 includes:

[0017] S201: Collect n output signals x = [x(1), x(2), ..., x(n)] and n input signals y = [y(1), y(2), ..., y(n)] of the power amplifier as training signals;

[0018] S202: Process the training signal to obtain a real signal I in =[I in (1),I in (2),…I in (n)]、I out =[I out (1),I out (2),…I out (n)] and the imaginary signal Q in =[Q in (1),Q in (2),…Q in (n)], Q out =[Q out (1),Q out (2),…Q out (n)];

[0019] S203: Combining the real and imaginary parts of the input signal into an input signal matrix X, and combining the real and imaginary parts of the output signal into an output signal matrix Y, where

[0020]

[0021] S204: Take 1 to m tap delay items in columns of X as inputs representing memory effects, and reassemble them into X, such that:

[0022]

[0023] Where m represents the memory depth, 0≤m<n;

[0024] S205: Normalize X and Y to obtain X n and Y n , the normalization formula is as follows:

[0025]

[0026] Among them, max(·) means finding the maximum value in the matrix, and min(·) means finding the minimum value in the matrix.

[0027] Furthermore, in S3, the constructed pre-distortion model includes an input layer, a hidden layer and an output layer, each neuron in the input layer is connected to each neuron in the hidden layer, and each neuron in the hidden layer is also connected to each neuron in the output layer, and there is a trainable activation function between each interconnected neuron.

[0028] Furthermore, when the pre-distortion model is a KAN neural network model, its activation function is defined by the formula φ(x)=w(b(x)+spline(x)), where w is a weight coefficient; b(x) is a basis function, and spline(x) is a spline function, wherein:

[0029]

[0030] Among them B i (x) is a B-spline curve, and its formula is as follows:

[0031] B i (x) = a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3

[0032] where a i 、b i 、c i d i are all trainable parameters; x is the input variable, x i is the reference point in the B-spline;

[0033] Multiple activation functions φ p (x) is added to form the input of the hidden layer, and multiple activation functions after the hidden layer output are added to form the output, so that each output of the network can be described as:

[0034]

[0035] Among them, n1 and n2 represent the number of neurons in the input layer and hidden layer respectively.

[0036] Furthermore, when the pre-distortion model is a KAN neural network model improved based on memory polynomials, its activation function replaces the B-spline curve combination with the basic term combination of memory polynomials, and its formula is as follows:

[0037]

[0038] where Z k (x) is the memory polynomial term, a k is a trainable parameter, k is an index, K represents the maximum order of the memory polynomial; each activation function is expressed as Each output of the network is described as:

[0039]

[0040] Among them, w is the weight coefficient, b(x) is the silu function, n1 and n2 represent the number of neurons in the input layer and hidden layer respectively.

[0041] Furthermore, the output layer of the predistortion model includes two neurons, which are used to output the real part and imaginary part of the predistortion signal respectively, and the input layer includes 2(m+1) neurons, which are used to receive the real part and imaginary part of the input signal and their corresponding tap delay terms, where m is the memory depth.

[0042] Furthermore, in S4, the model is trained using the LBFGS algorithm.

[0043] Furthermore, in S5, the inverse normalization transformation formula is as follows:

[0044] Y1=min(Y)+(max(Y)-min(Y))Y0

[0045] Where Y0 is the output signal of the model, min(Y) is the minimum value of the training signal, max(Y) is the maximum value of the training signal, Y1 represents the signal input to the power amplifier after the inverse normalization transformation, and Y is the training signal matrix.

[0046] The beneficial effects of the present invention are:

[0047] (1) The KAN neural network architecture is used to replace the traditional polynomial model. The nonlinear characteristics of the power amplifier are accurately fitted through a learnable spline activation function, breaking through the accuracy bottleneck of the existing polynomial model in complex distortion scenarios and achieving high-fidelity signal amplification.

[0048] (2) By taking advantage of the parameter efficiency of the KAN network, the same or higher accuracy can be achieved with far fewer parameters than traditional neural networks, reducing hardware resource consumption and providing feasibility for deployment on embedded devices.

[0049] (3) Two optimization models are provided: the basic KAN model ensures the highest accuracy, and the improved KAN model based on memory polynomial significantly shortens the training time. Users can flexibly choose according to real-time requirements.

[0050] (4) The original delay term reorganization and normalization processing method effectively captures the memory effect of the power amplifier; combined with the anti-normalization output mechanism, it ensures that the pre-distortion signal matches the power amplifier input characteristics, thereby improving the robustness of the system.

[0051] (5) Reduce the computational load while maintaining high linearity, and reduce the overall energy consumption of the RF system, which is especially suitable for high-density deployment scenarios such as 5G base stations.

[0052] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0054] Figure 1 Flowchart of the KAN neural network predistortion method of the present invention;

[0055] Figure 2 This is a structural diagram of a KAN neural network model according to an embodiment of the present invention;

[0056] Figure 3 This is the overall structural diagram of the predistortion system corresponding to the present invention. DETAILED DESCRIPTION

[0057] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0058] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0059] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0060] like Figure 1 As shown, a power amplifier predistortion method based on a KAN neural network includes the following steps:

[0061] S1: Collect the output signal and input signal of the power amplifier as training signals and process the training signals;

[0062] S2: Construct a KAN neural network model or a KAN neural network model improved based on memory polynomials;

[0063] S3: Use the processed signal to train the constructed model;

[0064] S4: After the new input signal undergoes the same processing as the training signal, it is sent to the trained model for output. The output signal is then denormalized to obtain a predistorted signal, which is then input into the power amplifier. At this point, the output signal of the power amplifier is the linearized output signal.

[0065] In this embodiment, the S1 first collects n output signals x = [x(1), x(2), ..., x(n)] and n input signals y = [y(1), y(2), ..., y(n)] of the power amplifier as training signals; then processes the training signals to obtain the real signal I in =[I in (1),I in (2),…I in (n)]、I out =[I out (1),I out (2),…I out(n)] and the imaginary signal Q in =[Q in (1),Q in (2),…Q in (n)], Q out =[Q out (1),Q out (2),…Q out (n)]; combine the real and imaginary parts of the input signal into X, combine the real and imaginary parts of the output signal into Y, then take 2 tap delay items from X by column, indicating a memory depth of 2, and reassemble into X, then:

[0066]

[0067] Finally, X and Y are normalized to obtain X n and Y n , the normalization formula is as follows:

[0068]

[0069] Among them, max(·) means finding the maximum value in the matrix, and min(·) means finding the minimum value in the matrix.

[0070] The KAN neural network model constructed in S2 or the KAN neural network model improved based on memory polynomials comprises an input layer, a hidden layer, and an output layer. Each neuron in the input layer is connected to each neuron in the hidden layer, and each neuron in the hidden layer is also connected to each neuron in the output layer. There is a trainable activation function between each connected neuron. Figure 2 As shown, the KAN neural network model constructed in this embodiment includes five hidden layer neurons. The output layer includes two neurons, each for outputting the real and imaginary parts of the predistorted signal. The input layer includes six neurons, each for receiving the real and imaginary parts of the input signal and their corresponding tap delay terms.

[0071] Furthermore, the activation function in the KAN neural network model constructed in S2 is defined by the formula φ(x)=w(b(x)+spline(x)), where:

[0072]

[0073] Among them B i (x) is a B-spline curve, and its formula is as follows:

[0074] B i (x) = a i +b i (xx i )+c i (xxi ) 2 +d i (xx i ) 3

[0075] where a i 、b i 、c i d i These are all trainable parameters;

[0076] Multiple activation functions φ p (x) is added to form the input of the hidden layer, and multiple activation functions after the hidden layer output are added to form the output. In this embodiment, the number of input layer neurons is 6 and the number of hidden layer neurons is 5, so each output of the network can be described as:

[0077]

[0078] The KAN neural network model can also be improved based on the memory polynomial, replacing the B-spline curve combination that constitutes the activation function with the basic term combination of the memory polynomial. The formula is as follows:

[0079]

[0080] Where K represents the maximum order of the memory polynomial. In this embodiment, the maximum order K is set to 3, then

[0081] Each activation function can be expressed as Since the number of neurons is 6 and the number of hidden layer neurons is 5, each output of the network can be described as:

[0082] After the pre-distortion model is constructed, the model is trained using the LBFGS algorithm. The new input signal is then processed and input into the pre-distortion model. The model output signal is processed using the following inverse normalization transformation formula and then input into the power amplifier:

[0083] Y1=min(Y)+(max(Y)-min(Y))Y0

[0084] Where Y0 is the output signal of the model, min(Y) is the minimum value of the training signal, max(Y) is the maximum value of the training signal, Y1 represents the signal input to the power amplifier after the inverse normalization transformation, and the output of the power amplifier Y2 is the orthogonal modulation signal after linearization. The overall structure of the predistortion system is as follows Figure 3 shown.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A power amplifier predistortion method based on a KAN neural network, characterized by: The following steps are involved: S1: Collect the input signal and output signal of the power amplifier as training signals; S2: performing time domain processing and normalization processing on the training signal to generate a model input matrix and an output matrix; S3: Construct a pre-distortion model, where the pre-distortion model is any one of the following: (a) KAN neural network model; (b) KAN neural network model improved based on memory polynomial; S4: training the predistortion model using the processed training signal; S5: The signal to be predistorted is input into the trained predistortion model after the same processing as S2, and the model output signal is denormalized to generate a predistorted signal; S6: Input the predistortion signal into a power amplifier, so that the power amplifier outputs a linearized signal.

2. The power amplifier predistortion method based on the KAN neural network according to claim 1, characterized in that: The S2 includes: S201: Collect n output signals x = [x(1), x(2), ..., x(n)] and n input signals y = [y(1), y(2), ..., y(n)] of the power amplifier as training signals; S202: Process the training signal to obtain a real signal I in =[I in (1),I in (2),…I in (n)]、I out =[I out (1),I out (2),…I out (n)] and the imaginary signal Q in =[Q in (1),Q in (2),…Q in (n)], Q out =[Q out (1),Q out (2),…Q out (n)]; S203: Combining the real and imaginary parts of the input signal into an input signal matrix X, and combining the real and imaginary parts of the output signal into an output signal matrix Y, where S204: Take 1 to m tap delay items in columns of X as inputs representing memory effects, and reassemble them into X, such that: Where m represents the memory depth, 0≤m<n; S205: Normalize X and Y to obtain X n and Y n , the normalization formula is as follows: Among them, max(·) means finding the maximum value in the matrix, and min(·) means finding the minimum value in the matrix.

3. The power amplifier predistortion method based on the KAN neural network according to claim 1, characterized in that: In S3, the constructed pre-distortion model includes an input layer, a hidden layer and an output layer. Each neuron in the input layer is connected to each neuron in the hidden layer, and each neuron in the hidden layer is also connected to each neuron in the output layer. There is a trainable activation function between each interconnected neuron.

4. The power amplifier predistortion method based on the KAN neural network according to claim 1 or 3, characterized in that: When the predistortion model is a KAN neural network model, its activation function is defined by the formula φ(x)=w(b(x)+spline(x)), where w is a weight coefficient; b(x) is a basis function, and spline(x) is a spline function, wherein: Among them B i (x) is a B-spline curve, and its formula is as follows: B i (x)=a i +b i (x-x i )+c i (x-x i ) 2 +d i (x-x i ) 3 where a i 、b i 、c i d i are all trainable parameters; x is the input variable, x i is the reference point in the B-spline; Multiple activation functions φ p (x) is added to form the input of the hidden layer, and multiple activation functions after the hidden layer output are added to form the output, so that each output of the network can be described as: Among them, n1 and n2 represent the number of neurons in the input layer and hidden layer respectively.

5. The power amplifier predistortion method based on the KAN neural network according to claim 1 or 3, characterized in that: When the predistortion model is a KAN neural network model improved based on memory polynomials, its activation function replaces the B-spline curve combination with the basic term combination of memory polynomials, and its formula is as follows: where Z k (x) is the memory polynomial term, a k is a trainable parameter, k is an index, and K represents the maximum order of the memory polynomial; Each activation function is expressed as Each output of the network is described as: Among them, w is the weight coefficient, b(x) is the silu function, n1 and n2 represent the number of neurons in the input layer and hidden layer respectively.

6. The power amplifier predistortion method based on the KAN neural network according to claim 3, characterized in that: The output layer of the predistortion model includes two neurons, which are used to output the real part and imaginary part of the predistortion signal respectively. The input layer includes 2(m+1) neurons, which are used to receive the real part and imaginary part of the input signal and their corresponding tap delay terms, where m is the memory depth.

7. The power amplifier predistortion method based on the KAN neural network according to claim 1, characterized in that: In S4, the model is trained using the LBFGS algorithm.

8. The power amplifier predistortion method based on the KAN neural network according to claim 1, characterized in that: In S5, the anti-normalization transformation formula is as follows: Y1=min(Y)+(max(Y)-min(Y))Y0 Where Y0 is the output signal of the model, min(Y) is the minimum value of the training signal, max(Y) is the maximum value of the training signal, Y1 represents the signal input to the power amplifier after denormalization, and Y is the training signal matrix.

Images