Power amplifier linearization method based on KAN network and digital pre-distortion system

Abstract

The invention discloses a power amplifier linearization method based on a KAN network and a digital pre-distortion system, and belongs to the technical field of wireless communication, and the method comprises the steps: introducing Kolmogorov-Arnold Networks (KAN) in a digital pre-distortion processing process, and constructing a real value time delay KAN digital pre-distortion network and a real value time delay KAN digital pre-distortion network; for a real-value time delay KAN digital pre-distortion network, input signals at the current moment and the previous moment are considered in the input process of the network, and the memorability of a power amplifier can be correspondingly compensated; for a vector decomposition time delay KAN digital pre-distortion network, on the basis of a real value time delay KAN digital pre-distortion network, a phase recovery layer and a full connection layer are further introduced, and under the condition that the performance similar to that of a previous KAN model is kept, the complexity of the network can be simplified; compared with an original digital pre-distortion method based on a neural network, the KAN digital pre-distortion method used in the invention uses fewer network parameters and realizes higher linearization capability.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication, and in particular relates to a power amplifier linearization method and a digital predistortion system based on a KAN network, which are applied to a radio frequency link at a transmitting end of a wireless communication system. Background Art

[0002] In the transmission system of wireless communication, the memory and nonlinearity of the power amplifier are the main reasons for the distortion of the transmitted signal. In order to prevent the amplified signal from being distorted, the power amplifier can be limited to operate in the linear region, but this will greatly reduce the energy efficiency of the power amplifier. Currently, the most effective linearization technology is digital pre-distortion technology.

[0003] Digital pre-distortion technology is to pre-distort the digital signal first. The pre-distorted digital signal enters the power amplifier after digital-to-analog conversion. After the nonlinear distortion of the power amplifier, the output signal of the power amplifier has only linear gain compared to the original input signal. In other words, the entire system of the digital pre-distorter and the power amplifier in cascade is a linear system. It is only necessary to construct the system function of the digital pre-distortion as the inverse function of the system function of the power amplifier to realize that the system after the two are cascaded is a linear system and to achieve compensation for the nonlinear distortion of the power amplifier.

[0004] Traditional digital pre-distortion methods require mathematical modeling of the behavioral characteristics of the power amplifier. Common modeling methods include polynomial models, Volterra series models, memory polynomial (MP) models, multi-box models, piecewise linear function models, etc. Then, an inverse model is constructed based on the mathematical modeling formula as the system function of digital pre-distortion. However, as the design of power amplifiers becomes more and more complex, their behavioral characteristics are becoming increasingly difficult to characterize well with mathematical models. With the development of deep learning, researchers have discovered that deep neural networks have powerful fitting capabilities and can better characterize various complex systems. Therefore, in recent years, many methods have emerged for modeling power amplifiers using deep neural networks and implementing digital pre-distortion.

[0005] Although neural networks have stronger nonlinear fitting capabilities than traditional mathematical modeling methods, and the linearization effect is better when applied to digital pre-distortion, there are still the following challenges before practical application: First, the scale of neural network coefficients is much higher than that of mathematical models, and the structure is complex. Engineering implementation will consume a lot of hardware resources, so a low-complexity digital pre-distortion neural network is needed. Secondly, the behavioral characteristics of the power amplifier will change according to the signal waveform, bandwidth, frequency, power, and temperature. In order to solve this problem, the existing methods are to store the digital pre-distortion coefficients corresponding to all nonlinear states, load different coefficients for different states, and recalculate the corresponding digital pre-distortion coefficients when the power amplifier state changes. However, the above two methods either take up a lot of storage space or consume a lot of computing resources. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the above-mentioned prior art and provide a power amplifier linearization method and a digital predistortion system based on a KAN network to solve the problem of large storage space occupied or consumption of computing resources in the digital predistortion processing process in the prior art. In order to achieve the above object, the present invention adopts the following technical solutions: A power amplifier linearization method based on a KAN network comprises the following steps: S1, obtaining an in-phase component and a quadrature component of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment minus a delay from the current moment; S2, processing the input signal through a real-valued time-delay KAN digital pre-distortion network, wherein the real-valued time-delay KAN digital pre-distortion network includes several layers of kanlayer, the input of the next layer is the output of the previous layer, and the output of each layer is the product of the input signal of the layer and the activation function of the layer; the kanlayer is used to compensate for the nonlinear distortion caused by the power amplifier; S3, outputs the in-phase component and the orthogonal component of the predistortion processed signal.

[0007] A further improvement of the present invention is: Preferably, in S1, the length of the time delay is the memory depth of the power amplifier.

[0008] A power amplifier linearization method based on a KAN network comprises the following steps: S1, obtaining the amplitude and phase of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment obtained by subtracting a delay from the current moment; S2, processing the input signal through a vector decomposition delay KAN digital pre-distortion network, wherein the vector decomposition delay KAN digital pre-distortion network includes a kanlayer, a phase recovery layer and a fully connected layer, wherein the phase recovery layer includes a plurality of cosPRBs and a plurality of sinPRBs, the kanlayer layer is used to compensate for the nonlinear distortion caused by the power amplifier, the phase recovery layer is used to recover the amplitude and phase into an in-phase component and an orthogonal component, and the fully connected layer is used to output the in-phase component and the orthogonal component; S3, outputs the in-phase component and the orthogonal component of the predistortion processed signal.

[0009] Preferably, in S2, the calculation formula in the cos PRB is:

[0010] Where M is the memory depth of the power amplifier, It is the Lth layer. Node, located at Phase recovery group; The calculation formula in the sinPRB is:

[0011] in, It is the Lth layer. Node, located at Phase recovery group.

[0012] Preferably, in S1, the calculation formula of the phase recovery layer is:

[0013] in, is the phase recovery vector.

[0014] Preferably, in S1, the size determination process of the real-valued delay KAN digital pre-distortion network or the vector decomposition delay KAN digital pre-distortion network is: firstly establish a dense network, and then prune the dense network to obtain a corresponding pre-distortion network.

[0015] Preferably, the real-valued delay KAN digital pre-distortion network or the vector decomposition delay KAN digital pre-distortion network is trained by a training set and a validation set, and verified by a test set; the input samples of the test set and the validation set are the output signals of the power amplifier, and the target samples are the input signals of the power amplifier; the input samples of the test set are the input signals of the power amplifier, and the target samples are the output signals of the power amplifier; During the training process, NMSE is used as the loss function. After the loss function converges, the training is completed.

[0016] Preferably, in S3, the in-phase component and the orthogonal component after predistortion processing are input to the power amplifier, and if a new state appears in the power amplifier, the coefficients of the real-valued delay KAN digital predistortion network or the vector decomposition delay KAN digital predistortion network are updated, including the following steps: S401, collecting the input signal and output signal of the power amplifier, and recording the data set obtained when the power amplifier appears in a new state for the Nth time as .

[0017] S402: The first N-1 random sampling data With new state data Merge to get a complete data set : S403, yes Random sampling The data as , covering the storage area ; S403, using the full dataset The real-value delay KAN digital pre-distortion network or the vector decomposition delay KAN digital pre-distortion network is trained, and the corresponding distortion network is updated after the training is completed.

[0018] A digital predistortion device based on KAN, comprising: An input module, used to obtain an in-phase component and an orthogonal component of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment when a delay is subtracted from the current moment; A digital pre-distortion module, used for processing an input signal through a real-valued time-delay KAN digital pre-distortion network, wherein the real-valued time-delay KAN digital pre-distortion network comprises several layers of kanlayer, the input of the next layer is the output of the previous layer, and the output of each layer is the product of the input signal of the layer and the activation function of the layer; the kanlayer is used to compensate for the nonlinear distortion caused by the power amplifier; The output module is used to output the in-phase component and the orthogonal component of the predistortion processed signal.

[0019] A digital predistortion device based on KAN, comprising: An input module, used to obtain the amplitude and phase of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment obtained by subtracting a delay from the current moment; A digital pre-distortion module, used for processing an input signal through a vector decomposition delay KAN digital pre-distortion network, wherein the vector decomposition delay KAN digital pre-distortion network includes a kanlayer, a phase recovery layer and a fully connected layer, wherein the phase recovery layer includes a plurality of cos PRBs and a plurality of sin PRBs, wherein the kanlayer layer is used for compensating for the nonlinear distortion caused by the power amplifier, wherein the phase recovery layer is used for recovering the amplitude and the phase into an in-phase component and an orthogonal component, and wherein the fully connected layer is used for outputting the in-phase component and the orthogonal component; The output module is used to output the in-phase component and the orthogonal component of the predistortion processed signal.

[0020] Compared with the prior art, the present invention has the following beneficial effects: The invention discloses a power amplifier linearization method based on a KAN network. In the process of digital predistortion processing, the method introduces Kolmogorov–Arnold Networks (KAN) to construct a real-valued delay KAN digital predistortion network and a real-valued delay KAN digital predistortion network; for the real-valued delay KAN digital predistortion network, the input process of the network considers the input signals at the current moment and the previous moment, and can compensate the memory of the power amplifier accordingly; for the vector decomposition delay KAN digital predistortion network, on the basis of the real-valued delay KAN digital predistortion network, a phase recovery layer and a fully connected layer are further introduced, and the complexity of the network can be simplified while maintaining similar performance to the previous KAN model; the KAN digital predistortion method used in the invention uses fewer network parameters and achieves higher linearization capability compared to the original digital predistortion method based on a neural network. At the same time, in view of the problem that the behavior characteristics of the power amplifier will change dynamically with changes in external conditions, a method for training and updating digital predistorter coefficients is provided based on the memory of KAN. Verification found that compared with the traditional fully connected neural network, a smaller network size can be used to achieve better fitting effects, and compared with the catastrophic forgetting of traditional neural networks, it also has better memory.

[0021] Furthermore, based on the memory of KAN, the present invention provides a method for training and updating digital predistorter coefficients. When a power amplifier has new behavior characteristics, the digital predistorter coefficients are trained and updated. The updated digital predistorter can not only compensate for the new nonlinear state well, but also retain the memory of the previous state. When the nonlinear state of the power amplifier changes to a state that has appeared before, the predistorter coefficients do not need to be updated and can be directly used for compensation. Thus, the nonlinearity of the power amplifier can be quickly compensated and computing resources or storage space can be saved. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1A digital pre-distortion architecture and training method based on KAN; Figure 2 is the network structure of the real-valued time-delay KAN digital predistorter; Figure 3 The network structure of the vector decomposition delay KAN digital predistorter; Figure 4 The structural block diagram of the digital predistorter performance evaluation method is shown in FIG. Figure 5 The power spectrum comparison between KAN-based digital predistorter and traditional digital predistorter is shown in Figure 2. Figure 6 Dynamically update the structure block diagram of the digital pre-distorter based on KAN. DETAILED DESCRIPTION

[0023] In the following, the terms "first", "second", "third", and "fourth" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, a feature defined as "first", "second", "third", and "fourth" may explicitly or implicitly include one or more of the features.

[0024] The co-shooting method provided in the embodiment of the present application can be applied to terminal devices such as mobile phones, tablet computers, wearable devices, vehicle-mounted devices, augmented reality (AR) / virtual reality (VR) devices, laptop computers, ultra-mobile personal computers (UMPC), netbooks, personal digital assistants (PDA), etc. The embodiment of the present application does not impose any restrictions on the specific type of the terminal device.

[0025] It should be noted that the terms "first", "second", etc. in the specification and drawings of the present invention are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] Example 1 This embodiment discloses a power amplifier linearization method based on a KAN network, comprising the following steps: S1, obtaining an in-phase component and a quadrature component of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment minus a delay from the current moment; S2, processing the input signal through a real-valued time-delay KAN digital pre-distortion network, wherein the real-valued time-delay KAN digital pre-distortion network includes several layers of kanlayer, the input of the next layer is the output of the previous layer, and the output of each layer is the product of the input signal of the layer and the activation function of the layer; each layer of kanlayer is used to compensate for the nonlinear distortion caused by the power amplifier; S3, outputs the in-phase component and the orthogonal component of the predistortion processed signal.

[0027] In S2, in the real-valued time delay KANKAN (RVTDKAN for short) digital pre-distortion network, the input dimension of the network is determined by the memory depth of the power amplifier, and the output dimension is 2, representing the output of I and Q signals.

[0028] The network input is the current I and Q data. , where I and Q data represent the same direction component and the orthogonal component respectively, and I and Q data at the past moment , the data delay length M is equal to the power amplifier memory depth, so the input dimension of the network is 2(M+1). The output of the network is I and Q data. , the relationship between network input and output can be expressed as:

[0029] RVTDKAN is a multi-layered structure, and each layer has the same structure. l Layer input nodes, there are output node, so l Layer Total activation function, It is l Scalar input to the i-th node of the layer To l +1st floor j Node input scalar The activation function is l Layer i Node input scalar To l +1st floor j Node input scalar It can be expressed as:

[0030] So the l Layer input vector To l Layer output vector It can be expressed as:

[0031] The above formula is defined as the representation of a layer of kanlayer, so the function expression of L-layer RVTDKAN is:

[0032] Among them, the network activation function is a spline curve, expressed as:

[0033] in, is the control point of the spline curve, which is also the trainable parameter of the network. The number of control points is , Spline k The order basis functions, k The order basis functions are calculated using the following iterative algorithm:

[0034] in, is the i-th node of the spline curve, and the number of nodes is + k +1, initialized to uniform sampling in the interval [-1,1].

[0035] Example 2 S1, obtaining the amplitude and phase of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment obtained by subtracting a delay from the current moment; S2, processing the input signal through a vector decomposition delay KAN digital pre-distortion network, wherein the vector decomposition delay KAN digital pre-distortion network includes a kanlayer, a phase recovery layer and a fully connected layer, wherein the phase recovery layer includes a plurality of cosPRBs and a plurality of sinPRBs, and the phase recovery layer is used to recover the amplitude and phase into an in-phase component and an orthogonal component; S3, outputs the in-phase component and the orthogonal component of the predistortion processed signal.

[0036] In this embodiment, the specific structure of the vector decomposition time delay KAN (VDTDKAN for short) digital pre-distortion network in S2 is as follows: 1) The network input is the amplitude and phase at the current moment , as well as the amplitude and phase at past moments , the data delay length M is equal to the power amplifier memory depth. The output of the network is IQ two-way data , the relationship between network input and output can be expressed as:

[0037] 2) The first layer input of the network is the amplitude of the current and past moments. Assuming that VDTDKAN has L layers of kanlayers to compensate for the distortion caused by the nonlinearity of the power amplifier, the first L layers of VDTDKAN are expressed as:

[0038] in .

[0039] 3) The L-layer kanlayer is connected to the phase recovery layer, which is used to recover the phase. The input of the phase recovery layer is the phase of the current and past moments and the output of the L-th layer of VDTDKAN. The phase recovery layer consists of multiple phase recovery modules, each of which inputs a phase value corresponding to an output node value of the L-th layer. The phase recovery module (phase recovery block, abbreviated as PRB) is divided into two types of cos PRB and sin PRB of equal quantity. M+1 cos PRBs and M+1 sin PRBs are defined as a phase recovery group, and the phase recovery layer has P phase recovery groups.

[0040] like No. The node is located at phase recovery group, the following operation is performed in the corresponding cos PRB, where The output of cos PRB is:

[0041] like No. The node is located at phase recovery group, the following operation is performed in the corresponding sin PRB, where The output of sin PRB is:

[0042] The entire phase recovery layer can be expressed by the following relationship:

[0043] The phase recovery vector It is formed by merging P basic phase recovery vectors:

[0044] The basic phase recovery vector is expressed as:

[0045] 3) The phase recovery layer is followed by a fully connected layer, which is used to divide the network output into I and Q forms. It is the last layer of VDTDKAN and is expressed as:

[0046] in, is the weight of the fully connected layer, is the bias of the fully connected layer.

[0047] Example 3 A digital predistortion device based on KAN, comprising: An input module, used to obtain an in-phase component and an orthogonal component of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment when a delay is subtracted from the current moment; A digital pre-distortion module, used for processing an input signal through a real-valued time-delay KAN digital pre-distortion network, wherein the real-valued time-delay KAN digital pre-distortion network comprises several layers of kanlayer, the input of the next layer is the output of the previous layer, and the output of each layer is the product of the input signal of the layer and the activation function of the layer; the kanlayer is used to compensate for the nonlinear distortion caused by the power amplifier; The output module is used to output the in-phase component and the orthogonal component of the predistortion processed signal.

[0048] Example 4 A digital predistortion device based on KAN, comprising: An input module, used to obtain the amplitude and phase of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment obtained by subtracting a delay from the current moment; A digital pre-distortion module, used for processing an input signal through a vector decomposition delay KAN digital pre-distortion network, wherein the vector decomposition delay KAN digital pre-distortion network includes a kanlayer, a phase recovery layer and a fully connected layer, wherein the phase recovery layer includes a plurality of cos PRBs and a plurality of sin PRBs, wherein the kanlayer layer is used for compensating for the nonlinear distortion caused by the power amplifier, wherein the phase recovery layer is used for recovering the amplitude and the phase into an in-phase component and an orthogonal component, and wherein the fully connected layer is used for outputting the in-phase component and the orthogonal component; The output module is used to output the in-phase component and the orthogonal component of the predistortion processed signal.

[0049] Example 5 In this embodiment, the training acquisition process of the above two RVTDKAN and VDTDKAN is disclosed.

[0050] S201, collecting the input signal and output signal of the power amplifier as a set of data sets, normalizing and time-synchronizing the data, and dividing the data into a training set, a validation set, and a test set in a ratio of 8:1:1. The input samples in the training set and the validation set are the output signals of the power amplifier, and the target samples are the input signals of the power amplifier. The input samples in the test set are the input signals of the power amplifier, and the target samples are the output signals of the power amplifier.

[0051] S202, build a real-valued time delay KAN (Real-Valued Time Delay KAN abbreviated as RVTDKAN) digital pre-distortion network or a vector decomposition time delay KAN (Vector Decomposition Time Delay KAN abbreviated as VDTDKAN) digital pre-distortion network, wherein the input dimensions of the two networks are determined by the memory depth of the power amplifier, and the output dimensions are both 2, representing the output of two IQ signals.

[0052] S203, determining the optimal size of RVTDKAN and VDTDKAN. First, a dense network is established in which each hidden layer has a sufficiently wide width, and then a sparse network is automatically obtained by training and pruning, and the sparse network is used as the final size of the digital predistorter.

[0053] S204, using the training set in step 1 to train the network with the size determined in step 4. NMSE is used as the loss function, and the training is completed when the loss function converges.

[0054] S205, using the trained network as a digital pre-distorter to compensate for the nonlinearity of the power amplifier, and measuring three pre-distortion effect indicators, namely, normalized minimum mean square error (NMSE), tie leakage ratio (ACPR), and error vector magnitude (EVM).

[0055] S206, a new nonlinear state will appear in the power amplifier during use, so it is necessary to update the coefficients in the digital predistorter to compensate for the new nonlinear state. First, the power amplifier state detection is used to determine whether the power amplifier has a new state. If no new state appears, the digital predistorter coefficients are not updated. If a new state appears, the digital predistorter coefficients are updated. Since KAN has memory, after updating the coefficients of KAN, not only the new nonlinear state can be compensated, but also the nonlinear state that has appeared before can be compensated, avoiding the need to update the digital predistorter coefficients when the power amplifier has a previous nonlinear state. At the same time, in order to further improve the memory, the data backtracking method is adopted during training, that is, part of the data of the previous state data set is added to the data set of the current state.

[0056] In S206, the method for training and updating the digital predistorter coefficients when the power amplifier has a new state is as follows: 1) Collect the input and output signals of the power amplifier in this state, process the data in the same way as step 1 and obtain a data set. The data set obtained when the power amplifier appears in a new state for the Nth time is recorded as .

[0057] 2) Use data backtracking to build a complete data set , that is, the previously stored data of the first N-1 random samples With new state data Merge to get a complete data set :

[0058]

[0059] in Represents a randomly sampled dataset D data.

[0060] 3) Yes Random Sampling The data as , and cover the storage area As the traceback data when the power amplifier appears in a new state for the N+1th time.

[0061] 4) Use the full dataset Train the KAN-based digital pre-distortion network, and update the coefficients of the digital pre-distorter after training.

[0062] The following is further described in conjunction with specific embodiments.

[0063] Example 1, digital pre-distortion architecture based on KAN Figure 1As shown in Figure 1, the position of digital predistortion in the wireless communication transmission chain is as follows: Figure 1 As shown in the figure, the signal enters the digital predistorter after baseband modulation and OFDM modulation. The predistorted signal compensates for the nonlinear characteristics of the power amplifier after passing through the power amplifier, ensuring that the output signal of the power amplifier will not be significantly distorted. The training method is also as follows Figure 1 As shown, an indirect learning method is adopted to collect the output of the power amplifier as the input of the learning inverse model. The loss function during training is the normalized mean square error between the output of the inverse model and the input of the power amplifier. When the loss function converges, the training is completed, and then the coefficients of the inverse model are copied to the digital predistorter.

[0064] In this embodiment, 16QAM modulation is adopted, the number of subcarriers is 4096, the cyclic prefix length is 288, the signal bandwidth is 100 MHz, the oversampling rate is 10, and the power amplifier is a Doherty-type power amplifier. Taking the transmission link as an example, the technical solution is clearly and completely described.

[0065] Step 1: collect the input signal and output signal of the power amplifier as a set of data sets, normalize the data and synchronize the time, and then divide the data into a training set, a validation set and a test set in a ratio of 8:1:1. The input samples in the training set and the validation set are the output signals of the power amplifier, and the target samples are the input signals of the power amplifier. The input samples in the test set are the input signals of the power amplifier, and the target samples are the output signals of the power amplifier.

[0066] Step 2: Build a real-valued time delay KAN (RVTDKAN for short) digital pre-distortion network and a vector decomposition time delay KAN (VDTDKAN for short) digital pre-distortion network.

[0067] Among them, the network structure of RVTDKAN is as follows Figure 2 As shown, since the memory depth of the power amplifier is 4, the input dimension of RVTDKAN is 10 and the output dimension is 2. The specific structure is expressed as follows: 1) Network input is the current I and Q data , where I and Q data represent the same direction component and the orthogonal component respectively, and I and Q data at the past moment , the data delay length M is equal to the power amplifier memory depth, so the input dimension of the network is 2(M+1). The output of the network is I and Q data. , the relationship between network input and output can be expressed as:

[0068] 2) RVTDKAN is a multi-layered structure, and each layer has the same structure. l Layer input nodes, there are output node, so l Layer Total activation function, It is l Scalar input to the i-th node of the layer To l +1st floor j Node input scalar The activation function is l Layer i Node input scalar To l +1st floor j Node input scalar It can be expressed as:

[0069] So the l Layer input vector To l Layer output vector It can be expressed as:

[0070] The above formula is defined as the representation of a layer of kanlayer, so the function expression of L-layer RVTDKAN is: 3) The network activation function is a spline curve, expressed as:

[0072] in, is the control point of the spline curve, which is also the trainable parameter of the network. The number of control points is , Spline k The order basis functions, k The order basis functions are calculated using the following iterative algorithm:

[0073] in, is the i-th node of the spline curve, and the number of nodes is + k +1, initialized to uniform sampling in the interval [-1,1].

[0074] The network structure of VDTDKAN is as follows: Figure 3 As shown, since the memory depth of the power amplifier is 4, the input dimension of VDTDKAN is 5 and the output dimension is 2. The specific structure is expressed as follows: 1) The network input is the amplitude and phase at the current moment , as well as the amplitude and phase at past moments , the data delay length M is equal to the power amplifier memory depth. The output of the network is IQ two-way data , the relationship between network input and output can be expressed as:

[0075] 2) The first layer input of the network is the amplitude of the current and past moments. Assuming that VDTDKAN has L layers of kanlayers to compensate for the distortion caused by the nonlinearity of the power amplifier, the first L layers of VDTDKAN are expressed as:

[0076] in .

[0077] 3) The L-layer kanlayer is connected to the phase recovery layer, which is used to recover the phase. The input of the phase recovery layer is the phase of the current and past moments and the output of the L-th layer of VDTDKAN. The phase recovery layer consists of multiple phase recovery modules, each of which inputs a phase value corresponding to an output node value of the L-th layer. The phase recovery module (phase recovery block, abbreviated as PRB) is divided into two types of cos PRB and sin PRB of equal quantity. M+1 cos PRBs and M+1 sin PRBs are defined as a phase recovery group, and the phase recovery layer has P phase recovery groups.

[0078] like No. The node is located at phase recovery group, the following operation is performed in the corresponding cos PRB, where The output of cos PRB is:

[0079] like No. The node is located at phase recovery group, the following operation is performed in the corresponding sin PRB, where The output of sin PRB is:

[0080] The entire phase recovery layer can be expressed by the following relationship:

[0081] The phase recovery vector It is formed by merging P basic phase recovery vectors:

[0082] The basic phase recovery vector is expressed as:

[0083] 3) The phase recovery layer is followed by a fully connected layer, which is used to divide the network output into I and Q forms. It is the last layer of VDTDKAN and is expressed as:

[0084] in, is the weight of the fully connected layer, is the bias of the fully connected layer.

[0085] Step 3, determine the optimal size of RVTDKAN and VDTDKAN. First, establish a dense network with each hidden layer wide enough, then automatically obtain a sparse network through training and pruning, and use the sparse network as the final size of the digital pre-distorter. For RVTDKAN, take the initial network size as (10,10,10,2) as an example. After pruning, the RVTDKAN network size is (10,6,5,2). For VDTDKAN, it is necessary to set the number of phase recovery groups p first. Taking p=1 as an example, the initial network size is set to (5,10,10), (10,2). After pruning, the VDTDKAN network size is (5,4,10), (10,2).

[0086] Step 4: Use the training set in step 1 to train the network with the size determined in step 4. Take NMSE as the loss function. The training is completed when the loss function converges. For example, when the LBFGS optimizer is used, the loss function converges after 100 iterations, and the training is considered completed.

[0087] Step 5: Use the trained network as a digital predistorter to compensate for the nonlinearity of the power amplifier. Calculate the normalized minimum mean square error (NMSE), ACPR, and EVM. The calculation method of each performance indicator is as follows: Figure 4Table 1 shows the comparison of the parameter scale and performance index of the two digital pre-distorters based on KAN proposed in the present invention and the digital pre-distorters without digital pre-distortion and three existing digital pre-distorters. It can be seen from Table 1 that compared with the traditional digital pre-distorters based on neural networks, the two digital pre-distorters based on KAN proposed in the present invention use fewer network parameters and achieve better linearization effect. At the same time, compared with RVTDKAN, VDTDKAN uses a smaller network scale and achieves a linearization effect similar to RVTDKAN.

[0088] Table 1 Coefficient scale and performance comparison between KAN-based digital predistorter and traditional digital predistorter

[0089] Figure 5 The power spectrum of the power amplifier output signal is shown. Because the nonlinear behavior of the power amplifier will cause the signal power spectrum to have skirts on both sides of the bandwidth, the lower the skirt power value around the bandwidth of the digital predistorter, the better the performance. Figure 5 It can be seen intuitively that the two KAN-based digital predistorters proposed in the present invention have the lowest skirt power values ​​around the bandwidth.

[0090] Step 6: During the use of the power amplifier, new nonlinear states will appear, so the coefficients in the digital predistorter need to be updated to compensate for the new nonlinear states. The structure diagram of the digital predistorter based on KAN is shown in the figure. Figure 6 As shown. First, the power amplifier state detection is used to determine whether the power amplifier has a new state. If no new state appears, the digital pre-distortion coefficient is not updated. If a new state appears, the digital pre-distortion coefficient is updated. Since KAN has memory, after updating the coefficient of KAN, not only the new nonlinear state can be compensated, but also the nonlinear state that has appeared before can be compensated, avoiding the need to update the digital pre-distortion coefficient when the power amplifier has a previous nonlinear state. At the same time, in order to further improve the memory, the data backtracking method is used during training, that is, adding part of the data of the previous state data set to the current state data set.

[0091] Among them, the method of training and updating the digital predistorter coefficients when the power amplifier appears in a new state is as follows: 1) Collect the input and output signals of the power amplifier in this state, process the data in the same way as step 1 and obtain a data set. The data set obtained when the power amplifier appears in a new state for the Nth time is recorded as .

[0092] 2) Use data backtracking to build a complete data set , that is, the previously stored data of the first N-1 random samples With new state data Merge to get a complete data set :

[0093]

[0094] in, Represents a randomly sampled dataset D The data is actually taken .

[0095] 3) Yes Random Sampling The data as , and cover the storage area As the traceback data when the power amplifier appears in a new state for the N+1th time.

[0096] 4) Use the full dataset Train the KAN-based digital pre-distortion network, and update the coefficients of the digital pre-distorter after training.

[0097] Table 2 compares the memory performance of the digital predistorter based on KAN and the original digital predistorter based on a fully connected network. Since the power amplifier has different nonlinear characteristics under different signal bandwidths, the nonlinear characteristics of the power amplifier when the signal bandwidth is 100MHz, 40MHz, and 15MHz are taken as states 1, 2, and 3. When the power amplifier sequentially experiences states 1, 2, and 3, the coefficients of the above two digital predistorters are trained and updated respectively according to step 6. After reaching state 3 and updating the coefficients three times, the performance indicators of the two digital predistorters in the three states are calculated using the latest updated coefficients. It can be seen from Table 2 that the enhanced real-valued time-delay neural network digital predistorter, that is, the digital predistorter based on the fully connected network, has good linearization capability only in state 3, and no linearization capability in states 1 and 2, that is, the network has no memory, and the network coefficients need to be updated again to compensate for the power amplifier in state 1 or 2. Correspondingly, the digital predistorter based on KAN has good linearization capability in all three states, which verifies its memory performance.

[0098] Table 2 Comparison of memory between KAN-based digital predistorter and traditional digital predistorter.

[0099]

[0100] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A power amplifier linearization method based on a KAN network, characterized in that: The following steps are involved: S1, obtaining an in-phase component and a quadrature component of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment minus a delay from the current moment; S2, processing the input signal through a real-valued time-delay KAN digital pre-distortion network, wherein the real-valued time-delay KAN digital pre-distortion network includes several layers of kanlayer, the input of the next layer is the output of the previous layer, and the output of each layer is the product of the input signal of the layer and the activation function of the layer; the kanlayer is used to compensate for the nonlinear distortion caused by the power amplifier; S3, outputs the in-phase component and the orthogonal component of the predistortion processed signal.

2. A power amplifier linearization method based on a KAN network according to claim 1, characterized in that: In S1, the length of the time delay is the memory depth of the power amplifier.

3. A power amplifier linearization method based on a KAN network, characterized in that: The following steps are involved: S1, obtaining the amplitude and phase of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment obtained by subtracting a delay from the current moment; S2, processing the input signal through a vector decomposition delay KAN digital pre-distortion network, wherein the vector decomposition delay KAN digital pre-distortion network includes a kanlayer, a phase recovery layer and a fully connected layer, wherein the phase recovery layer includes a plurality of cos PRBs and a plurality of sin PRBs, the kanlayer layer is used to compensate for the nonlinear distortion caused by the power amplifier, the phase recovery layer is used to recover the amplitude and phase into an in-phase component and an orthogonal component, and the fully connected layer is used to output the in-phase component and the orthogonal component; S3, outputs the in-phase component and the orthogonal component of the predistortion processed signal.

4. A power amplifier linearization method based on a KAN network according to claim 3, characterized in that: In S2, the calculation formula in the cos PRB is: Where M is the memory depth of the power amplifier, It is the Lth layer. Node, located at Phase recovery group; The calculation formula in the sinPRB is: in, It is the Lth layer. Node, located at Phase recovery group.

5. The power amplifier linearization method based on KAN network according to claim 3, characterized in that: In S1, the calculation formula of the phase recovery layer is: in, is the phase recovery vector.

6. A power amplifier linearization method based on a KAN network according to claim 1 or 3, characterized in that: In S1, the size determination process of the real-valued time-delay KAN digital pre-distortion network or the vector decomposition time-delay KAN digital pre-distortion network is as follows: firstly, a dense network is established, and then the dense network is pruned to obtain a corresponding pre-distortion network.

7. A power amplifier linearization method based on a KAN network according to claim 1 or 3, characterized in that: The real-valued delay KAN digital pre-distortion network or the vector decomposition delay KAN digital pre-distortion network is trained by a training set and a validation set, and verified by a test set; the input samples of the test set and the validation set are the output signals of the power amplifier, and the target samples are the input signals of the power amplifier; the input samples of the test set are the input signals of the power amplifier, and the target samples are the output signals of the power amplifier; During the training process, NMSE is used as the loss function. After the loss function converges, the training is completed.

8. The power amplifier linearization method based on KAN network according to claim 7, characterized in that: In S3, the in-phase component and the orthogonal component after predistortion processing are input to the power amplifier. If a new state appears in the power amplifier, the coefficients of the real-valued delay KAN digital predistortion network or the vector decomposition delay KAN digital predistortion network are updated, including the following steps: S401, collecting the input signal and output signal of the power amplifier, and recording the data set obtained when the power amplifier appears in a new state for the Nth time as ; S402: The first N-1 random sampling data With new state data Merge to get a complete data set : S403, yes Random sampling The data as , covering the storage area ; S403, using the full dataset The real-value delay KAN digital pre-distortion network or the vector decomposition delay KAN digital pre-distortion network is trained, and the corresponding distortion network is updated after the training is completed.

9. A digital predistortion device based on KAN, characterized in that: include: An input module, used to obtain an in-phase component and an orthogonal component of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment when a delay is subtracted from the current moment; A digital pre-distortion module, used for processing an input signal through a real-valued time-delay KAN digital pre-distortion network, wherein the real-valued time-delay KAN digital pre-distortion network comprises several layers of kanlayer, the input of the next layer is the output of the previous layer, and the output of each layer is the product of the input signal of the layer and the activation function of the layer; the kanlayer is used to compensate for the nonlinear distortion caused by the power amplifier; The output module is used to output the in-phase component and the orthogonal component of the predistortion processed signal.

10. A digital predistortion device based on KAN, characterized in that: include: An input module, used to obtain the amplitude and phase of an input signal, wherein the input signal includes a signal between a current moment and a past moment, wherein the past moment is a moment obtained by subtracting a delay from the current moment; A digital pre-distortion module, used for processing an input signal through a vector decomposition delay KAN digital pre-distortion network, wherein the vector decomposition delay KAN digital pre-distortion network includes a kanlayer, a phase recovery layer and a fully connected layer, wherein the phase recovery layer includes a plurality of cos PRBs and a plurality of sin PRBs, wherein the kanlayer layer is used for compensating for the nonlinear distortion caused by the power amplifier, wherein the phase recovery layer is used for recovering the amplitude and the phase into an in-phase component and an orthogonal component, and wherein the fully connected layer is used for outputting the in-phase component and the orthogonal component; The output module is used to output the in-phase component and the orthogonal component of the predistortion processed signal.

Images