A three-layer cascaded digital pre-distortion method and system for multi-antenna crosstalk

Abstract

The application belongs to the technical field of multi-antenna power amplifier array, and proposes a three-layer cascaded digital pre-distortion method and system for multi-antenna crosstalk, which comprises the following steps: acquiring multiple antenna input signals and constructing a candidate digital pre-distortion base function set; performing crosstalk item screening on the set to generate an effective base function set; extracting SED feature vectors according to the nonlinear distortion errors of each power amplifier branch, and clustering each power amplifier branch according to the distance between the SED feature vectors; based on the effective base function set and the clustering result, a digital pre-distortion model is constructed and a pre-distortion signal of the current observation data block is output; the distributed recursive prediction error algorithm is used to recursively update the digital pre-distortion parameters; the pre-distortion signal of the next observation data block is generated and output; and the updating is cyclically executed until the subsequent observation data block processing is completed. The application solves the defects of the existing MIMO system full-amount DPD model, such as huge hardware resource consumption, high matrix inversion calculation complexity and easy divergence.

Description

Technical Field

[0001] This invention belongs to the field of multi-antenna power amplifier array technology, and particularly relates to a three-layer cascaded digital predistortion method and system for multi-antenna crosstalk. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of 5G-Advanced and future 6G technologies, massive MIMO (Multiple-Input Multiple-Output) antenna arrays have been widely used in base stations, millimeter-wave communication, and high-capacity wireless transmission systems. Power amplifiers, as key components in multi-antenna transmission links, often operate in near-saturation high-efficiency regions, making them prone to nonlinear problems such as distortion. Digital pre-distortion (DPD) is a technique for inversely compensating for the nonlinear distortion of RF power amplifiers in the digital baseband domain. Existing DPD methods often employ Volterra models, memory polynomial models, etc., and estimate the pre-distortion coefficients using least-squares algorithms.

[0004] However, in large-scale MIMO multi-antenna transmission systems, the implementation difficulty of DPD is significantly increased: on the one hand, there are spatial coupling, radio frequency leakage and crosstalk between antennas in different antenna branches, and the output of the target channel will be interfered with by signals from other channels; on the other hand, the nonlinear characteristics of the power amplifiers at different positions of the array are different. If a full-order DPD model is built independently for each power amplifier, it will bring extremely high computation, storage and hardware implementation overhead.

[0005] Existing technologies suffer from the following main shortcomings: high overhead of basis functions and hardware resources; the number of basis functions increases with the number of antennas and the order of nonlinearity; independent parameter configuration consumes a significant amount of hardware, which is not conducive to low-power real-time implementation; high computational complexity of traditional coefficient solving; high-dimensional operations put pressure on storage and are also affected by hardware errors, reducing the stability of coefficient updates and limiting the online update capability of DPD; full modeling introduces a large number of low-contribution redundant parameters; modeling low-power crosstalk terms increases the amount of computation and causes overfitting, reducing generalization ability and affecting out-of-band spectrum suppression effect; the nonlinear similarity between multiple power amplifiers is not fully utilized; traditional independent modeling causes parameter redundancy, further increasing the implementation complexity in large-scale array scenarios. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a three-layer cascaded digital predistortion method and system for multi-antenna crosstalk, which solves the problems of basis function redundancy, high parameter dimensionality and complex online updates caused by traditional full-scale modeling of DPD in multi-antenna power amplifier arrays with high isolation or weak high-order crosstalk contribution.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a three-layer cascaded digital predistortion method for multi-antenna crosstalk, comprising: S1: Acquire the input signals from multiple antennas and construct a set of candidate digital predistortion basis functions; S2: Perform crosstalk term filtering on the basis function set, block the higher-order crosstalk basis functions formed by cross-antenna crosstalk paths, dynamically filter the remaining first-order crosstalk terms, retain the first-order crosstalk terms whose physical cross power exceeds a preset threshold, and generate an effective basis function set. S3: Extract SED feature vectors based on the nonlinear distortion error between the output feedback signal of each power amplifier branch and the corresponding ideal linear output, and cluster each power amplifier branch according to the distance between the SED feature vectors. S4: Based on the effective basis function set and clustering results, construct a low-order independent and high-order shared digital predistortion model and output the predistortion signal of the current observation data block; S5: Obtain the current output feedback signal of each power amplifier branch, calculate the prior prediction error, and use a distributed recursive prediction error algorithm to recursively update the digital predistortion parameters in the digital predistortion model, including low-order independent parameters, retained first-order crosstalk term parameters, and high-order shared parameters; generate and output the predistortion signal of the next observation data block; S5 is executed repeatedly until the subsequent observation data blocks are processed.

[0008] As a further technical solution, crosstalk term screening is performed on the basis function set, specifically including: firstly, based on the power transfer relationship of the crosstalk path, high-order crosstalk terms formed by multiple cross-antenna coupling superposition are eliminated, and only first-order crosstalk terms generated by direct coupling from other antennas are retained; then, for the retained first-order crosstalk terms, the physical cross power corresponding to each first-order crosstalk term is calculated, and weak crosstalk terms whose cross power does not reach the preset threshold are eliminated, and finally the effective basis function set after screening is obtained.

[0009] As a further technical solution, the physical cross power is defined as:

[0010] in, and These are the ideal baseband signals for the main antenna and the interfering antenna, respectively. The spatial coupling coefficient between the two is... A regularization constant is used to prevent zero overflow in hardware division. This represents the number of sampling points contained in the current observation data block.

[0011] As a further technical solution, the nonlinear distortion error is defined as:

[0012] in, Indicates the first The power amplifier output feedback signal of the root antenna branch in the uncompensated state or the current observation state; Indicates the first The ideal linear output reference corresponding to the root antenna branch; This is the estimated linear gain for this branch.

[0013] As a further technical solution, clustering of each power amplifier branch based on the distance between the SED feature vectors includes: extracting the SED features of each power amplifier branch based on the nonlinear distortion error of each power amplifier branch, and representing the SED features as SED feature vectors; pre-setting the number of power amplifier clusters and initializing the cluster centers of each power amplifier cluster; calculating the squared Euclidean distance between the SED feature vector of each power amplifier branch and each cluster center, and assigning each power amplifier branch to the power amplifier cluster corresponding to the cluster center with the smallest distance; updating the cluster centers based on the SED feature vectors of the power amplifier branches within each power amplifier cluster, and repeating the distance calculation, power amplifier branch classification, and cluster center update until the preset clustering convergence condition is met; each power amplifier branch assigned to the power amplifier cluster corresponding to the same cluster center is determined to belong to the same power amplifier cluster; after clustering, constructing a digital predistortion parameter structure with independent low-order parameters and shared high-order parameters within each power amplifier cluster.

[0014] As a further technical solution, the digital predistortion model is as follows:

[0015] in, For the first The digital predistortion output signal of the root antenna branch; It is a set of low-order basis functions; It is a set of higher-order basis functions; The set of valid first-order crosstalk basis functions that are retained; For the first Independent low-order parameters of each antenna branch; For the first High-order parameters shared within each cluster; The parameters for the first-order crosstalk term to be retained; Indicates the first power amplifier The values ​​of the low-order basis functions Indicates the first The first power amplifier Values ​​of higher-order nonlinear basis functions Indicates the first The power amplifier retains the first term after crosstalk filtering. The first-order crosstalk basis function values.

[0016] As a further technical solution, a distributed recursive prediction error algorithm is used to recursively update the digital predistortion parameters in the digital predistortion model, specifically including: The prior prediction error is calculated based on the input basis function vector, target reference signal, and digital predistortion parameters of a power amplifier cluster at a certain moment. The recursive gain vector is calculated based on the input basis function vector, inverse covariance matrix, and forgetting factor of a power amplifier cluster at a certain moment. Based on the prior prediction error and recursive gain vector, the low-order independent parameters and retained first-order crosstalk parameters of each power amplifier branch, as well as the high-order shared parameters of each power amplifier cluster, are updated, and the inverse covariance matrix corresponding to the aforementioned digital predistortion parameters is updated.

[0017] A second aspect of the present invention provides a three-layer cascaded digital predistortion system for multi-antenna crosstalk, comprising: The multi-input signal module is configured to: acquire multiple antenna input signals and construct a set of candidate digital predistortion basis functions; The crosstalk filtering module is configured to: filter crosstalk terms on the basis function set, block higher-order crosstalk basis functions formed by cross-antenna crosstalk paths, dynamically filter the remaining first-order crosstalk terms, retain first-order crosstalk terms whose physical cross power exceeds a preset threshold, and generate an effective basis function set. The SED clustering initialization and shared configuration module is configured to: extract SED feature vectors based on the nonlinear distortion error between the output feedback signal of each power amplifier branch and the corresponding ideal linear output, and cluster each power amplifier branch based on the distance between the SED feature vectors. The intra-cluster low-order independent / high-order shared DPD module is configured to: construct a low-order independent and high-order shared digital predistortion model based on the effective basis function set and clustering results, and output the predistortion signal of the current observation data block; The distributed RPEM parameter update module is configured to: acquire the current output feedback signal of each power amplifier branch, calculate the prior prediction error, and use a distributed recursive prediction error algorithm to recursively update the digital predistortion parameters in the digital predistortion model, including low-order independent parameters, retained first-order crosstalk term parameters, and high-order shared parameters; generate and output the predistortion signal of the next observation data block. The DPD coefficient update module is configured to repeatedly execute the steps in the distributed RPEM parameter update module until the subsequent observation data blocks are processed.

[0018] A third aspect of the invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of a three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in the first aspect of the invention.

[0019] The fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in the first aspect of the present invention.

[0020] The above one or more technical solutions have the following beneficial effects: This invention employs a two-stage crosstalk term selection mechanism driven by physical contribution. First, it blocks higher-order crosstalk terms formed by cross-antenna crosstalk paths, preventing them from entering the basis function generation and parameter estimation process. Second, it calculates the normalized physical cross power for the remaining first-order crosstalk terms and filters them according to a preset threshold, retaining only first-order crosstalk basis functions that effectively compensate for the target antenna branch, thus reducing the complexity of DPD modeling. Furthermore, it adopts a SED-based clustering-based order parameter sharing approach to cluster power amplifier branches with similar nonlinear distortion characteristics, and within each cluster, it uses a parameter organization structure where low-order parameters are independent and high-order parameters are shared. This approach reduces redundant high-order parameter configurations in multi-antenna arrays, decreasing the number of model parameters and multiplier usage.

[0021] Before parameter solving, this invention reduces the number of basis functions and parameters to be estimated by blocking higher-order crosstalk terms, filtering first-order crosstalk terms, and sharing higher-order parameters within the cluster. In this low-dimensional parameter space, the system employs a distributed RPEM recursive update method to update channel-level low-order parameters and cluster-level shared higher-order parameters, avoiding the large-scale matrix inversion process in traditional full-scale least squares methods, thus improving parameter solving efficiency and online update capability.

[0022] This invention employs a sub-parameter structure of "independent low-order and shared high-order parameters." Low-order parameters are independently maintained by each antenna branch, used to compensate for differences in gain, phase, time delay, and linear memory between different branches. High-order parameters are shared within a power amplifier cluster with similar nonlinear distortion characteristics, used to reduce redundancy of high-order nonlinear parameters. This structure balances reduced complexity with linearization compensation performance, achieving a good trade-off between complexity and linearization performance while preserving the individual compensation capabilities of each antenna branch and utilizing the high-order nonlinear similarity between power amplifier branches within the same cluster.

[0023] This invention employs feedback signals to extract nonlinear distortion errors and can periodically or event-triggeredly update clustering results and DPD parameters based on feedback errors, changes in transmit power, temperature variations, or changes in linearization indices. Combined with a distributed RPEM recursive update mechanism, it can adapt to dynamic operating conditions such as power amplifier characteristic drift, changes in service load, and changes in the array's working environment, exhibiting a certain degree of online adaptive capability and making it suitable for dynamic operating scenarios of multi-antenna power amplifier arrays.

[0024] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0026] Figure 1 Here is a block diagram of the three-layer cascaded low-complexity DPD system of the first embodiment; Figure 2 The flowchart of the three-layer cascaded digital predistortion method in the first embodiment is shown below; Figure 3 This is a comparison chart of the output power spectral density of the radio frequency transmitter in the first embodiment; Figure 4 This is a schematic diagram of a three-layer cascaded digital predistortion system module for the second embodiment. Detailed Implementation

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0029] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0030] Terminology Explanation: Multiple-input multiple-output (MIMO) technology transmits and receives signals through multiple antennas. By utilizing multiple transmit and receive antennas and employing techniques such as spatial multiplexing, beamforming, and spatial diversity, it can increase data rate and channel capacity without increasing bandwidth.

[0031] A power amplifier (PA) is an electronic device used to enhance the power of radio frequency signals. It is a core component in fields such as wireless communication, radar, and satellite communication, responsible for amplifying low-power radio frequency signals to high power levels to meet the requirements of transmission distance and signal coverage.

[0032] Digital pre-distortion (DPD) is a commonly used power amplifier linearization technique. Its basic idea is to add digital compensation at the input of the power amplifier that is opposite to the nonlinear characteristics of the power amplifier, so that the output signal of the power amplifier is as close as possible to the ideal linear output.

[0033] SED (Squared Error Distribution): The squared error distribution characterizes the severity and distribution characteristics of nonlinear distortion in each power amplifier under uncompensated conditions.

[0034] LC-DPD (Low-Complexity DPD) refers to a matrix reconstruction architecture that employs "low-order independence and high-order sharing" in the implementation examples.

[0035] RPEM (Recursive Prediction Error Method) is an adaptive filtering algorithm that uses scalar iteration instead of high-dimensional matrix inversion.

[0036] To address the shortcomings of existing MIMO systems, such as excessive hardware resource consumption, high computational complexity of matrix inversion, and susceptibility to divergence in full-scale DPD models, this invention provides a three-layer cascaded low-complexity digital predistortion method and system for multi-antenna crosstalk. It proposes a three-step joint dimensionality reduction architecture involving two-stage physical space truncation, dynamic clustering of error features, and scalar iterative solution. This architecture effectively filters low-contribution crosstalk terms and combines power amplifier clustering sharing and low-dimensional recursive updates to reduce the complexity of DPD implementation for multi-antenna power amplifier arrays. The specific implementation is illustrated in the following examples.

[0037] Example 1 like Figure 1 As shown, this embodiment discloses a three-layer cascaded digital predistortion method for multi-antenna crosstalk. First, an effective basis function set for FCS is generated by physically blocking higher-order crosstalk terms and dynamically thresholding first-order crosstalk terms. Then, the power amplifier channels are clustered based on nonlinear distortion error and SED similarity, and a low-order independent and high-order shared DPD model is constructed. Finally, a distributed RPEM is used to recursively update the DPD parameters to generate a predistortion signal, which can be iteratively processed for the next data block. Figure 1The solid line represents the forward signal transmission path, and the dashed line represents the feedback observation path or parameter configuration update path. Specifically, it includes the following steps: Step S1: Acquire the input signals from multiple antennas and construct a set of candidate digital predistortion basis functions; Step S2: Filter the basis function set for crosstalk terms, block the higher-order crosstalk basis functions formed by cross-antenna crosstalk paths, dynamically filter the remaining first-order crosstalk terms, retain the first-order crosstalk terms whose physical cross power exceeds a preset threshold, and generate an effective basis function set.

[0038] Specifically: In multi-antenna power amplifier arrays, spatial coupling and crosstalk exist between antennas. Modeling all cross-antenna crosstalk terms and their higher-order nonlinear combinations would rapidly increase the number of DPD basis functions and parameters to be estimated. To reduce model complexity, we first block the higher-order crosstalk basis functions formed by cross-antenna crosstalk paths, preventing them from entering subsequent basis function generation, coefficient estimation, and multiplication accumulation processes.

[0039] After blocking higher-order crosstalk terms, the system further dynamically filters the remaining first-order linear crosstalk terms. For the... root antenna branch to the first The normalized physical cross-power of the first-order crosstalk path generated by the root target antenna branch is defined as:

[0040] In the formula, and These are the ideal baseband signals for the main antenna and the interfering antenna, respectively. The spatial coupling coefficient between the two is... A regularization constant to prevent zero overflow during hardware division; This represents the number of sampling points contained in the current observation data block.

[0041] Will With preset crosstalk power threshold Compare. When: When the first-order crosstalk term is determined to have effective compensation significance for the target antenna branch, the basis function corresponding to the first-order crosstalk term is retained; when the following conditions are met: When the physical contribution of the first-order crosstalk term is deemed small, the system does not generate the basis function corresponding to that term. If the threshold is expressed in dB, it can also be represented as: ,in, This is the crosstalk power threshold expressed in dB. This threshold can be set according to system noise floor, antenna isolation, error vector magnitude (EVM) requirements, adjacent channel leakage ratio (ACLR) requirements, or hardware complexity constraints.

[0042] After the above processing, the system forms an effective basis function set for FCS. This set removes higher-order crosstalk basis functions and first-order crosstalk basis functions below the threshold, retaining only the effective nonlinear basis functions of the target power amplifier itself and the first-order crosstalk basis functions with compensation significance, thereby reducing the dimensionality of subsequent DPD modeling and parameter estimation.

[0043] In this embodiment, this step is particularly applicable to multi-antenna power amplifier array scenarios with high antenna spatial isolation and low physical contribution of cross-antenna high-order crosstalk terms. In such scenarios, the power of high-order crosstalk terms formed by far-end antennas or multi-stage crosstalk paths, after spatial coupling attenuation and power amplifier nonlinear path attenuation, is usually lower than the effective compensation threshold or close to the system noise floor. If such high-order crosstalk terms are still fully modeled, it will significantly increase the number of DPD basis functions, the number of parameters to be estimated, and the hardware multiplier consumption, while contributing little to the final linearization performance. Therefore, under the condition of satisfying the above physical contribution, cross-antenna high-order crosstalk basis functions are not generated in the basis function generation stage, and the remaining first-order crosstalk terms are further normalized and screened by physical cross power, retaining only first-order crosstalk terms higher than the preset threshold to participate in subsequent DPD modeling, so as to reduce model complexity.

[0044] It should be noted that what is blocked are the higher-order crosstalk terms formed by cross-antenna crosstalk paths, not the higher-order nonlinear compensation terms of the target power amplifier itself. The third-order, fifth-order, and other higher-order nonlinear terms of the target power amplifier itself can still be retained as needed by the DPD model to compensate for the intrinsic nonlinear distortion of the power amplifier; As another implementation, when the crosstalk coupling strength between antennas in a multi-antenna power amplifier array is high, or when some high-order crosstalk terms formed by neighboring channels still effectively contribute to linearization compensation, the strategy of not generating high-order crosstalk terms in this embodiment can be replaced by a selective retention strategy of high-order crosstalk terms based on physical contribution. Specifically, the normalized physical contribution or equivalent error contribution of cross-antenna high-order crosstalk terms can be calculated. When the contribution of the high-order crosstalk term is lower than a preset effective compensation threshold, the corresponding high-order crosstalk basis function is not generated; when the contribution of the high-order crosstalk term is higher than the preset effective compensation threshold, some of its corresponding neighboring high-order crosstalk basis functions are retained to participate in subsequent DPD modeling. Thus, in scenarios with strong crosstalk or obvious neighboring coupling, excessive deletion of effective basis functions can be avoided, while still reducing modeling redundancy caused by far-end or low-contribution high-order crosstalk terms.

[0045] Step S3: Extract SED (Spectral Envelope Difference) feature vectors based on the nonlinear distortion error between the output feedback signal of each power amplifier branch and the corresponding ideal linear output, and cluster each power amplifier branch according to the distance between the SED feature vectors.

[0046] Specifically: After reducing the number of crosstalk basis functions in step S2, parameter redundancy is further reduced by utilizing the nonlinear distortion similarity between different power amplifier branches. In this embodiment, clustering is not performed directly based on the input signals of each antenna branch, but rather based on the nonlinear distortion error between the power amplifier output signal and the ideal linear output.

[0047] For the For a single antenna branch, its nonlinear distortion error is defined as:

[0048] in, Indicates the first The power amplifier output feedback signal of the root antenna branch in the uncompensated state or the current observation state; Indicates the first The ideal linear output reference corresponding to the root antenna branch; This is the estimated linear gain for this branch.

[0049] Based on the nonlinear distortion error, extract the first... SED characteristics of the root antenna branch.

[0050] This feature may include one or more of the following: error energy, error amplitude distribution, error phase characteristics, error spectrum characteristics, or memory effect characteristics. In one embodiment, error energy can be expressed as:

[0051] Furthermore, the SED features of each branch can be represented as feature vectors. The squared Euclidean distance is used to measure the distortion similarity between different power amplifier branches:

[0052] in, Indicates the first root antenna branch and the first Distortion characteristic distance between root antenna branches.

[0053] Multiple antenna branches are clustered based on the aforementioned SED characteristic distance. In one implementation, the K-Means algorithm can be used, with the clustering objective function being:

[0054] in, Indicates the first A power amplifier cluster, Indicates the number of clusters. Indicates the first The cluster centers of each cluster. The clustering process specifically includes: SED features of each power amplifier branch are extracted based on the nonlinear distortion error of each branch, and the SED features are represented as SED feature vectors. The number of power amplifier clusters is preset, and the cluster centers of each power amplifier cluster are initialized. The squared Euclidean distance between the SED feature vector of each power amplifier branch and each cluster center is calculated, and each power amplifier branch is assigned to the power amplifier cluster corresponding to the cluster center with the smallest distance. The cluster centers are updated based on the SED feature vectors of the power amplifier branches within each power amplifier cluster, and the distance calculation, power amplifier branch classification, and cluster center update are repeated until the preset clustering convergence condition is met. Power amplifier branches that are assigned to the same power amplifier cluster are determined to belong to the same power amplifier cluster.

[0055] In this embodiment, SED clustering can be implemented by combining offline initialization and online updates. During the offline initialization phase, the system extracts the nonlinear distortion error of each power amplifier branch based on the training data and completes the initial clustering. During the online operation phase, the system can periodically or event-triggeredly update the clustering results based on feedback error, temperature changes, transmit power changes, or linearization index changes.

[0056] In another implementation, in addition to using the K-Means clustering algorithm, it can also be replaced with density-based spatial clustering algorithm DBSCAN, Gaussian mixture model GMM, hierarchical clustering algorithm or other unsupervised clustering methods that can complete feature grouping based on the similarity, correlation or distance metric relationship between uncompensated nonlinear error features, depending on the distribution form of the uncompensated nonlinear error features and the hardware implementation requirements.

[0057] S4: Based on the effective basis function set and clustering results, construct a low-order independent and high-order shared digital predistortion model and output the predistortion signal of the current observation data block.

[0058] After clustering, a DPD parameter structure with independent low-order parameters and shared high-order parameters is constructed within each power amplifier cluster. For low-order terms, such as first-order terms, linear memory terms, or main linear delay terms, independent parameters are reserved for each antenna branch to accommodate differences in gain, phase, time delay, routing, and impedance between different branches. For high-order terms, such as third-order, fifth-order, and higher-order nonlinear terms, the system shares a set of high-order DPD parameters within the same cluster to leverage the similarity in nonlinear distortion characteristics among power amplifiers in the same cluster.

[0059] Therefore, for those belonging to the first The first cluster The digital predistortion model of the root antenna branch can be expressed as:

[0060] in, For the first The DPD output signal of the root antenna branch; It is a set of low-order basis functions; It is a set of higher-order basis functions; This is the set of valid first-order crosstalk basis functions retained in step S1; For the first Independent low-order parameters of each antenna branch; For the first High-order parameters shared within each cluster; The parameters for the first-order crosstalk term to be retained; Indicates the first power amplifier The values ​​of the low-order basis functions Indicates the first The first power amplifier Values ​​of higher-order nonlinear basis functions Indicates the first The power amplifier retains the first term after crosstalk filtering. The first-order crosstalk basis function values.

[0061] Through this hierarchical sharing structure, this embodiment does not allow antennas within the same cluster to share complete DPD parameters, but rather allows them to share high-order nonlinear parameters while keeping low-order parameters independent, thereby reducing the number of parameters while preserving the individual compensation capability of each antenna branch.

[0062] In another implementation, the intra-cluster high-order parameter sharing method can be adjusted based on the nonlinear distortion similarity between power amplifier branches within the same cluster. When the SED distance between power amplifier branches within the same cluster is less than a preset sharing threshold, a DPD parameter structure with independent low-order parameters and intra-cluster sharing of high-order parameters is adopted. When the SED distance between power amplifier branches within the same cluster is greater than the preset sharing threshold, a method with independent third-order parameters and shared fifth-order and higher-order parameters can be adopted, or a method of superimposing high-order shared master parameters onto channels with independent parameter correction can be adopted to reduce compensation mismatch caused by excessive sharing of high-order parameters.

[0063] Furthermore, the basis function set and shared structure can be dynamically adjusted based on online linearization performance metrics. For example, when the observed NMSE, ACLR, or prediction error exceeds a preset threshold, some blocked nearest-neighbor third-order crosstalk basis functions can be released, or the SED features can be recalculated and the power amplifier clustering results updated. When linearization performance meets the requirements and the contribution of some basis functions remains below the threshold, these low-contribution basis functions can be re-masked. In this way, an adaptive trade-off between complexity and linearization performance can be achieved under different crosstalk intensities, different power amplifier differences, and different hardware resource constraints.

[0064] Step S5: Obtain the current output feedback signal of each power amplifier branch, calculate the prior prediction error, and use a distributed recursive prediction error algorithm to recursively update the digital predistortion parameters in the digital predistortion model, including low-order independent parameters, retained first-order crosstalk term parameters, and high-order shared parameters; generate and output the predistortion signal for the next observation data block. Specifically: After steps S2-S4, the number of basis functions and the number of parameters to be estimated in the DPD model have been reduced. As one implementation method, based on this sparse shared parameter structure, the distributed recursive prediction error method (RPEM) is used to update the DPD parameters to avoid the large-scale matrix inversion in traditional least squares methods.

[0065] The DPD parameters are updated using a distributed recursive prediction error algorithm, including: The prior prediction error is calculated based on the input basis function vector, target reference signal, and digital predistortion parameters of a power amplifier cluster at a given moment. The recursive gain vector is calculated based on the input basis function vector, inverse covariance matrix, and forgetting factor of the power amplifier cluster at a given moment. Based on the prior prediction error and recursive gain vector, the low-order independent parameters and retained first-order crosstalk parameters of each power amplifier branch, as well as the high-order shared parameters of each power amplifier cluster, are updated, and the inverse covariance matrix corresponding to the aforementioned digital predistortion parameters is also updated. Specifically: Let the current iteration time be... , No. The row vector of the input basis function corresponding to each power amplifier cluster is: (Its dimension has been reduced from S1 to the lowest), the target reference signal is ,in, The DPD parameter vector to be updated is determined by the effective basis functions retained in step S2 and the low-order independent and high-order shared parameter structures determined in step S4; the system updates the DPD parameter vector to be updated in closed loop according to the following formula. : Calculate the current prior error :

[0066] in, The digital predistortion parameters are from the previous time step.

[0067] Calculate the Kalman gain vector (recursive gain vector) obtained at the current iteration time, and calculate the complexity factor. The shortening leads to an exponential decrease:

[0068] in, Let be the inverse covariance matrix of the previous time step. Forgetting factor, It is the conjugate transpose of the input basis function vector.

[0069] Update the DPD parameter vector to be updated for this amplifier branch:

[0070] Updating the inverse covariance matrix yields the inverse covariance matrix at the current iteration time. :

[0071] Since step S2 reduces invalid crosstalk basis functions, and steps S3 and S4 reduce high-order parameter redundancy, the dimensions of both the input basis function vector and the DPD parameter vector to be updated are smaller than the corresponding vector dimensions in the traditional full-matrix DPD model. Compared with the least squares solution based on the full matrix, this step can reduce the multiplication, addition, and storage overhead required for parameter updates, and is suitable for parallel recursive updates of multi-amplifier clusters.

[0072] In another implementation, under different hardware resource constraints, real-time requirements, and compensation accuracy requirements, the RPEM algorithm can be replaced by the Recursive Least Squares (RLS) algorithm, the Normalized Minimum Mean Square Error (NLMS) algorithm, or other adaptive filtering update algorithms. After replacement, it can still be combined with the preceding feature truncation, feature clustering, and cluster sharing dimensionality reduction architecture to reduce the overall parameter update dimension and computational complexity while maintaining nonlinear compensation capability.

[0073] Step S6: Repeat step S5 until all data blocks have been processed.

[0074] In this embodiment, steps S2-S5 are progressive: step S2 reduces invalid crosstalk basis functions entering the model by filtering based on physical contribution; steps S3-S4, based on the filtered effective features, establish a low-order independent and high-order shared DPD parameter structure using the power amplifier nonlinear distortion similarity; step S5 uses distributed RPEM for recursive updates in the low-dimensional parameter space. These three steps work together to reduce the implementation complexity of the multi-antenna digital predistortion system from three aspects: the number of basis functions, the number of parameters to be estimated, and the complexity of parameter solving, while maintaining good NMSE and ACLR performance with relatively small linearization performance loss.

[0075] In this embodiment, as shown in Table 1, the resource consumption and performance are compared with those of the existing DPD model.

[0076] Table 1. Comparison of resource consumption and performance between this method and existing DPD models.

[0077] Simulation results show that the proposed joint architecture maintains linearization performance close to that of the full model while significantly reducing the number of equivalent parameters. Under weak crosstalk coupling conditions, the NMSE and ACLR of the proposed joint architecture are not significantly lower than those of the full FOCA model, with NMSE losing only about 0.57 dB and ACLR losing only about 0.74 dB, indicating that the removed cross-antenna higher-order crosstalk terms contribute little to linearization performance in this scenario. Reducing the multiplier consumption from 144 to 40 represents a reduction of approximately 72.2% in multiplier resource consumption, effectively alleviating the pressure on DSP slice resources in hardware platforms such as FPGAs and improving the engineering feasibility of large-scale multi-antenna DPD systems.

[0078] like Figure 3 As shown, without DPD, the power amplifier output signal exhibits significant spectral broadening in the adjacent channel region. Using traditional independent FOCA and the full model, adjacent channel leakage is significantly suppressed. Although the proposed joint architecture has slightly lower out-of-band suppression capability than the full model, its overall spectral shape remains close. Combined with the multiplier consumption results, it can be seen that the method in this embodiment achieves a significant reduction in complexity while introducing only a small loss in spectral performance.

[0079] Therefore, the core of this embodiment lies in the joint application of "physical blocking of higher-order crosstalk terms and dynamic screening of first-order crosstalk terms", "power amplifier clustering and graded parameter sharing modeling based on nonlinear distortion error" and "distributed RPEM recursive update based on sparse shared parameter structure" to the multi-antenna digital predistortion method, forming a three-layer joint low-complexity architecture from basis function screening, parameter organization to parameter solution.

[0080] Specifically, a physical contribution screening mechanism is first introduced during the basis function generation stage. For higher-order crosstalk terms formed by cross-antenna crosstalk paths, due to their weak physical contribution and high modeling complexity, corresponding higher-order crosstalk basis functions are not generated. For the remaining first-order crosstalk terms, their normalized physical cross power is further calculated and compared with a preset threshold; only first-order crosstalk basis functions above the threshold are retained. Thus, before proceeding to subsequent DPD modeling, a set of effective basis functions for the FCS is formed, reducing the parameter dimension and multiplication computation caused by invalid or low-contribution crosstalk terms.

[0081] Based on the aforementioned effective basis function set, parameter recombination is further performed using the similarity of nonlinear distortion among the power amplifier branches. Instead of clustering the power amplifier branches based on the main input signal, SED features are extracted based on the nonlinear distortion error between the power amplifier output and the ideal linear reference, and the power amplifier branches are clustered based on these features. After clustering, the power amplifier branches within the same cluster adopt a DPD parameter structure where low-order parameters are independent and high-order parameters are shared. That is, low-order parameters are still independently maintained by each antenna branch, while high-order nonlinear parameters are shared within the cluster. This approach preserves the individual compensation capability of each branch while reducing high-order parameter redundancy in multi-amplifier systems.

[0082] In the parameter update phase, instead of performing traditional least squares solutions in the full DPD basis function and independent parameter space, a distributed RPEM is used to recursively update channel-level low-order parameters and cluster-level high-order shared parameters in a low-dimensional parameter space after crosstalk term filtering and intra-cluster high-order sharing. Since the first two stages of processing have reduced the number of basis functions and parameters to be estimated, the dimensions of the input vector and parameter vector in the RPEM update process are correspondingly reduced, thereby reducing multiplication, addition, and storage overhead, and avoiding large-scale matrix inversions in traditional least squares methods.

[0083] Therefore, the three main parts of this embodiment are not simply a stack of algorithm modules, but have a clear progressive relationship: the first-level crosstalk term screening provides low-dimensional effective basis function inputs for subsequent clustering and modeling; clustering and order sharing further reduce the number of high-order parameters to be estimated; and the parameter structure after the first two levels of dimensionality reduction provides conditions for low-complexity recursive updates of distributed RPEM. The three work together to reduce the implementation complexity of the multi-antenna DPD system from three aspects: the number of basis functions, the number of parameters, and the solution complexity, while maintaining good power amplifier linearization performance with relatively small NMSE and ACLR performance losses.

[0084] Example 2 This embodiment discloses a three-layer cascaded digital predistortion system for multi-antenna crosstalk, including: The multi-input signal module is configured to: acquire multiple antenna input signals and construct a set of candidate digital predistortion basis functions; The crosstalk filtering module is configured to: filter crosstalk terms in the basis function set, block higher-order crosstalk basis functions formed by cross-antenna crosstalk paths, dynamically filter the remaining first-order crosstalk terms, retain first-order crosstalk terms whose physical cross power exceeds a preset threshold, and generate an effective basis function set. The SED clustering initialization and shared configuration module is configured to: extract SED feature vectors based on the nonlinear distortion error between the output feedback signal of each power amplifier branch and the corresponding ideal linear output, and cluster each power amplifier branch based on the distance between the SED feature vectors. The intra-cluster low-order independent / high-order shared DPD module is configured to: construct a low-order independent and high-order shared digital predistortion model based on the effective basis function set and clustering results, and output the predistortion signal of the current observation data block; The distributed RPEM parameter update module is configured to: acquire the current output feedback signal of each power amplifier branch, calculate the prior prediction error, and use a distributed recursive prediction error algorithm to recursively update the digital predistortion parameters in the digital predistortion model, including low-order independent parameters, retained first-order crosstalk term parameters, and high-order shared parameters; generate and output the predistortion signal of the next observation data block. The DPD coefficient update module is configured to repeatedly execute the steps in the distributed RPEM parameter update module until the subsequent observation data blocks are processed.

[0085] like Figure 4 As shown, the upper part represents the main transmit link of the multi-antenna DPD, and the lower part represents the feedback observation and parameter update link. The main link first physically blocks higher-order crosstalk terms, then performs power threshold filtering on the first-order crosstalk terms to form a set of effective basis functions for the FCS. Subsequently, based on the clustering results, it constructs a low-order independent and high-order shared DPD model and outputs a pre-distortion signal. The feedback link acquires the observation signal from the power amplifier output or antenna coupling end, extracts the nonlinear distortion error after ADC conversion, and uses it for SED feature extraction and power amplifier clustering, as well as for distributed RPEM parameter updates. The updated DPD coefficients are then written back to the model.

[0086] This embodiment is not a simple superposition of multiple independent modules, but rather forms a joint low-complexity digital predistortion architecture for heterogeneous large-scale MIMO systems, with a progressive joint dimensionality reduction relationship among the three core modules: (1) The simplified feature matrix output in step S1 can reduce the interference of weak crosstalk terms on the clustering results and improve the array clustering accuracy in step S2. (2) The feature dimensionality reduction in step S1 and the parameter sharing in step S2 together reduce the parameter dimension and the size of the covariance matrix; (3) Step S3 is based on the above-mentioned dimensionality reduction structure to realize low-complexity online parameter update, thereby improving the system real-time performance and FPGA implementation efficiency.

[0087] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0088] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of a three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in Embodiment 1 of this disclosure.

[0089] Example 4 The purpose of this embodiment is to provide an electronic device.

[0090] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in Embodiment 1 of this disclosure.

[0091] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0092] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0093] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A three-layer cascaded digital predistortion method for multi-antenna crosstalk, characterized in that, include: S1: Acquire the input signals from multiple antennas and construct a set of candidate digital predistortion basis functions; S2: Perform crosstalk term filtering on the basis function set, block the higher-order crosstalk basis functions formed by cross-antenna crosstalk paths, dynamically filter the remaining first-order crosstalk terms, retain the first-order crosstalk terms whose physical cross power exceeds a preset threshold, and generate an effective basis function set. S3: Extract SED feature vectors based on the nonlinear distortion error between the output feedback signal of each power amplifier branch and the corresponding ideal linear output, and cluster each power amplifier branch according to the distance between the SED feature vectors. S4: Based on the effective basis function set and clustering results, construct a low-order independent and high-order shared digital predistortion model and output the predistortion signal of the current observation data block; S5: Obtain the current output feedback signal of each power amplifier branch, calculate the prior prediction error, and use a distributed recursive prediction error algorithm to recursively update the digital predistortion parameters in the digital predistortion model, including low-order independent parameters, retained first-order crosstalk term parameters, and high-order shared parameters; generate and output the predistortion signal of the next observation data block; S5 is executed repeatedly until the subsequent observation data blocks are processed.

2. The three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in claim 1, characterized in that, The crosstalk term selection for the basis function set specifically includes: firstly, based on the power transfer relationship of the crosstalk path, removing higher-order crosstalk basis functions formed by multiple cross-antenna coupling superpositions, retaining only first-order crosstalk terms generated by direct coupling from other antennas; then, for the retained first-order crosstalk terms, calculating the physical cross-power corresponding to each first-order crosstalk term, removing weak crosstalk terms whose cross-power does not reach a preset threshold, and finally obtaining the selected effective basis function set.

3. The three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in claim 2, characterized in that, Define the physical cross power as: in, and These are the ideal baseband signals for the main antenna and the interfering antenna, respectively. The spatial coupling coefficient between the two is... A regularization constant is used to prevent zero overflow in hardware division. This represents the number of sampling points contained in the current observation data block.

4. The three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in claim 1, characterized in that, The nonlinear distortion error is defined as: in, Indicates the first The power amplifier output feedback signal of the root antenna branch in the uncompensated state or the current observation state; Indicates the first The ideal linear output reference corresponding to the root antenna branch; This is the estimated linear gain for this branch.

5. A three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in claim 1, characterized in that, Clustering of each power amplifier branch based on the distance between the SED feature vectors includes: extracting the SED features of each power amplifier branch based on the nonlinear distortion error of each power amplifier branch, and representing the SED features as SED feature vectors; pre-setting the number of power amplifier clusters and initializing the cluster centers of each power amplifier cluster; calculating the squared Euclidean distance between the SED feature vector of each power amplifier branch and each cluster center, and assigning each power amplifier branch to the power amplifier cluster corresponding to the cluster center with the smallest distance; updating the cluster centers based on the SED feature vectors of the power amplifier branches within each power amplifier cluster, and repeating the distance calculation, power amplifier branch classification, and cluster center update until the preset clustering convergence condition is met; power amplifier branches that are assigned to the same power amplifier cluster are determined to belong to the same power amplifier cluster; after clustering, constructing a digital predistortion parameter structure with independent low-order parameters and shared high-order parameters within each power amplifier cluster.

6. A three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in claim 1, characterized in that, The digital predistortion model is as follows: in, For the first The digital predistortion output signal of the root antenna branch; It is a set of low-order basis functions; It is a set of higher-order basis functions; The set of valid first-order crosstalk basis functions that are retained; For the first Independent low-order parameters of each antenna branch; For the first High-order parameters shared within each cluster; The parameters for the first-order crosstalk term to be retained; Indicates the first power amplifier The values ​​of the low-order basis functions Indicates the first The first power amplifier Values ​​of higher-order nonlinear basis functions Indicates the first The power amplifier retains the first term after crosstalk filtering. The first-order crosstalk basis function values.

7. A three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in claim 1, characterized in that, The digital predistortion parameters in the digital predistortion model are updated iteratively using a distributed recursive prediction error algorithm, specifically including: The prior prediction error is calculated based on the input basis function vector, target reference signal, and digital predistortion parameters of a power amplifier cluster at a certain moment. The recursive gain vector is calculated based on the input basis function vector, inverse covariance matrix, and forgetting factor of a power amplifier cluster at a certain moment. Based on the prior prediction error and recursive gain vector, the low-order independent parameters and retained first-order crosstalk parameters of each power amplifier branch, as well as the high-order shared parameters of each power amplifier cluster, are updated, and the inverse covariance matrix corresponding to the aforementioned digital predistortion parameters is updated.

8. A three-layer cascaded digital predistortion system for multi-antenna crosstalk, characterized in that, include: The multi-input signal module is configured to: acquire multiple antenna input signals and construct a set of candidate digital predistortion basis functions; The crosstalk filtering module is configured to: filter crosstalk terms on the basis function set, block higher-order crosstalk basis functions formed by cross-antenna crosstalk paths, dynamically filter the remaining first-order crosstalk terms, retain first-order crosstalk terms whose physical cross power exceeds a preset threshold, and generate an effective basis function set. The SED clustering initialization and shared configuration module is configured to: extract SED feature vectors based on the nonlinear distortion error between the output feedback signal of each power amplifier branch and the corresponding ideal linear output, and cluster each power amplifier branch based on the distance between the SED feature vectors. The intra-cluster low-order independent / high-order shared DPD module is configured to: construct a low-order independent and high-order shared digital predistortion model based on the effective basis function set and clustering results, and output the predistortion signal of the current observation data block; The distributed RPEM parameter update module is configured to: acquire the current output feedback signal of each power amplifier branch, calculate the prior prediction error, and use a distributed recursive prediction error algorithm to recursively update the digital predistortion parameters in the digital predistortion model, including low-order independent parameters, retained first-order crosstalk term parameters, and high-order shared parameters; generate and output the predistortion signal of the next observation data block. The DPD coefficient update module is configured to repeatedly execute the steps in the distributed RPEM parameter update module until the subsequent observation data blocks are processed.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps of a three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the three-layer cascaded digital predistortion method for multi-antenna crosstalk as described in any one of claims 1-7.

Images