A method and system for sparsification and online updating of a Volterra equalizer
By combining sparse Bayesian learning and Kalman filtering, the dominant kernel is selected and the Volterra equalizer model is updated online, which solves the problems of large model size and parameter update mismatch, achieves efficient nonlinear compensation and stable bit error rate performance, and adapts to channel changes.
Patent Information
- Application Number
- CN202610955915.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-30
- Publication Date
- 2026-11-17
AI Technical Summary
Existing Volterra equalizer models are large in scale, making hardware deployment difficult. The parameter update strategy after pruning lacks effective connection with the pruning stage, resulting in insufficient channel tracking capability and degraded equalization performance, especially in time-varying or bursty noise environments.
By constructing an initial regression matrix and introducing a Gaussian prior, sparse Bayesian learning is performed iteratively to select the dominant kernel and remove redundant kernels, the Kalman filter is initialized, and state prediction, prediction error calculation and Kalman gain correction are performed in the online update stage to output a sparse Volterra equalizer model.
It significantly reduces model complexity and online update overhead while maintaining good nonlinear compensation capability and bit error rate performance, improves the ability to track dominant kernel coefficients in time-varying channel environments, and enhances robustness and the ability to quickly adapt to channel changes.
Smart Images

Figure CN123123081A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of optical communication digital signal processing technology, and in particular to a method and system for sparsification and online updating of a Volterra equalizer. Background Technology
[0002] With the development of data center interconnects and high-speed passive optical networks, intensity modulation / direct detection (IM / DD) based optical transmission systems have been widely adopted due to their simple structure and low cost. However, when the signal rate increases and higher-order pulse amplitude modulation (PAM) is used, factors such as nonlinearity of the transmitter modulator, square-law detection effect of the receiver photodetector, and fiber link damage can lead to severe nonlinear distortion and memory effects, causing inter-symbol interference and deterioration of the bit error rate. Traditional linear equalizers are difficult to effectively compensate for these effects. Volterra nonlinear equalizers (VNEs) can simultaneously describe the nonlinearity and memory characteristics of the system and achieve good compensation performance, but the number of cores increases explosively with the order of nonlinearity and the length of memory, resulting in a huge model size and extremely high computational complexity, making it difficult to deploy on real-time platforms such as FPGAs.
[0003] In existing technologies, the Volterra equalizer pruning method based on Sparse Bayesian learning (SBL) is a typical approach. This method assigns autocorrelation-based priors to kernel coefficients, uses iterative inference to shrink redundant kernel coefficients to zero, completes the selection of dominant kernels and model compression, and then combines traditional adaptive algorithms to update the retained kernel coefficients online. This method can significantly reduce the number of effective kernels and the subsequent update dimensionality. However, its drawbacks are: the dominant kernel and posterior uncertainty information obtained in the SBL stage lack effective connection with the subsequent online update algorithm, and the conventional recursive strategy does not fully utilize the inferred parameter distribution and noise statistics, resulting in insufficient parameter tracking ability of the pruned model, especially in time-varying channel or burst noise environments, where the equalization performance is significantly degraded.
[0004] In view of this, there is an urgent need in the field for a technical solution that deeply integrates the kernel selection capability of SBL with the parameter recursive update capability of Kalman Filter (KF) to achieve rapid convergence and stable tracking by making full use of the parameter statistical information inferred during the pruning stage while compressing the model size and reducing the update complexity. This would overcome the problems of disconnect between pruning and updating, performance degradation after pruning, and difficulty in real-time deployment in the existing technology. Summary of the Invention
[0005] In view of this, embodiments of the present invention provide a sparsification and online update method and system for Volterra equalizers to solve the technical problems in the prior art, such as the difficulty in hardware deployment due to the large scale of Volterra equalizer models, and the lack of effective connection between the parameter update strategy after pruning and the pruning stage, which leads to insufficient channel tracking capability and degraded equalization performance.
[0006] One aspect of the present invention provides a method for sparsification and online updating of a Volterra equalizer, the method comprising the following steps: The training sample sequence of the received signal and the corresponding expected output signal are obtained, and the initial regression matrix of the Volterra equalizer is constructed based on the training sample sequence. The initial regression matrix contains all candidate kernel terms of the Volterra equalizer. Gaussian priors are introduced for the coefficients of all candidate kernel terms. Sparse Bayesian learning is performed iteratively until convergence based on the initial regression matrix and the expected output signal to obtain the posterior mean vector and posterior covariance matrix of the coefficients of each candidate kernel term. The candidate kernel terms are screened based on the posterior mean vector and the posterior covariance matrix, retaining the dominant kernel and eliminating redundant kernels to obtain a set of retained kernels; the corresponding columns are extracted from the initial regression matrix according to the set of retained kernels to obtain a sparse regression matrix, and the components corresponding to the set of retained kernels are extracted from the posterior mean vector to obtain a sparse kernel coefficient vector; The Kalman filter is initialized by using the sparse kernel coefficient vector as the initial state vector and the submatrix in the posterior covariance matrix corresponding to the set of retained kernels as the initial covariance matrix. Entering the online update phase, the current sparse regression vector is extracted from the current input signal according to the construction rules of the Volterra kernel terms corresponding to the reserved kernel set. Prediction is performed based on the state vector, covariance matrix, and process noise covariance of the previous time step to obtain the prior state vector and prior covariance matrix of the current time step. The prediction error is calculated based on the current sparse regression vector, the prior state vector, and the expected output value corresponding to the current sample, and the observation noise variance is determined based on the variance estimate of the prediction error. The Kalman gain is calculated based on the prior covariance matrix, the current sparse regression vector, and the observation noise variance. The prior state vector is corrected using the Kalman gain and the prediction error to obtain the updated state vector, which represents the updated coefficients of each kernel term in the reserved kernel set. The prior covariance matrix is then corrected based on the Kalman gain and the current sparse regression vector to obtain the updated covariance matrix. The output consists of a sparse Volterra equalizer model composed of the set of retained kernels and their updated coefficients.
[0007] In some embodiments of the present invention, constructing the initial regression matrix of the Volterra equalizer based on the training sample sequence includes: Based on the preset memory length and preset nonlinear order, linear terms, second-order nonlinear terms, and third-order nonlinear terms are extracted from the received signals corresponding to each sampling time in the training sample sequence to form the Volterra regression vector corresponding to each sampling time. The Volterra regression vectors at each sampling time point are stacked row by row to construct the initial regression matrix; wherein the rows of the initial regression matrix correspond to each sampling time point in the training sample sequence, and the columns correspond to all candidate kernel terms.
[0008] In some embodiments of the present invention, a Gaussian prior is introduced for each candidate kernel term coefficient, including: The coefficients of each candidate kernel term are assigned a zero-mean Gaussian prior, the covariance matrix of which is a diagonal matrix, and each diagonal element of which is controlled by a hyperparameter corresponding to each candidate kernel term, the hyperparameter representing the prior variance of the corresponding kernel term coefficient. Based on the initial regression matrix and the expected output signal, and combined with the Gaussian prior, the posterior mean vector and the posterior covariance matrix are calculated.
[0009] In some embodiments of the present invention, during the sparse Bayesian learning iteration process, the hyperparameters and noise variance are updated synchronously until convergence. The hyperparameters are updated based on the ratio of the squared modulus of the posterior mean component of the corresponding candidate kernel term to the sum of the corresponding diagonal elements of the posterior covariance matrix and a preset constant term. The calculation formula is as follows: ; in, Indicates the first The hyperparameters corresponding to the coefficients of each kernel term; The first element of the posterior mean vector represents the... One component; Denotes the first of the posterior covariance matrices. One diagonal element; A preset constant term to prevent division by zero; The noise variance is iteratively updated using the current residual term and the posterior uncertainty represented by the posterior covariance matrix, and the calculation formula is: ; in, This represents the updated noise variance; This represents the training target vector composed of the desired output signal; This represents the initial regression matrix; This represents the number of samples in the training sample sequence; This indicates the total number of all candidate kernel terms; Represents the posterior mean vector; Denotes the first of the posterior covariance matrices. One diagonal element; Indicates the first The hyperparameters corresponding to the kernel coefficients.
[0010] In some embodiments of the present invention, the prediction based on the state vector of the previous time step, the covariance matrix of the previous time step, and the process noise covariance includes: Based on the state vector and unit state transition matrix of the previous time step, the prior state vector of the current time step is predicted. The calculation formula of the state prediction equation is as follows: ; in, This represents the unit state transition matrix; This represents the state vector of the previous time step; This represents the prior state vector at the current moment; By combining the process noise covariance to predict the covariance matrix of the previous time step, the prior covariance matrix of the current time step is obtained. The calculation formula of the covariance prediction equation is as follows: ; in, This represents the covariance matrix of the previous time step; Represents the prior covariance matrix at the current moment; This represents the process noise covariance matrix; Wherein, the process noise covariance matrix It decreases as the sample number increases, and Process noise figure The attenuation formula is: ; in, Indicates the initial process noise figure; Indicates the sample sequence number; Indicates the attenuation control parameters; This represents the identity matrix.
[0011] In some embodiments of the present invention, the determination of the observation noise variance at the current moment based on the variance estimation of the prediction error includes: Set a smoothing factor; recursively estimate the observation noise variance at the current time based on the observed noise variance of the previous time step and the squared magnitude of the current prediction error, using the following formula: ; in, This represents the variance of the observation noise at the current moment; This represents the variance of the observation noise at the previous moment; This represents the smoothing factor; This indicates the current prediction error; The formula for calculating the Kalman gain is: ; in, Indicates the Kalman gain; Denotes the prior covariance matrix; This represents the current sparse regression vector; This indicates the preset weight parameters; This represents the variance of the observation noise.
[0012] In some embodiments of the present invention, the step of correcting the prior state vector using the Kalman gain and the prediction error includes: The prior state vector is corrected using the Kalman gain and the prediction error, and the calculation formula is as follows: ; in, This represents the updated state vector; Represents the prior state vector; Indicates the Kalman gain, This represents the prediction error; The prior covariance matrix is corrected based on the Kalman gain and the current sparse regression vector to obtain the updated covariance matrix, and the updated state vector is used as the updated coefficient of each kernel term in the retained kernel set.
[0013] On the other hand, the present invention also provides a sparsification and online update system for a Volterra equalizer, including a processor, a memory, and a computer program or instructions stored in the memory, wherein the processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the system implements the steps of the method described in any of the above.
[0014] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the method as described in any of the preceding claims.
[0015] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of any of the methods described above.
[0016] This invention provides a sparsification and online update method and system for a Volterra equalizer. It constructs an initial regression matrix and introduces a Gaussian prior to all candidate kernel coefficients. Sparse Bayesian learning iterations are then performed to obtain the posterior mean vector and posterior covariance matrix. Based on this, dominant kernels are selected and redundant kernels are removed to obtain a set of retained kernels, a sparse regression matrix, and a sparse kernel coefficient vector. Subsequently, a Kalman filter is initialized with the sparse kernel coefficient vector and the corresponding posterior covariance submatrix. During the online update phase, a sparse regression vector is constructed sample-by-sample, and state prediction, prediction error calculation, observation noise variance determination, Kalman gain calculation, and recursive correction of the state vector and covariance matrix are completed. The output is a sparse Volterra equalizer model composed of the set of retained kernels and their updated coefficients. This invention combines offline kernel selection using sparse Bayesian learning with online recursive updates using Kalman filtering. This allows the initial state vector to directly inherit the posterior estimation results, and the initial covariance matrix to directly reflect the corresponding uncertainty. This avoids the initial convergence error caused by recursion from zero state. While significantly reducing model complexity and online update overhead, it maintains good nonlinear compensation capability and bit error rate performance, and improves the tracking capability of dominant kernel coefficients in time-varying channel environments.
[0017] Furthermore, this invention employs a process noise covariance that decays inversely with the sample number during the Kalman filter recursion process. This allows the Kalman filter to have a large adjustment bandwidth in the initial stage to quickly adapt to channel changes, and gradually tends to a stable estimate in the later stage, thus balancing convergence speed and steady-state accuracy.
[0018] Furthermore, this invention estimates the observation noise variance through smooth recursion, enabling it to adaptively track changes in environmental noise levels and improve the robustness of the Volterra equalizer in time-varying channel scenarios. At the same time, the weighting parameters introduced in the Kalman gain calculation can flexibly adjust the contribution ratio of the observation noise variance in the correction process, making the dynamic trade-off between prior estimation and observation correction more refined.
[0019] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.
[0020] Those skilled in the art will understand that the objectives and advantages achievable with this invention are not limited to those specifically described above, and that the above and other objectives achievable with this invention will become clearer from the following detailed description. Attached Figure Description
[0021] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.
[0022] Figure 1 This is a flowchart illustrating the sparsity reduction and online update method of the Volterra equalizer according to an embodiment of the present invention.
[0023] Figure 2 This is a flowchart illustrating the sparsity reduction and online update method of the Volterra equalizer according to an embodiment of the present invention.
[0024] Figure 3 This is a comparison chart of the bit error rate performance of the Full-Model Volterra Equalizer (Full-VNE), the Pruned Volterra Equalizer Based on Sparse Bayesian Learning and Recursive Least Squares (SBL-RLS-VNE), and the Pruned Volterra Equalizer Based on Sparse Bayesian Learning and Kalman Filtering (SBL-KF-VNE) under different received optical powers, as described in another embodiment of the present invention. Detailed Implementation
[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.
[0026] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.
[0027] In existing technologies, while the complete Volterra nonlinear equalizer possesses strong nonlinear compensation capabilities, the number of model kernels rapidly expands with increasing nonlinear order and memory length, leading to a significant increase in computational complexity and hardware implementation overhead, making it difficult to directly apply to real-time high-speed optical communication systems. Furthermore, while existing pruning methods based on sparse Bayesian learning can compress model size, the adaptive update strategy of retaining kernel coefficients after pruning is often ineffective. In time-varying channel or burst noise environments, it is prone to insufficient parameter tracking capability and degraded equalization performance, hindering the practical application of sparse equalizers.
[0028] In view of this, the present invention provides a method for sparsification and online updating of a Volterra equalizer, such as... Figure 1 As shown, the method includes the following steps S101~S106: S101: Obtain the training sample sequence of the received signal and the corresponding expected output signal, and construct the initial regression matrix of the Volterra equalizer based on the training sample sequence. The initial regression matrix contains all candidate kernel terms of the Volterra equalizer.
[0029] S102: Introduce a Gaussian prior to the coefficients of all candidate kernel terms, and perform sparse Bayesian learning iterations based on the initial regression matrix and the expected output signal until convergence, to obtain the posterior mean vector and posterior covariance matrix of the coefficients of each candidate kernel term.
[0030] S103: Based on the posterior mean vector and posterior covariance matrix, candidate kernel terms are screened, dominant kernels are retained and redundant kernels are removed to obtain a set of retained kernels; the corresponding columns are extracted from the initial regression matrix according to the set of retained kernels to obtain a sparse regression matrix, and the components corresponding to the set of retained kernels are extracted from the posterior mean vector to obtain a sparse kernel coefficient vector.
[0031] S104: Initialize the Kalman filter by using the sparse kernel coefficient vector as the initial state vector and the submatrix in the posterior covariance matrix corresponding to the set of retained kernels as the initial covariance matrix.
[0032] S105: Entering the online update phase, the current sparse regression vector is extracted from the current input signal according to the construction rules of the Volterra kernel terms corresponding to the retained kernel set; prediction is performed based on the state vector, covariance matrix, and process noise covariance of the previous time step to obtain the prior state vector and prior covariance matrix of the current time step; the prediction error is calculated based on the current sparse regression vector, prior state vector, and expected output value corresponding to the current sample, and the observation noise variance of the current time step is determined based on the variance estimation of the prediction error; the Kalman gain is calculated based on the prior covariance matrix, current sparse regression vector, and observation noise variance; the prior state vector is corrected using the Kalman gain and prediction error to obtain the updated state vector, which represents the updated coefficients of each kernel term in the retained kernel set, and the prior covariance matrix is corrected based on the Kalman gain and current sparse regression vector to obtain the updated covariance matrix.
[0033] S106: Output a sparse Volterra equalizer model consisting of the retained kernel set and its updated coefficients.
[0034] In step S101, the training sample sequence of the received signal and the corresponding expected output signal are first obtained. The training sample sequence contains received signal samples at multiple sampling times, and the expected output signal is the known transmitting end signal corresponding to each sampling time.
[0035] Subsequently, the initial regression matrix of the Volterra equalizer is constructed based on the training sample sequence. The Volterra equalizer is a digital signal processing algorithm used to simultaneously compensate for nonlinear distortion and memory effects in communication systems. It reconstructs the original transmitted signal by linearly combining received signal samples from the current and historical moments and introducing product terms to characterize nonlinear distortion.
[0036] In some embodiments, the Volterra equalizer employs a third-order model, and its equalization output is expressed as follows: ; in, This represents the received signal sample at the nth sampling time. Indicates balanced output; Indicates the DC bias term; Indicates the noise term; Indicates the memory length; i, j, and k represent delay indices, ranging from 0 to L-1, representing the delay amount relative to the current sampling time n; Represents the first-order kernel coefficients with delay i; Let i represent the second-order kernel coefficients with delays of i and j respectively; Let represent the third-order kernel coefficients with delays of i, j, and k. In the above expression, For linear terms, It is a second-order nonlinear term. These are third-order nonlinear terms, and these linear terms and nonlinear terms of each order are collectively referred to as Volterra kernel terms.
[0037] Furthermore, all kernel coefficients are arranged in order to form a coefficient vector, expressed as: ; in, Represents the coefficient vector of all candidate kernel terms; This represents a subvector of first-order kernel coefficients composed of all first-order kernel coefficients arranged in order. This represents a subvector of second-order kernel coefficients composed of all second-order kernel coefficients arranged in order; This represents a subvector of third-order kernel coefficients composed of all third-order kernel coefficients arranged in order; When written in vector form, each kernel term is first organized into an inner product of the regression sub-vectors and the coefficient sub-vectors. That is, the balanced output is represented as the sum of the inner products of each order plus the noise term. The specific expression is as follows: ; Where p represents the order index, which takes the value 1, 2, or 3; This represents the regressor subvector composed of the p-th order kernel terms; Represents a coefficient subvector consisting of the p-th order kernel coefficients; Indicates the first All kernel terms at each sampling time are organized into a Volterra regression vector, which contains a constant term 1 and regression subvectors of each order; the superscript T indicates the transpose operation.
[0038] In constructing the initial regression matrix, linear, second-order, and third-order nonlinear terms are extracted from the received signal samples at each sampling time in the training sample sequence to form the Vorterra regression vector for each sampling time. The Vorterra regression vectors at each sampling time are then stacked row-wise to construct the initial regression matrix, where the rows correspond to each sampling time in the training sample sequence, and the columns correspond to all candidate kernel terms. Simultaneously, the vector composed of the desired output signals at each sampling time is denoted as the training target vector.
[0039] In some embodiments, the specific extraction range of the linear terms and nonlinear terms of each order is determined according to the preset memory length and the preset nonlinear order, and an initial regression matrix is constructed by stacking rows.
[0040] Thus, step S101 completes the acquisition of training data and the construction of the initial regression matrix, resulting in an initial regression matrix containing all candidate kernel terms and a training target vector composed of the expected output signals at each sampling time.
[0041] In step S102, a Gaussian prior is introduced into the coefficients of each candidate kernel term, and sparse Bayesian learning iteration is performed based on the initial regression matrix and the expected output signal to obtain the posterior mean vector and posterior covariance matrix of the coefficients of each candidate kernel term.
[0042] Among them, sparse Bayesian learning is a sparsification method based on Bayesian inference. It assigns a sparse prior distribution to the parameters to be estimated, calculates the posterior distribution of the parameters using the observed data, and automatically shrinks the prior variance of unimportant parameters to near zero during the iteration process, thereby achieving automatic parameter selection.
[0043] In some embodiments, the coefficient vector of all candidate kernel terms is first generated. Assuming a zero-mean Gaussian prior distribution, the expression is: ; in, This represents the prior covariance matrix, which is a diagonal matrix, with each diagonal element... Let represent the prior variance, or hyperparameter, corresponding to the coefficient of the i-th kernel term; M represents the total number of candidate kernel terms. Each candidate kernel term has an independent hyperparameter, and the priors of different kernel term coefficients are independent of each other. Before the iteration begins, each hyperparameter and noise variance is assigned initial values.
[0044] Given the prior covariance matrix and noise variance, based on Bayes' theorem, and combining the initial regression matrix and the training target vector, the coefficient vector is calculated. The posterior distribution.
[0045] In some embodiments, the posterior covariance matrix is calculated as follows: ; in, Represent the posterior covariance matrix; Represents the initial regression matrix; Indicates the noise variance; Denotes the prior covariance matrix. represents the inverse of the prior covariance matrix.
[0046] In some embodiments, the posterior mean vector is calculated as follows: ; in, Represents the posterior mean vector; Represent the posterior covariance matrix; This represents the training target vector consisting of the desired output signal.
[0047] In each iteration, based on the current posterior mean vector and posterior covariance matrix Simultaneously update each hyperparameter and noise variance.
[0048] In some embodiments, the update formula for each hyperparameter is: ; in, This represents the hyperparameter corresponding to the updated i-th kernel term coefficient; This represents the i-th component of the posterior mean vector; This represents the i-th diagonal element of the posterior covariance matrix; To prevent division by zero, a preset constant term is used. When the square of the modulus of the posterior mean component of a certain kernel term... Larger and diagonal elements of the posterior covariance matrix When it is small, its hyperparameters As the posterior mean component increases, the prior constraint on the core term weakens, and the coefficient is retained and effectively estimated; conversely, when the square of the modulus of the posterior mean component approaches zero, its hyperparameter will decrease, the prior constraint will strengthen, the coefficient will be compressed to near zero, and the core term will be gradually eliminated.
[0049] In some embodiments, the update formula for noise variance is: ; in, This represents the updated noise variance; This represents the training target vector composed of the desired output signal; Represents the initial regression matrix; represents the posterior mean vector; N represents the number of samples in the training sample sequence; M represents the total number of candidate kernel terms; This represents the i-th diagonal element of the posterior covariance matrix; Let represent the hyperparameter corresponding to the coefficient of the i-th kernel term. The numerator of this update formula... The L2 norm square of the residual vector reflects the model's fitting error to the training data; the summation term in the denominator It reflects the posterior uncertainty of each core term coefficient and incorporates the posterior uncertainty into the estimation of noise variance, so that the estimation of noise variance can adaptively reflect the complexity of the current model.
[0050] Repeat the above process of posterior distribution calculation, hyperparameter update, and noise variance update until the posterior mean vector converges. During the iteration, the hyperparameters corresponding to insignificant kernel terms gradually shrink to near zero, while the hyperparameters corresponding to dominant kernel terms remain valid, thus achieving automatic selection of kernel terms. After convergence, the posterior mean vector and posterior covariance matrix of the coefficients of each candidate kernel term are obtained.
[0051] Thus, step S102 completes the iterative process of sparse Bayesian learning, obtaining the posterior mean vector and posterior covariance matrix of all candidate kernel coefficients. Each component in the posterior mean vector reflects the estimated value of the corresponding kernel coefficient, and each diagonal element of the posterior covariance matrix reflects the uncertainty of the corresponding estimated value.
[0052] In step S103, candidate kernel terms are screened, retaining the dominant kernel and removing redundant kernels to obtain a set of retained kernels. Based on this set, the corresponding sparse regression matrix and sparse kernel coefficient vector are extracted from the initial regression matrix and the posterior mean vector. Through threshold determination and secondary pruning, kernel terms that play a dominant role in the equilibrium performance are determined from the candidate kernel terms evaluated by sparse Bayesian learning, and a sparse model structure containing only the dominant kernel is constructed.
[0053] Specifically, after the iterative convergence in step S102, the hyperparameters of each candidate kernel term have differentiated: the hyperparameters corresponding to the dominant kernel term remain valid, the squared modulus of the corresponding posterior mean component is large, and the diagonal elements of the posterior covariance matrix are small; the hyperparameters corresponding to redundant kernel terms shrink to near zero, and the squared modulus of the corresponding posterior mean component approaches zero. Based on this, a threshold screening is performed by combining the hyperparameters and posterior mean components of each candidate kernel term. Kernel terms whose hyperparameters shrink to near zero or whose squared modulus of the posterior mean component is lower than a preset threshold are identified as redundant kernels and removed. Kernel terms whose hyperparameters remain valid and whose squared modulus of the posterior mean component is higher than a preset threshold are retained, thus obtaining the candidate kernel set.
[0054] Subsequently, a second pruning process is performed on the candidate kernel set. By setting a more stringent threshold, marginal contribution kernels are removed from the candidate kernel set, and only the dominant kernels that make substantial contributions to the equilibrium performance are retained, resulting in the retained kernel set.
[0055] After determining the set of kernels to be retained, the corresponding columns are extracted from the initial regression matrix according to the index position of each kernel term in the retained kernel set to form a sparse regression matrix. The number of columns in the sparse regression matrix is reduced from the total number of candidate kernel terms to the number of kernels in the retained kernel set. At the same time, the components corresponding to the retained kernel set are extracted from the posterior mean vector to obtain the sparse kernel coefficient vector. The dimension of the sparse kernel coefficient vector is consistent with the number of kernels in the retained kernel set.
[0056] In some embodiments, the expression for the pruned Volterra equalizer is: ; in, This indicates the balanced output after pruning; This represents a sparse regression vector consisting of the Volterra kernel terms corresponding to the retained kernel set; Represents the sparse kernel coefficient vector; This indicates the noise term.
[0057] Thus far, step S103 has completed the kernel selection and sparse model construction, resulting in the retained kernel set, sparse regression matrix, and sparse kernel coefficient vector.
[0058] In step S104, the sparse kernel coefficient vector obtained in step S103 is used as the initial state vector of the Kalman filter, the submatrix corresponding to the set of retained kernels in the posterior covariance matrix obtained in step S102 is used as the initial covariance matrix, and the state transition matrix is set as the identity matrix to complete the initialization of the Kalman filter.
[0059] The Kalman filter is a recursive estimation algorithm that achieves recursive estimation of time-varying parameters by alternating between two steps: predicting the system state and correcting for observations. Before starting the recursive process, an initial state vector and an initial covariance matrix need to be set. The initial state vector represents the filter's initial estimate of the system state before the recursion begins, while the initial covariance matrix characterizes the uncertainty of this initial estimate.
[0060] In this step, the Kalman filter is used to recursively update the dominant kernel coefficients retained after pruning. Its initial state vector directly adopts the sparse kernel coefficient vector obtained in step S103. ,Right now Sparse kernel coefficient vector It is the component corresponding to the retained kernel set extracted from the posterior mean vector after the sparse Bayesian learning iteration converges. It reflects the estimated value of each dominant kernel coefficient during the training phase. Thus, the Kalman filter has an effective initial estimate of each dominant kernel coefficient before entering the online update phase.
[0061] For the initial covariance matrix, the posterior covariance matrix obtained in step S102 In the process, the rows and columns corresponding to each kernel item in the retained kernel set are extracted to form a submatrix, which serves as the initial covariance matrix of the Kalman filter. Posterior covariance matrix The diagonal elements represent the uncertainty of the corresponding kernel term coefficient estimates. The extracted submatrix contains only the posterior uncertainty information of each kernel term in the retained kernel set, and its dimension is equal to the number of kernels in the retained kernel set. This initial covariance matrix... This provides an initial uncertainty baseline for the Kalman filter to dynamically balance predicted and observed values in subsequent recursive processes.
[0062] Through the above initialization operations, the Kalman filter directly inherits the parameter estimation results and uncertainty assessment information of sparse Bayesian learning. It does not need to start from zero state or preset initial value for recursion. In the early stage of online update, it has effective estimation of each dominant kernel coefficient and corresponding estimation reliability information.
[0063] At this point, step S104 completes the initialization of the Kalman filter and obtains the initial state vector. Initial covariance matrix And the unit state transition matrix.
[0064] In step S105, the online update phase begins. For the input signal received sample by sample, the coefficients of each kernel term in the retained kernel set are recursively updated using a Kalman filter. During the online update phase, a complete recursive process from prediction to correction is executed every time a new input signal sample is received, ensuring that the dominant kernel coefficients can track channel changes in real time.
[0065] For the current sampling time m, firstly, according to the Volterra kernel term construction rule corresponding to the preserved kernel set, extract the current sparse regression vector from the current input signal. The sparse regression vector contains only the values of each kernel term in the kernel set at the current sampling time, and its dimension is equal to the number of kernels in the kernel set.
[0066] Then, a prediction step is performed, based on the state vector from the previous time step. covariance matrix of the previous time step Combined with the process noise covariance matrix Calculate the prior state vector at the current time. and prior covariance matrix .
[0067] In some embodiments, the state prediction equation is calculated as follows: ; in, Let represent the state transition matrix, and , i.e., the identity matrix, means that the predicted prior state vector remains unchanged from the previous state vector when there is no new observation information; This represents the state vector from the previous moment; This represents the prior state vector at the current moment.
[0068] In some embodiments, the covariance prediction equation is calculated as follows: ; in, Represents the covariance matrix of the previous time step; This represents the prior covariance matrix at the current moment; This represents the process noise covariance matrix.
[0069] Process noise covariance matrix Used to characterize the uncertainty of the evolution of the state from the previous moment to the current moment.
[0070] In some embodiments, the process noise covariance matrix decreases as the sample number increases, and is expressed as a scalar multiplication of an identity matrix. Process noise figure The attenuation calculation formula is: ; in, Indicates the initial process noise figure; Indicates the attenuation control parameters; Indicates the sample sequence number; This represents the identity matrix. As the recursion progresses, the process noise figure gradually decreases, allowing the filter to have a large tracking bandwidth in the initial stage to quickly adapt to the channel, and then tends to stabilize in the later stages.
[0071] After obtaining the prior state vector and the prior covariance matrix, the prediction error at the current sampling time is calculated. The prediction error is the deviation between the expected output value corresponding to the current sample and the predicted output obtained based on the prior state vector and the current sparse regression vector.
[0072] Based on the prediction error, the variance of the observation noise at the current moment is determined to reflect the uncertainty of the current observation.
[0073] In some embodiments, a smoothing factor is set and recursively estimated based on the variance of the observation noise at the previous time step and the squared modulus of the current prediction error, calculated as follows: ; in, This represents the variance of the observation noise at the current moment; This represents the variance of the observation noise at the previous moment; This represents the preset smoothing factor. Through smoothing estimation, the observed noise variance can adaptively track changes in the environmental noise level.
[0074] After obtaining the observation noise variance, the Kalman gain is calculated. The Kalman gain determines the extent to which the prediction error is utilized when correcting the prior state vector.
[0075] In some embodiments, the Kalman gain is calculated as follows: ; in, Represents the Kalman gain vector; Represent the prior covariance matrix; This represents the current sparse regression vector; This represents the preset weighting parameter, used to adjust the contribution ratio of observation noise variance in gain calculation.
[0076] The prior state vector is corrected using Kalman gain and prediction error to obtain the updated state vector at the current time step. .
[0077] In some embodiments, the expression for state update is: ; in, This represents the updated state vector, where each component is the updated coefficient of each kernel term in the kernel set retained at the current sampling time. Simultaneously, the prior covariance matrix is corrected based on the Kalman gain and the current sparse regression vector to obtain the updated covariance matrix. The updated covariance matrix reflects the residual uncertainty of the corrected state estimate, providing input for the recursion at the next sampling time.
[0078] For each newly arrived sampling time, the above recursive process is repeated to achieve online updates of the retained kernel coefficients on a sample-by-sample basis.
[0079] Thus, step S105 completes one recursive process of the online update phase, obtaining the updated state vector at the current moment. and the updated covariance matrix The updated state vector is the updated coefficient of each kernel term in the kernel set.
[0080] In step S106, the updated state vector at the current time is... Each component is used as the updated coefficient of each kernel term in the retained kernel set, and combined with the retained kernel set to output the sparse Volterra equalizer model at the current time. As the online update phase continues, an updated sparse Volterra equalizer model is output for each new input signal at a new sampling time.
[0081] At this point, step S106 completes the output of the sparse Volterra equalizer model, resulting in a sparse Volterra equalizer model consisting of the retained kernel set and its updated coefficients.
[0082] On the other hand, the present invention also provides a sparsification and online update system for a Volterra equalizer, including a processor, a memory, and a computer program or instructions stored in the memory. The processor is used to execute the computer program or instructions, and when the computer program or instructions are executed, the system implements the steps of any of the methods described above.
[0083] On the other hand, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of any of the methods described above.
[0084] On the other hand, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of any of the above methods.
[0085] The present invention will now be described with reference to a specific embodiment: This embodiment uses a 20 Gbaud PAM4 signal intensity modulation / direct detection optical communication system as an application scenario to fully demonstrate the entire process from initial regression matrix construction, sparse Bayesian learning offline pruning, Kalman filter initialization to online recursive update. Experiments are conducted to compare the performance of the method of this invention with the full-model Volterra equalizer and the recursive least squares-based pruning equalizer in terms of model complexity and bit error rate.
[0086] I. Sparse Algorithm for Volterra Equalizer (VNE) Based on Sparse Bayesian Learning (SBL) and Kalman Filtering (KF) The VNE sparse algorithm flow based on SBL and KF in this embodiment is as follows: Figure 2 As shown, the process is divided into two stages: the first stage uses Sparse Bayesian Learning (SBL) to select the dominant kernel in the complete Volterra model; the second stage uses Kalman Filtering (KF) to recursively update the coefficients of the retained kernels on the pruned sparse regression matrix, thereby constructing a sparse Volterra equalizer.
[0087] Figure 2 This paper illustrates the overall flow of a Volterra equalizer sparsification and online update method based on sparse Bayesian learning and Kalman filtering according to an embodiment of the present invention. The flow is generally divided into two stages: the dashed box on the left represents the offline pruning stage, and the dashed box on the right represents the online update stage.
[0088] Specifically, in the offline pruning stage, the training target vector is first... and the initial regression matrix As input, the training target vector consists of the expected output signal at each sampling time in the training sample sequence of the received signal. The initial regression matrix is constructed from the above training sample sequence, with rows corresponding to training samples and columns corresponding to all candidate kernel terms. The number of columns is equal to the number of kernels in the complete Volterra model. Then, it proceeds to the sparse Bayesian learning iteration, first processing the prior covariance matrix... and noise variance Assign an initial value, denoted as and ,in This is a diagonal matrix, where each diagonal element represents the initial hyperparameters of the candidate kernel term coefficients. Given the initial values, the initial regression matrix, and the training target vector, the posterior distribution of the coefficient vector is calculated to obtain the posterior covariance matrix. and posterior mean vector Then, based on the current posterior statistic, each hyperparameter is updated synchronously. and noise variance The updated hyperparameters and noise variance are then used again for the next round of posterior distribution calculation, thus forming an iterative cycle. When the posterior mean vector... While the convergence is not yet complete, repeat the process of calculating the posterior mean and updating the parameters. Once the posterior mean vector converges, combine the hyperparameters and the posterior mean to perform threshold filtering to obtain a set of candidate kernels. Then, a second pruning process is performed on the candidate kernel set to obtain the retained kernel set. Based on this, the corresponding columns are extracted from the initial regression matrix to output the sparse regression matrix. .
[0089] During the online update phase, the sparse kernel coefficient vector obtained in the sparse Bayesian learning phase is first used as the initial state vector. The submatrix in the posterior covariance matrix corresponding to the preserved kernel set is used as the initial covariance matrix. This completes the initialization of the Kalman filter. Then, the sample-by-sample recursive process begins. For each newly arrived sampling time, covariance prediction is first performed based on the covariance matrix of the previous time step and the process noise covariance to obtain the prior covariance matrix for the current time step. Then, the prediction error is calculated based on the current sparse regression vector, the prior state vector, and the expected output value corresponding to the current sample. The innovation variance at the current time is obtained by smoothing and recursively estimating based on the squared magnitude of the prediction error. Then, based on the prior covariance matrix, the current sparse regression vector, and the aforementioned innovation variance, the Kalman gain is calculated. Finally, the prior state vector is corrected using Kalman gain and prediction error to obtain the updated state vector. Each component of the state vector is the updated coefficient of each kernel term in the kernel set retained at the current sampling time. The above recursive process is repeated at each new sampling time until the sparse Volterra equalizer model is output.
[0090] set up To receive discrete-time sampled signals, To achieve a balanced output, a third-order Volterra equalizer (VNE) can be represented as: ; in, This represents the received signal sample at the nth sampling time. Indicates balanced output; Indicates the DC bias term; Indicates the noise term; Indicates the memory length; i, j, and k represent delay indices, ranging from 0 to L-1, representing the delay amount relative to the current sampling time n; Represents the first-order kernel coefficients with delay i; Let i represent the second-order kernel coefficients with delays of i and j respectively; Let i represent the third-order kernel coefficients with delays of i, j, and k.
[0091] Further orders: ; in, Represents the coefficient vector of all candidate kernel terms; This represents a subvector of first-order kernel coefficients composed of all first-order kernel coefficients arranged in order. This represents a subvector of second-order kernel coefficients composed of all second-order kernel coefficients arranged in order; This represents a subvector of third-order kernel coefficients composed of all third-order kernel coefficients arranged in order; The linear, second-order, and third-order terms are combined to form the Volterra regression vector. The third-order Volterra equalizer (VNE) can then be written in vector form, with the following expression: ; Where p represents the order index, which takes the value 1, 2, or 3; This represents the regressor subvector composed of the p-th order kernel terms; Represents a coefficient subvector consisting of the p-th order kernel coefficients; Indicates the first All kernel terms at each sampling time are organized into a Volterra regression vector, which contains a constant term 1 and regression subvectors of each order; the superscript T indicates the transpose operation.
[0092] As the nonlinear order and memory length increase, the number of kernels in the complete Volterra model increases rapidly. However, in actual applications, not all kernels play an equal role in nonlinear compensation. Therefore, its sparsity can be utilized for kernel extraction and model compression.
[0093] Let the complete Volterra regression matrix be... The training target vector is ,in, The number of training samples, This represents the kernel number of the complete Volterra model. In the Sparse Bayesian Learning (SBL) phase, the coefficient vector is first processed... Assign a Gaussian prior distribution: ; in The prior covariance matrix is a diagonal. This represents the prior variance corresponding to the i-th kernel coefficient. Given the noise variance... Under the given conditions, the formula for calculating the posterior covariance matrix is: ; in, Represent the posterior covariance matrix; Represents the initial regression matrix; Indicates the noise variance; Denotes the prior covariance matrix. represents the inverse of the prior covariance matrix.
[0094] The formula for calculating the posterior mean vector is: ; in, Represents the posterior mean vector; Represent the posterior covariance matrix; This represents the training target vector consisting of the desired output signal.
[0095] when Larger and When the value is small, it indicates that the corresponding kernel has high importance and estimation reliability, therefore its corresponding hyperparameters are... It is also relatively large, so the updated formula is: ; in, This represents the hyperparameter corresponding to the updated i-th kernel term coefficient; This represents the i-th component of the posterior mean vector; This represents the i-th diagonal element of the posterior covariance matrix; A preset constant term to prevent division by zero.
[0096] Simultaneous noise variance Iterative updates using the current residual term and posterior uncertainty are expressed as follows: ; in, This represents the updated noise variance; This represents the training target vector composed of the desired output signal; Represents the initial regression matrix; represents the posterior mean vector; N represents the number of samples in the training sample sequence; M represents the total number of candidate kernel terms; This represents the i-th diagonal element of the posterior covariance matrix; This represents the hyperparameter corresponding to the coefficient of the i-th kernel term.
[0097] During the above iterations, the hyperparameters corresponding to insignificant kernels gradually shrink to near zero, while the dominant kernels remain valid terms. After the posterior mean converges, a threshold is used to filter candidates by combining the hyperparameters and the posterior mean, resulting in a set of candidate kernels. A further threshold can be set to perform secondary pruning, yielding a sparse kernel coefficient vector. and the corresponding sparse regression vector The expression for the pruned Volterra equalizer (VNE) is then: ; in, This indicates the balanced output after pruning; This represents a sparse regression vector consisting of the Volterra kernel terms corresponding to the retained kernel set; Represents the sparse kernel coefficient vector; This indicates the noise term.
[0098] After obtaining the dominant kernel set, a Kalman filter (KF) is used to recursively estimate the retained kernel coefficients sample by sample. The initial state vector of the Kalman filter (KF) is shown below. The initial covariance matrix is given directly from the output of Sparse Bayesian Learning (SBL). Constructed from the posterior covariance matrix. For the m-th training sample, the expression for state prediction is: ; in, Let represent the state transition matrix, and , i.e., the identity matrix, means that the predicted prior state vector remains unchanged from the previous state vector when there is no new observation information; This represents the state vector from the previous moment; This represents the prior state vector at the current moment.
[0099] set up The initial process noise figure, As the attenuation control parameter, the process noise covariance matrix takes a decreasing form, and the expression for the covariance prediction process is: ; in, Represents the covariance matrix of the previous time step; This represents the prior covariance matrix at the current moment; Represents the process noise covariance matrix; Indicates the initial process noise figure; Indicates the attenuation control parameters; Indicates the sample sequence number; Represents the identity matrix.
[0100] Meanwhile, to improve noise tracking capabilities, the prediction error for the current sample is defined as... Set the smoothing factor And recursively update the innovation variance: ; in, This represents the variance of the observation noise at the current moment; This represents the variance of the observation noise at the previous moment; This indicates the preset smoothing factor.
[0101] Based on this, set the weight parameters. The Kalman gain is calculated as follows: ; in, Represents the Kalman gain vector; Represent the prior covariance matrix; This represents the current sparse regression vector; This represents the preset weighting parameter, used to adjust the contribution ratio of observation noise variance in gain calculation.
[0102] Finally, the preserved kernel coefficients are updated: ; in, This represents the updated state vector; Through the above recursive process, stable online parameter updates can be achieved on the pruned sparse model, thereby improving the tracking ability of the dominant kernel coefficients and reducing the performance loss caused by pruning.
[0103] II. Performance Analysis Table 1 compares the complexity of the pruning algorithm based on Sparse Bayesian Learning (SBL), the full-model Vorterra equalizer (VNE) nonlinear equalization-recursive least squares (RLS) algorithm, and the post-pruned Kalman filter (KF) update algorithm. It can be seen that the online update complexity of the full-model VNE-RLS algorithm is... The complexity of the online update algorithm based on Kalman filtering (KF) after pruning is reduced to... Due to the retention of core count Much smaller than the number of cores in the full model The proposed method significantly reduces the online parameter update overhead after model compression. Taking a third-order Volterra equalizer (VNE) with a memory length of 11 and a fixed training sequence length of 2000 as an example, on an intensity modulation / direct detection (IM / DD) optical communication experimental platform with a working wavelength of 1565nm and transmitting 20Gbaud pulse amplitude modulation (PAM4) signals, after sparse Bayesian learning (SBL) pruning, the average number of retained kernels decreased from 1465 to approximately 140, a reduction of 90.43%. Due to the significant reduction in the number of dominant kernels, the normalized online update complexity was further reduced by 99.08%. These results demonstrate that the proposed sparse Volterra equalization scheme based on sparse Bayesian learning (SBL) and Kalman filtering (KF) can effectively alleviate the high complexity problem of the complete Volterra equalizer (VNE) while ensuring nonlinear compensation capabilities, making it more suitable for low-complexity hardware implementation at the receiver.
[0104] Table 1. Complexity of VNE sparse algorithms based on SBL and KF like Figure 3 The figure shows the bit error rate performance of a 20G pulse amplitude modulation (PAM4) intensity modulation / direct detection (IM / DD) optical communication system at a wavelength of 1565nm for the full-model Volterra equalizer (Full-VNE), the pruned Volterra equalizer based on sparse Bayesian learning (SBL) and recursive least squares (RLS) (SBL-RLS-VNE), and the proposed pruned Volterra equalizer based on sparse Bayesian learning (SBL) and Kalman filtering (KF) (SBL-KF-VNE) at different received optical powers.
[0105] It can be seen that although pruning operations cause some performance loss, the proposed SBL-KF-VNE consistently outperforms the SBL-RLS-VNE across the entire received optical power range, with a particularly significant advantage in the medium-to-high received optical power region. At the forward error correction threshold... and In the vicinity, the required received optical power for the SBL-KF-VNE is lower than that for the SBL-RLS-VNE, and... Near the threshold, the performance remains close to that of the full-model Volterra equalizer (Full-VNE). This indicates that Kalman filter (KF) updates can more effectively maintain the tracking ability of dominant nonlinear features after pruning, thereby maintaining good bit error rate performance while significantly reducing complexity.
[0106] In summary, this invention provides a sparsification and online update method and system for a Volterra equalizer. It constructs an initial regression matrix and introduces a Gaussian prior to all candidate kernel coefficients. Sparse Bayesian learning iterations are then performed to obtain the posterior mean vector and posterior covariance matrix. Based on this, dominant kernels are selected and redundant kernels are removed to obtain a set of retained kernels, a sparse regression matrix, and a sparse kernel coefficient vector. Subsequently, the Kalman filter is initialized with the sparse kernel coefficient vector and the corresponding posterior covariance submatrix. During the online update phase, a sparse regression vector is constructed sample by sample, and state prediction, prediction error calculation, observation noise variance determination, Kalman gain calculation, and recursive correction of the state vector and covariance matrix are completed. The output is a sparse Volterra equalizer model composed of the set of retained kernels and their updated coefficients. This invention combines offline kernel selection using sparse Bayesian learning with online recursive updates using Kalman filtering. This allows the initial state vector to directly inherit the posterior estimation results, and the initial covariance matrix to directly reflect the corresponding uncertainty. This avoids the initial convergence error caused by recursion from zero state. While significantly reducing model complexity and online update overhead, it maintains good nonlinear compensation capability and bit error rate performance, and improves the tracking capability of dominant kernel coefficients in time-varying channel environments.
[0107] Furthermore, this invention employs a process noise covariance that decays inversely with the sample number during the Kalman filter recursion process. This allows the Kalman filter to have a large adjustment bandwidth in the initial stage to quickly adapt to channel changes, and gradually tends to a stable estimate in the later stage, thus balancing convergence speed and steady-state accuracy.
[0108] Furthermore, this invention estimates the observation noise variance through smooth recursion, enabling it to adaptively track changes in environmental noise levels and improve the robustness of the Volterra equalizer in time-varying channel scenarios. At the same time, the weighting parameters introduced in the Kalman gain calculation can flexibly adjust the contribution ratio of the observation noise variance in the correction process, making the dynamic trade-off between prior estimation and observation correction more refined.
[0109] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.
[0110] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.
[0111] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.
[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for sparsification and online updating of a Volterra equalizer, characterized in that, The method includes: The training sample sequence of the received signal and the corresponding expected output signal are obtained, and the initial regression matrix of the Volterra equalizer is constructed based on the training sample sequence. The initial regression matrix contains all candidate kernel terms of the Volterra equalizer. Gaussian priors are introduced for the coefficients of all candidate kernel terms. Sparse Bayesian learning is performed iteratively until convergence based on the initial regression matrix and the expected output signal to obtain the posterior mean vector and posterior covariance matrix of the coefficients of each candidate kernel term. The candidate kernel terms are screened based on the posterior mean vector and the posterior covariance matrix, retaining the dominant kernel and eliminating redundant kernels to obtain a set of retained kernels; the corresponding columns are extracted from the initial regression matrix according to the set of retained kernels to obtain a sparse regression matrix, and the components corresponding to the set of retained kernels are extracted from the posterior mean vector to obtain a sparse kernel coefficient vector; The Kalman filter is initialized by using the sparse kernel coefficient vector as the initial state vector and the submatrix in the posterior covariance matrix corresponding to the set of retained kernels as the initial covariance matrix. Entering the online update phase, the current sparse regression vector is extracted from the current input signal according to the construction rules of the Volterra kernel terms corresponding to the reserved kernel set. Prediction is performed based on the state vector, covariance matrix, and process noise covariance of the previous time step to obtain the prior state vector and prior covariance matrix of the current time step. The prediction error is calculated based on the current sparse regression vector, the prior state vector, and the expected output value corresponding to the current sample, and the observation noise variance is determined based on the variance estimate of the prediction error. The Kalman gain is calculated based on the prior covariance matrix, the current sparse regression vector, and the observation noise variance. The prior state vector is corrected using the Kalman gain and the prediction error to obtain the updated state vector, which represents the updated coefficients of each kernel term in the reserved kernel set. The prior covariance matrix is then corrected based on the Kalman gain and the current sparse regression vector to obtain the updated covariance matrix. The output consists of a sparse Volterra equalizer model composed of the set of retained kernels and their updated coefficients.
2. The method for sparsification and online updating of the Volterra equalizer according to claim 1, characterized in that, The construction of the initial regression matrix of the Volterra equalizer based on the training sample sequence includes: Based on the preset memory length and preset nonlinear order, linear terms, second-order nonlinear terms, and third-order nonlinear terms are extracted from the received signals corresponding to each sampling time in the training sample sequence to form the Volterra regression vector corresponding to each sampling time. The Volterra regression vectors at each sampling time point are stacked row by row to construct the initial regression matrix; wherein the rows of the initial regression matrix correspond to each sampling time point in the training sample sequence, and the columns correspond to all candidate kernel terms.
3. The method for sparsification and online updating of the Volterra equalizer according to claim 1, characterized in that, Gaussian priors are introduced for the coefficients of each candidate kernel term, including: The coefficients of each candidate kernel term are assigned a zero-mean Gaussian prior, the covariance matrix of which is a diagonal matrix, and each diagonal element of which is controlled by a hyperparameter corresponding to each candidate kernel term. The hyperparameter represents the prior variance of the corresponding kernel term coefficient. Based on the initial regression matrix and the expected output signal, and combined with the Gaussian prior, the posterior mean vector and the posterior covariance matrix are calculated.
4. The method for sparsification and online updating of the Volterra equalizer according to claim 3, characterized in that, During the sparse Bayesian learning iteration process, the hyperparameters and noise variance are updated synchronously until convergence. The hyperparameters are updated based on the ratio of the squared modulus of the posterior mean component of the corresponding candidate kernel term to the sum of the corresponding diagonal elements of the posterior covariance matrix and a preset constant term. The calculation formula is as follows: ; in, Indicates the first The hyperparameters corresponding to the coefficients of each kernel term; The first element of the posterior mean vector represents the... One component; Denotes the first posterior covariance matrix. One diagonal element; A preset constant term to prevent division by zero; The noise variance is iteratively updated using the current residual term and the posterior uncertainty represented by the posterior covariance matrix, and the calculation formula is: ; in, This represents the updated noise variance; This represents the training target vector composed of the desired output signal; This represents the initial regression matrix; This represents the number of samples in the training sample sequence; This indicates the total number of all candidate kernel terms; Represents the posterior mean vector; Denotes the first posterior covariance matrix. One diagonal element; Indicates the first The hyperparameters corresponding to the kernel coefficients.
5. The method for sparsification and online updating of the Volterra equalizer according to claim 1, characterized in that, The prediction based on the state vector of the previous time step, the covariance matrix of the previous time step, and the process noise covariance includes: Based on the state vector and unit state transition matrix of the previous time step, the prior state vector of the current time step is predicted. The calculation formula of the state prediction equation is as follows: ; in, This represents the unit state transition matrix; This represents the state vector of the previous time step; This represents the prior state vector at the current moment; By combining the process noise covariance to predict the covariance matrix of the previous time step, the prior covariance matrix of the current time step is obtained. The calculation formula of the covariance prediction equation is as follows: ; in, This represents the covariance matrix of the previous time step; Represents the prior covariance matrix at the current moment; This represents the process noise covariance matrix; Wherein, the process noise covariance matrix It decreases as the sample number increases, and Process noise figure The attenuation formula is: ; in, Indicates the initial process noise figure; Indicates the sample sequence number; Indicates the attenuation control parameters; This represents the identity matrix.
6. The method for sparsification and online updating of the Volterra equalizer according to claim 1, characterized in that, The variance estimation based on the prediction error to determine the observation noise variance at the current time includes: Set a smoothing factor; recursively estimate the observation noise variance at the current time based on the observed noise variance of the previous time step and the squared magnitude of the current prediction error, using the following formula: ; in, This represents the variance of the observation noise at the current moment; This represents the variance of the observation noise at the previous moment; This represents the smoothing factor; This indicates the current prediction error; The formula for calculating the Kalman gain is: ; in, Indicates the Kalman gain; Denotes the prior covariance matrix; This represents the current sparse regression vector; This indicates the preset weight parameters; This represents the variance of the observation noise.
7. The method for sparsification and online updating of the Volterra equalizer according to claim 1, characterized in that, The step of correcting the prior state vector using the Kalman gain and the prediction error includes: The prior state vector is corrected using the Kalman gain and the prediction error, and the calculation formula is as follows: ; in, This represents the updated state vector; Represents the prior state vector; Indicates the Kalman gain, This represents the prediction error; The prior covariance matrix is corrected based on the Kalman gain and the current sparse regression vector to obtain the updated covariance matrix, and the updated state vector is used as the updated coefficient of each kernel term in the retained kernel set.
8. A sparsity reduction and online update system for a Volterra equalizer, comprising a processor, a memory, and a computer program or instructions stored in the memory, characterized in that, The processor is configured to execute the computer program or instructions, and when the computer program or instructions are executed, the system implements the steps of the method as described in any one of claims 1 to 7.
9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method as described in any one of claims 1 to 7.
10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 7.
