Frequency domain non-linear dustpan line spline adaptive filtering method
Through the frequency domain nonlinear slime-twined spline adaptive filtering method, the generalized slime-twined line function and stochastic gradient operation are used to solve the problem of nonlinear adaptive filters under high computational complexity and non-Gaussian noise, and realizes the non-convergence of non-linear adaptive filters. Low complexity and high robustness filter design.
Patent Information
- Application Number
- CN202411001367.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-08
AI Technical Summary
Existing nonlinear adaptive filters have high computational complexity when processing large amounts of data or non-Gaussian noise environments, and may not converge under non-Gaussian noise, limiting their application in real-time and complex scenarios.
Adaptive filtering method for frequency domain nonlinear slime lines is adopted. By constructing a loss function based on generalized slime lines function, combining fast Fourier transform and stochastic gradient operation, iterative update method is optimized, reducing the computational complexity and improving the Lu Great.
It effectively reduces the computational complexity of the nonlinear adaptive filter, and maintains good convergence performance in non-Gaussian noise environments, significantly improving stability and real-time performance in complex scenarios.
Smart Images

Figure CN118921041A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of nonlinear adaptive filters, and in particular relates to a frequency domain nonlinear spur spline adaptive filtering method. Background Art
[0002] In the field of signal processing, nonlinear adaptive filters are an important system modeling method and are widely used in complex scenarios such as echo cancellation and channel equalization. Although traditional filtering methods, such as kernel filters and neural network filters, perform well in processing nonlinear signals, their computational complexity usually increases exponentially with the increase in the number of filter taps, which greatly limits their application in real-time or high-performance systems.
[0003] To overcome this challenge, researchers have been exploring filtering algorithms with lower computational complexity. Nonlinear spline filters have received widespread attention due to their simple structure. While maintaining certain filtering performance, this type of filter has lower computational complexity than other nonlinear filters. However, the computational complexity of nonlinear spline filters will still increase with the increase in the number of filter taps. For systems that need to process large amounts of data or have strict real-time requirements, this is still a problem that needs to be solved.
[0004] In recent years, some innovative filtering technologies have attracted the attention of the academic community, such as the paper "Frequency domain spline adaptive filters" published in the journal Signal Processing. This method proposes a method of applying Fourier transform and its inverse transform to the nonlinear spline adaptive filtering architecture. By introducing Fourier transform and its inverse transform, the nonlinear spline adaptive filtering architecture is optimized, which significantly reduces the computational complexity of the algorithm.
[0005] However, although this method has achieved remarkable results in reducing computational complexity, the algorithm may not converge in non-Gaussian noise environments, which limits its widespread use in practical applications.
[0006] Therefore, how to maintain the excellent performance of nonlinear adaptive filters while further reducing their computational complexity and improving their stability in non-Gaussian noise environments has become an urgent problem to be solved in the current signal processing field. Summary of the invention
[0007] The purpose of the present invention is to provide a frequency domain nonlinear sigmoid spline adaptive filtering method, which can reduce the computational complexity and still show good robustness in a non-Gaussian noise environment.
[0008] The technical solution adopted by the present invention is as follows:
[0009] A frequency domain nonlinear spur spline adaptive filtering method comprises the following steps:
[0010] A: Build a robust optimization strategy, including:
[0011] A1: Construct a loss function based on the generalized dustpan function;
[0012] A2: Solve the derivative of the loss function with respect to the error;
[0013] B: Construct the filter architecture of frequency domain nonlinear spline adaptive filtering method;
[0014] The frequency domain nonlinear spline adaptive filtering architecture consists of a frequency domain linear part and a time domain nonlinear part;
[0015] Included are:
[0016] B1: construct the frequency domain linear part;
[0017] B2: construct the time domain nonlinear part;
[0018] C: Iterative update method through stochastic gradient calculation;
[0019] C1: weight iterative update;
[0020] C2: Iterative update of control points.
[0021] The technical effects achieved by the present invention are:
[0022] The present invention performs block operations on a nonlinear filtering method through fast Fourier transform and its inverse transform, combines the FIR linear part with the spline basis function nonlinear part, and uses random gradient operations in the optimization iteration process through a spur line optimization strategy. A nonlinear adaptive filter with low computational complexity can be designed while effectively suppressing non-Gaussian noise interference, and can also maintain good convergence performance while reducing the algorithm computational complexity under Gaussian signals. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 It is a principle flow chart of the present invention;
[0024] Figure 2 It is the loss function diagram of the optimization strategy of the present invention;
[0025] Figure 3 is a gradient graph of the loss function of the optimization strategy of the present invention with respect to the error;
[0026] Figure 4It is a curve diagram of the computational complexity ratio between the frequency domain nonlinear slat spline adaptive filtering algorithm and the corresponding time domain algorithm in the present invention;
[0027] Figure 5 It is a comparison diagram of the mean square error MSE curve under non-Gaussian noise in the present invention;
[0028] Figure 6 It is a comparison diagram of the mean square error MSE curve under non-Gaussian noise in the present invention. DETAILED DESCRIPTION
[0029] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.
[0030] like Figure 1-6 As shown, a frequency domain nonlinear sieve spline adaptive filtering method, the filtering method includes: (1) constructing a frequency domain-based nonlinear spline adaptive filter; (2) using a random gradient algorithm to iteratively update the frequency domain nonlinear spline adaptive filter; (3) obtaining a loss function based on the sieve spline function as an optimization strategy for the frequency domain nonlinear spline adaptive filter. The present invention can reduce the computational complexity of the nonlinear spline adaptive filter, and is generally applicable to nonlinear system modeling due to the real-time characteristics of the filtering architecture. At the same time, it still has good convergence in a non-Gaussian noise environment due to the robustness of the optimization strategy of the inventive method;
[0031] The following steps are involved:
[0032] A: Build a robust optimization strategy, including:
[0033] A1: Construct a loss function based on the generalized dustpan function;
[0034] The generalized scimitar function is expressed as:
[0035]
[0036] Where e(n) represents the error, a>0 represents the radius of the circle generated by the scissor function, the center of mass of which is located at (0, a), r>0 represents the shape parameter of the generalized scissor function, and τ=(2a) -r ;
[0037] At the same time, as the parameter a decreases, the dustpan-tongue function becomes more and more gentle, and as the parameter a increases, the dustpan-tongue function becomes more and more steep;
[0038] The loss function based on the generalized helical function is expressed as:
[0039]
[0040] The cost function of the weight optimization strategy omits the constant factor 2a, and E[·] represents the expected operator symbol.
[0041] A2: Solve the derivative of the loss function with respect to the error;
[0042] The derivative of the loss function with respect to the error is expressed as:
[0043]
[0044] When there is non-Gaussian pulse interference in the environment, the error presents a larger value. Figure 2 Optimization strategy dustpan line loss function diagram and Figure 3 From the gradient graph of the error of the optimization strategy dustpan loss function, it can be observed that when there is a large outlier in the system error, the derivative of the loss function will tend to 0, thereby stopping the update of the loss function and suppressing the pulse interference with large outliers. When there is no non-Gaussian pulse interference with large outliers in the system or the interference is relatively small, the derivative of the loss function operates normally, and the loss function performs the adaptive strategy operation normally, thereby improving the robustness of the filtering method.
[0045] B: Construct the filter architecture of frequency domain nonlinear spline adaptive filtering method;
[0046] The frequency domain nonlinear spline adaptive filtering architecture consists of a frequency domain linear part and a time domain nonlinear part;
[0047] Included are:
[0048] B1: construct the frequency domain linear part;
[0049] The input data of the system is expressed as:
[0050] x(k)=[x(kM+1),x(kM+2),...,x(kM+M)] T
[0051] Using the 50% overlapping storage method, at the moment n=kM+M, the input signal is expressed as:
[0052]
[0053] Where k represents the data cache in the time domain, represents the data buffer in the frequency domain, and FFT[·] represents the fast Fourier transform;
[0054] Intermediate output of the linear section T(k), also used as the input of the nonlinear part, is transformed into the time domain through IFFT[·], i.e., inverse fast Fourier transform, and is expressed as:
[0055]
[0056] The last M elements of
[0057] Among them, ⊙ represents the element product of the matrix, Represents the weight of the linear part of the filter.
[0058] B2: construct the time domain nonlinear part;
[0059] According to the intermediate output of the linear part, the normalized horizontal coordinate u is obtained j , expressed as:
[0060]
[0061] And get the local interval index i j , expressed as:
[0062]
[0063] After the data enters the buffer, the horizontal coordinate u is normalized in the frequency domain j Vectorized to It is expressed as:
[0064] U(k)=[u1,u2,u3,…,u M ]
[0065] in, It means that the local lateral quantity is obtained according to the projection of the linear output s(kM+j);
[0066] Similar to the vectorized horizontal coordinate, the control point vectorization is defined as Represents the spline interpolation control point matrix, which is expressed in vector form as follows:
[0067] P(k)=[p i1 ,p i2 ,p i3 ,…,p iM ]
[0068] Among them, the control point is represented by p ij =[p ij-1 ,p ij ,p ij+1 ,p ij+2 ] T , according to the matrix definition of the buffer, the output vector y(k) of the system is obtained, which is expressed as:
[0069] y(k)=[y(kM+1),...,y(kM+M)] T
[0070] =(U T (k)·C·P(k)) ii ,i=1,2,...,M
[0071] =sum r (U T (k)·C⊙P T (k))
[0072] =sum c (C·P(k)⊙U(k)) T
[0073] Where y(kM+j) represents the output of the nonlinear partial spline interpolation at time n=kM+j, y(k) represents the M-element output from n=kM+1 to n=kM+M, and (·) ii represents a column vector consisting of the diagonal elements of the matrix, sum r (·) represents the column vector obtained by summing the elements of each row of the matrix, and the row vector obtained by summing the elements of each column of the matrix is c (·) indicates that C represents the CR spline basic matrix, which is expressed as:
[0074]
[0075] The system error is obtained from the expected signal and output signal of the unknown system and is expressed as:
[0076] e(k)=[e(kM+1),e(kM+2),...,e(kM+M)] T
[0077] =d(k)-y(k)
[0078] Among them, the system error at time kM+j is expressed as:
[0079] e(kM+j)=d(kM+j)-y(kM+j)
[0080] The expected signal of the buffer module is expressed as d(k) = [d(kM+1), d(kM+2), ..., d(kM+M)] T At moment kM+j.
[0081] C: According to the filtering architecture of the present invention, an iterative update method is obtained by stochastic gradient calculation, including:
[0082] C1: weight iterative update;
[0083] Linear part of the tongue cost function JMVC The stochastic gradient operation is performed on the weight w(k), and the result is expressed as:
[0084]
[0085] Among them, the back propagation vector is defined as:
[0086]
[0087] Converted to the frequency domain through fast Fourier transform, expressed as:
[0088]
[0089] in, is a zero vector of length M, and the time domain of the weight gradient vector is transformed from its frequency domain by inverse fast Fourier transform, expressed as:
[0090] The first M elements of
[0091] The matrix superscript H represents the Hermitian transpose, that is, the complex conjugate transpose, the inverse fast Fourier transform The length is 2M;
[0092] By fast Fourier transforming the difference between the iterative updates of the weights, we can obtain the update learning method in the weight frequency domain, which is expressed as:
[0093]
[0094] Among them, the learning rate of the frequency domain weight is given by μ w The zero vector is represented by And its length is M.
[0095] C2: Iterative update of control points.
[0096] The nonlinear part is iteratively updated about the control points, and the same spur function is used as the loss function of the frequency domain nonlinear spur spline adaptive filtering; the nonlinear part cost function J MVC About control point p i The derivative of is expressed as follows through stochastic gradient operation:
[0097]
[0098] The stochastic gradient operation of the nonlinear part cost function about the control point is vectorized and expressed as:
[0099]
[0100] Among them, the system error of the frequency domain nonlinear adaptive filter constructed based on the helical line is:
[0101]
[0102] The iterative update formula of the nonlinear partial spline interpolation control point matrix is:
[0103]
[0104] For convenience of representation, the constant term in the above formula is combined with the learning rate, μ p The learning rate expressed as a matrix of control points.
[0105] The present invention uses the computational complexity ratio to verify the effectiveness of the proposed algorithm in reducing computational complexity. For fairness, the same spur cost function is used in different filtering architectures. The computational complexity of the frequency domain nonlinear spur spline adaptive filtering method and the corresponding time domain filtering method is as follows: Figure 4 As shown, as the number of taps of the filtering method increases, the computational complexity tends to decrease, indicating that the computational complexity of the filtering method is effectively reduced.
[0106] According to the specific implementation of the present invention, nonlinear system identification simulation is performed to verify the performance of the filtering method of the present invention in a non-Gaussian noise environment with a large number of taps.
[0107] In this simulation, the input signal x(n) is a random process, uniformly distributed in [-1,1]. Non-Gaussian noise is represented by alpha stable distribution, and the number of samples is set to 100,000.
[0108] Implementation Case 1:
[0109] The unknown nonlinear system to be identified is expressed as:
[0110]
[0111] The nonlinear part consists of control points with an equal spacing of 0.2, and the number of taps is set to 100.
[0112] Can get as Figure 5 The MSE curve comparison diagram of each algorithm is shown, where FDSAF-MVC represents the filtering method of the present invention, and is compared with SAF-MVC representing the corresponding time domain method of the present invention and the existing FDSAF-LMS method. It can be seen that the FDSAF-LMS method does not converge, and the method of the present invention has better convergence performance in a non-Gaussian noise environment and effectively shortens the running time of the filtering method.
[0113] Implementation Case 2:
[0114] In the filtering method with a longer number of taps, implementation case 2 is designed to verify the effectiveness of the method designed by the present invention. In this case, the number of taps is set to 1000, and the linear part of the unknown system is expressed as:
[0115]
[0116] The nonlinear part is expressed as:
[0117] f(s(k))=arctan(2s(k))
[0118] Figure 6 The mean square error (MSE) curve of the FDSAF-MVC method proposed by the present invention and the corresponding time domain SAF-MVC method is shown in FIG. As can be seen from the figure, the method proposed by the present invention can suppress non-Gaussian interference in the case of a longer number of taps and effectively reduce the running time of the filtering method.
[0119] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.
Claims
1. A frequency domain nonlinear spline adaptive filtering method, It is characterized in that The following steps are involved: A: Build a robust optimization strategy, including: A1: Construct a loss function based on the generalized dustpan function; A2: Solve the derivative of the loss function with respect to the error; B: Construct the filter architecture of frequency domain nonlinear spline adaptive filtering method; The frequency domain nonlinear spline adaptive filtering architecture consists of a frequency domain linear part and a time domain nonlinear part; Included are: B1: construct the frequency domain linear part; B2: construct the time domain nonlinear part; C: Iterative update method through stochastic gradient calculation, including: C1: weight iterative update; C2: Iterative update of control points.
2. The adaptive filtering method according to claim 1, Features: In A1, the generalized helical function is expressed as: Where e(n) represents the error, a>0 represents the radius of the circle generated by the scissor function, the center of mass of which is located at (0, a), r>0 represents the shape parameter of the generalized scissor function, and τ=(2a) -r ; At the same time, as the parameter a decreases, the dustpan-tongue function becomes more and more gentle, and as the parameter a increases, the dustpan-tongue function becomes more and more steep; The loss function based on the generalized helical function is expressed as: The cost function of the weight optimization strategy omits the constant factor 2a, and E[·] represents the expected operator symbol.
3. The adaptive filtering method according to claim 1, Features: In A2, the derivative of the loss function with respect to the error is expressed as:
4. The adaptive filtering method according to claim 1, Features: In B1, the input data of the system is expressed as: x(k)=[x(kM+1),x(kM+2),...,x(kM+M)] T Using the 50% overlapping storage method, at the moment n=kM+M, the input signal is expressed as: Where k represents the data cache in the time domain, represents the data buffer in the frequency domain, and FFT[·] represents the fast Fourier transform; Intermediate output of the linear section T (k), also used as the input of the nonlinear part, is transformed into the time domain through IFFT[·], i.e., inverse fast Fourier transform, and is expressed as: The last M elements of Among them, ⊙ represents the element product of the matrix, Represents the weight of the linear part of the filter.
5. The adaptive filtering method according to claim 1, Features: In B2, the normalized horizontal coordinate u is obtained according to the intermediate output of the linear part. j , expressed as: And get the local interval index i j , expressed as: After the data enters the buffer, the horizontal coordinate u is normalized in the frequency domain j Vectorized to It is expressed as: U(k)=[u 1 ,in 2 ,in 3 ,…,in M ] in, It means that the local lateral quantity is obtained according to the projection of the linear output s(kM+j); Similar to the vectorized horizontal coordinate, the control point vectorization is defined as Represents the spline interpolation control point matrix, which is expressed in vector form as follows: P(k)=[p i1 ,p i2 ,p i3 ,…,p iM ] Among them, the control point is represented by p ij =[p ij-1 ,p ij ,p ij+1 ,p ij+2 ] T , according to the matrix definition of the buffer, the output vector y(k) of the system is obtained, which is expressed as: y(k)=[y(kM+1),...,y(kM+M)] T =(U T (k)·C·P(k)) ii ,i=1,2,...,M =sum r (U T (k)·C⊙P T (k)) =sum c (C·P(k)⊙U(k)) T Where y(kM+j) represents the output of the nonlinear partial spline interpolation at time n=kM+j, y(k) represents the M-element output from n=kM+1 to n=kM+M, and (·) ii represents a column vector consisting of the diagonal elements of the matrix, sum r (·) represents the column vector obtained by summing the elements of each row of the matrix, and the row vector obtained by summing the elements of each column of the matrix is c (·) indicates that C represents the CR spline basic matrix, which is expressed as:
6. The adaptive filtering method according to claim 1, Features: In C1, the linear part of the cost function J MVC The stochastic gradient operation is performed on the weight w(k), and the result is expressed as: Among them, the back propagation vector is defined as: Converted to the frequency domain through fast Fourier transform, expressed as: in, is a zero vector of length M, and the time domain of the weight gradient vector is transformed from its frequency domain by inverse fast Fourier transform, expressed as: The first M elements of The matrix superscript H represents the Hermitian transpose, that is, the complex conjugate transpose, the inverse fast Fourier transform The length is 2M; By fast Fourier transforming the difference between the iterative updates of the weights, we can obtain the update learning method in the weight frequency domain, which is expressed as: Among them, the learning rate of the frequency domain weight is given by μ w The zero vector is represented by And its length is M.
7. The adaptive filtering method according to claim 1, Features: In the nonlinear part of C2, the iterative update of the control points also uses the sigmoid function as the loss function of the frequency domain nonlinear sigmoid spline adaptive filtering; the nonlinear part cost function J MVC About control point p i The derivative of is expressed as follows through stochastic gradient operation: The stochastic gradient operation of the nonlinear part cost function about the control point is vectorized and expressed as: Among them, the system error of the frequency domain nonlinear adaptive filter constructed based on the helical line is: The iterative update formula of the nonlinear partial spline interpolation control point matrix is: Combining the constant term in the above formula with the learning rate, μ p The learning rate expressed as a matrix of control points.
8. The adaptive filtering method according to claim 5, Features: In B2, the system error is obtained by the expected signal and output signal of the unknown system, which is expressed as: e(k)=[e(kM+1),e(kM+2),...,e(kM+M)] T =d(k)-y(k) Among them, the system error at time kM+j is expressed as: e(kM+j)=d(kM+j)-y(kM+j) The expected signal of the buffer module is expressed as d(k) = [d(kM+1), d(kM+2), ..., d(kM+M)] T At moment kM+j.
