Heuristic preprocessing nonlinear spline adaptive filtering identification system and method
Through heuristic preprocessing and maximum correlation entropy criterion, the nonlinear spline filtering model is optimized, and the problems of high computational complexity and poor interference adaptability of the nonlinear adaptive filtering algorithm are solved, achieving faster convergence speed and wider application scenarios .
Patent Information
- Application Number
- CN202411001482.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-22
AI Technical Summary
The existing nonlinear adaptive filtering algorithms have high computational complexity when processing nonlinear systems and are difficult to adapt to interference in the actual environment, resulting in unsatisfactory convergence effect, especially in the presence of pulse interference.
Using heuristic preprocessing method, first optimize the weight parameters of the nonlinear system through a linear FIR filter to reduce the computational complexity, and iteratively optimize the filtering system using the maximum correlation entropy criterion and stochastic gradient algorithm to improve the estimation of the nonlinear spline filtering model Accuracy.
It improves the convergence speed and robustness of the system, can better adapt to pulse interference in the actual environment, and expands the application scope of nonlinear system identification.
Smart Images

Figure CN119010841A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of nonlinear adaptive filtering system design, and in particular relates to a heuristic preprocessing nonlinear spline adaptive filtering identification system and method. Background Art
[0002] System identification is divided into linear systems and nonlinear systems according to whether the system conforms to the superposition principle; nonlinear elements are ubiquitous in real life; system identification methods based on idealized and simplified linear system modeling cannot accurately describe the working mechanism of nonlinear systems, therefore, nonlinear system identification methods have received widespread attention; filtering algorithms, as a more general system identification method, have been further studied.
[0003] The adaptive filtering algorithms for nonlinear structures are complex and diverse, and there is no universal theoretical system architecture, so various types of nonlinear filtering algorithms have been designed; among several widely popular nonlinear adaptive filtering algorithms: the polynomial filtering algorithm based on Volterra series expansion, its computational complexity grows in the form of geometric series as the nonlinear order increases, and its low-order nonlinear structure model is usually used as a filtering algorithm to process nonlinear systems in practical applications; the filtering algorithm based on neural network has been valued as a general nonlinear model in nonlinear signal processing, but the main disadvantage of the filtering algorithm based on neural network model is that it requires a very high computational complexity during the implementation process; the kernel adaptive filtering algorithm with a simple architecture is relatively easy to implement, but the network structure based on the kernel adaptive filtering algorithm will have the problem of increasing computational complexity as the number of iterations increases; in recent years, the adaptive filtering algorithm based on nonlinear spline structure design has excellent performance in describing nonlinear systems, and it has a simple architecture itself; however, the adaptive filtering algorithm of nonlinear spline structure has the problem of estimating approximation performance that needs to be solved; it can be seen that exploring nonlinear adaptive filters with low complexity and strong nonlinear estimation ability is still a recognized core problem in the field of nonlinear system identification.
[0004] When designing a nonlinear adaptive filtering algorithm, in order to reduce the computational complexity of the filtering algorithm, it is usually assumed that the research on the filtering algorithm is based on a Gaussian environment; however, in actual engineering applications, the system will inevitably have correlated interference that destroys the Gaussian assumption due to its own faults or external uncertain factors; in the presence of interference, the adaptive filtering algorithm based on the Gaussian distribution assumption does not match the interference environment in actual application, resulting in unsatisfactory convergence of the filtering algorithm; therefore, for the identification system of practical significance, it is worth studying how to design an adaptive filtering algorithm for nonlinear structures with pulse interference so that the system identification has better performance; for this purpose, we propose a heuristic preprocessing nonlinear spline adaptive filtering identification system and method. Summary of the invention
[0005] The purpose of the present invention is to provide a heuristic preprocessing nonlinear spline adaptive filtering identification system and method. First, a preprocessing operation is performed. Compared with the traditional nonlinear filtering system, in order to improve the estimation accuracy of the nonlinear filtering model, the weight parameters of the filtering system are initialized before the nonlinear filtering is performed; during the initialization process, the entire nonlinear system is iteratively optimized only through the linear part; invariance is a commonly used property in nonlinear system identification. This property makes it possible to identify the linear part of the Wiener model without compensating or estimating the static nonlinearity in the model; invariance is that the cross-correlation function obtained after the signal undergoes nonlinear amplitude distortion is the same as the cross-correlation function obtained before the distortion except for the scale factor. ; According to the invariant characteristics of the filtering system, before the nonlinear spline filtering process is carried out, the weight items of the linear part of the filtering system need to be preprocessed and optimized first; the step of optimizing the weights first uses only the linear FIR filter to identify the entire unknown system, that is, the weights of the FIR filter are used to optimize the entire unknown system first; more specifically, according to the invariant characteristics of the nonlinear filtering system, the linear and nonlinear parts of the unknown system are optimized only by the linear FIR filter, and then the weights of the initial optimized FIR filter of the entire system are obtained; therefore, the entire unknown system is first optimized only by the FIR filter to obtain the initial value of the weight, and finally the nonlinear spline filtering model after iterative optimization is obtained;
[0006] Then in the preprocessing stage, the output of the FIR filter in the iterative optimization weight step for the nonlinear unknown system is expressed as:
[0007] s (0) (n) = w (0) (n)x(n);
[0008] Among them, the input signal of the preprocessing operation weight process is consistent with the input signal in the nonlinear filtering system and is defined as:
[0009] x(n)=[x(n),x(n-1),...,x(n-M+1)] T ,
[0010] The weight vector s (0) (n) is the output representing the iterative update weight of the linear part; finally, the parameters are adjusted continuously by iteration until the identification of the set nonlinear spline filter model is completed; and the final output result is y(n).
[0011] The technical solution adopted by the present invention is as follows:
[0012] A heuristic preprocessing nonlinear spline adaptive filtering identification system comprises a preprocessing module, an optimization processing module and an output module. The preprocessing module first obtains an input signal x(n); and preprocesses the input signal x(n) to obtain an output signal y(n), an error signal e(n) and an expected signal d(n); a nonlinear spline filtering model is constructed inside the optimization processing module; the nonlinear spline filtering model is used to adjust the control point of the filtering identification system according to the input signal x(n) and the system error, and then the initialization intermediate output is obtained, and the system error and the system weight are combined to calculate the system error according to the expected signal d(n) and the output signal y(n); finally, an adaptive optimization strategy is designed according to the system error, and finally the input signal x(n) is optimized by the adaptive optimization strategy to obtain the output signal y(n).
[0013] A heuristic preprocessing nonlinear spline adaptive filtering identification method specifically comprises the following steps:
[0014] S1: First, construct a nonlinear spline filter model;
[0015] S101: first construct the nonlinear filtering part;
[0016] S102: constructing the linear part;
[0017] S2: preprocess the input signal x(n);
[0018] S3: Calculate the system error;
[0019] S4: Iteratively optimize the control points and weights in the filtering system using the maximum correlation entropy criterion;
[0020] S5: Use the stochastic gradient algorithm to iterate the weights of the preprocessing process and the control points and weights of the filtering process;
[0021] S6: Perform multiple iterations and optimizations to finally obtain an output signal y(n).
[0022] Preferably, in step S101, at time n, the input signal of the nonlinear filtering system is x(n), which is expanded to be:
[0023] x(n)=[x(n),x(n-1),...,x(n-M+1)] T ;
[0024] The number of taps of the filter is represented by M; the input signal x(n) is obtained by the lookup table module to obtain the local parameter U in the nonlinear part representation. n and index i:
[0025]
[0026] Where Δx is the uniform interval between two control points, and Q represents the number of control points in nonlinear spline interpolation. Based on the local parameters and index, the nonlinear part output s(n) is obtained:
[0027] s(n)=u(n) T Cp i (n);
[0028] In the formula, C is represented as the spline basis matrix, where p i =[p x,i ,p s,i ] T is the control point, p x,i Indicates the corresponding x-axis coordinate of the control point. The value range of i is 0-N, indicating N+1 control points.
[0029] Preferably, in step S102; the nonlinear spline part in the nonlinear spline filtering model is connected to the linear filtering part; the nonlinear part output s(n) is used as the input of the linear filtering part, and the output y(n) of the linear part is obtained through an FIR filter, and the output y(n) of the linear part is used as the output of the nonlinear spline filtering model.
[0030] Preferably, regarding the linear part in the nonlinear spline filtering model, the weight of the FIR filter constructed therein is defined as w(n)=[w(0),w(1),…,w(n-1)] T ;
[0031] y(n)=w T (n)s(n).
[0032] Preferably, in step S2, firstly, an input signal x(n) is obtained; and the input signal x(n)x(n) is preprocessed to obtain an output signal y(n), an error signal e(n) and a desired signal d(n).
[0033] Preferably, in step S3, the system error is obtained by subtracting the system output S obtained by iteratively updating the weights of the linear part from the system expected signal d(n). (0)(n) Obtain the preprocessing operating system error signal e*(n):
[0034] e*(n)=d(n)-S (0) (n);
[0035] In the nonlinear filtering process, the system error e(n) is obtained by subtracting the output signal y(n) from the system expected signal d(n):
[0036] Preferably, in step S4, the control points and weights in the nonlinear spline filter model and the weights in the preprocessing operation are iterated using the maximum relevant entropy criterion; the maximum relevant entropy criterion cost function is expressed as:
[0037]
[0038] Where σ>0 represents the kernel length parameter; in the preprocessing operation, the cost function is based on the system error signal e*(n), and in the filtering process, the loss function is represented by the system error e(n).
[0039] Preferably, in step S5, a random gradient algorithm is used to iterate the weights of the preprocessing process and the control points and weights of the filtering process; in the preprocessing priority optimization weight step, a simple structured maximum correlation entropy adaptive strategy is used to iteratively update the weights of the FIR filter, thereby obtaining an expression for iterative update of the preprocessing priority optimization FIR filter weights:
[0040] w (0) (n+1)=w (0) (n+1)+μe*(n)x(n);
[0041] The learning rate μ of the weights is about the FIR filter during preprocessing;
[0042] w(n+1)=w(n)+μ w e*(n)s(n);
[0043] p(n+1)=p(n)+μ p e*(n)C T u(n)w(n).
[0044] The technical effects achieved by the present invention are:
[0045] In the present invention, the preprocessing process improves the convergence speed of the system: the present invention adopts a preprocessing operation. Before the nonlinear filtering identification system, the parameters obtained by preprocessing can be directly used as the initial values in the nonlinear process, thereby effectively improving the convergence speed of the identification process.
[0046] In the present invention, the maximum correlation entropy criterion is robust: the cost functions used in the preprocessing process and the nonlinear filtering process of the present invention are all robust maximum correlation entropy criterion, thereby, the constructed nonlinear system identification can better adapt to the pulse interference existing in the environment.
[0047] In the present invention, the scope of application is improved: the heuristic nonlinear filtering system identification method proposed in the present invention has a wider range of application scenarios, because compared with the linear system identification description, the present invention is closer to the nonlinear description of the actual operating environment; the present invention designs a preprocessing operation to improve the convergence speed of the identification system; the present invention adopts a cost function with robustness, and takes the actual existing pulse interference environment into consideration. Therefore, the scope of application of the present invention is improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is a system block diagram of a heuristic preprocessing nonlinear spline adaptive filtering identification system in the first embodiment of the present invention;
[0049] Figure 2 It is a flow chart of a heuristic preprocessing nonlinear spline adaptive filtering identification method in the second embodiment of the present invention;
[0050] Figure 3 is a schematic diagram of the preprocessing operation of the adaptive filter in the system identification application in the first embodiment of the present invention;
[0051] Figure 4 is a schematic diagram of the filtering process operation of the adaptive filter in the system identification application in the first embodiment of the present invention;
[0052] Figure 5 is a schematic diagram of the cascade connection of an adaptive filter in Embodiment 2 of the present invention;
[0053] Figure 6 is a diagram of a nonlinear spline adaptive filtering architecture in Embodiment 2 of the present invention;
[0054] Figure 7 is an iterative cycle flow chart of the heuristic preprocessing nonlinear spline adaptive filtering method in the second embodiment of the present invention;
[0055] Figure 8 is a gradient graph of the maximum relevant entropy criterion loss function with respect to the error in the second embodiment of the present invention;
[0056] Fig. 9 It is an MMSE curve diagram of system identification in the second embodiment of the present invention. DETAILED DESCRIPTION
[0057] 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.
[0058] like Figure 1 as well as Figure 3 As shown, a heuristic preprocessing nonlinear spline adaptive filtering identification system includes a preprocessing module, an optimization processing module and an output module. The preprocessing module first obtains an input signal x(n); and preprocesses the input signal x(n) to obtain an output signal y(n), an error signal e(n) and an expected signal d(n); a nonlinear spline filtering model is constructed inside the optimization processing module; the nonlinear spline filtering model is used to adjust the control point of the filtering identification system according to the input signal x(n) and the system error, and then the initialization intermediate output is obtained, combined with the system error and the system weight, and then the system error is calculated according to the expected signal d(n) and the output signal y(n); finally, an adaptive optimization strategy is designed according to the system error, and finally the input signal x(n) is optimized using the adaptive optimization strategy to obtain the output signal y(n).
[0059] In the present invention, the preprocessing process improves the convergence speed of the system: the present invention adopts a preprocessing operation. Before the nonlinear filtering identification system, the parameters obtained by preprocessing can be directly used as the initial values in the nonlinear process, thereby effectively improving the convergence speed of the identification process.
[0060] Embodiment 2
[0061] like Figure 2 as well as Figure 7 As shown, a heuristic preprocessing nonlinear spline adaptive filtering identification method specifically includes the following steps:
[0062] S1: First, construct a nonlinear spline filter model;
[0063] S101: first construct the nonlinear filtering part;
[0064] S102: constructing the linear part;
[0065] S2: preprocess the input signal x(n);
[0066] S3: Calculate the system error;
[0067] S4: Iteratively optimize the control points and weights in the filtering system using the maximum correlation entropy criterion;
[0068] S5: Use the stochastic gradient algorithm to iterate the weights of the preprocessing process and the control points and weights of the filtering process;
[0069] S6: Perform multiple iterations and optimizations to finally obtain an output signal y(n).
[0070] like Figure 4 As shown, in step S101, at time n, the input signal of the nonlinear filtering system is x(n), which is expanded to be:
[0071] x(n)=[x(n),x(n-1),...,x(n-M+1)] T ;
[0072] The number of taps of the filter is represented by M; the input signal x(n) is obtained by the lookup table module to obtain the local parameter U in the nonlinear part representation. n and index i:
[0073]
[0074] Where Δx is the uniform interval between two control points, and Q represents the number of control points in nonlinear spline interpolation. Based on the local parameters and index, the nonlinear part output s(n) is obtained:
[0075] s(n)=u(n) T Cp i (n);
[0076] In the formula, C is represented as the spline basis matrix, where p i =[p x,i ,p s,i ] T is the control point, p x,i Indicates the corresponding x-axis coordinate of the control point. The value range of i is 0-N, indicating N+1 control points. Taking four adjacent control points as an example, the horizontal coordinates are normalized:
[0077] u(n)=[u 3 ,u 2 ,u,1] T ;
[0078] The basic matrix is represented by C. There are two commonly used matrices: B-spline and CR-spline. The present invention selects and applies B-spline to represent the spline basis C matrix according to actual conditions.
[0079] like Figure 5 As shown in Figure 6, it is shown that the nonlinear spline filtering model of the present invention is composed of a nonlinear spline part cascaded with a linear filtering part. In step S102, the nonlinear spline part in the nonlinear spline filtering model is connected to the linear filtering part. The nonlinear part output s(n) is used as the input of the linear filtering part, and the output y(n) of the linear part is obtained through an FIR filter. The output y(n) of the linear part is used as the output of the nonlinear spline filtering model.
[0080] Preferably, regarding the linear part in the nonlinear spline filtering model, the weight of the FIR filter constructed therein is defined as w(n)=[w(0),w(1),…,w(n-1)] T ;
[0081] y(n)=w T (n)s(n).
[0082] This nonlinear spline filtering model uses the nonlinear part output s(n) as the input of the linear part; the representation of this part of the output is consistent with the system output calculation method in the system preprocessing operation, and the preprocessing parameters obtained are directly applied to the filtering system, thereby reducing the number of convergence iterations of the filtering system.
[0083] In the present invention, the maximum correlation entropy criterion is robust: the cost functions used in the preprocessing process and the nonlinear filtering process of the present invention are all robust maximum correlation entropy criterion, thereby, the constructed nonlinear system identification can better adapt to the pulse interference existing in the environment
[0084] like Figure 1 As shown, in step S2, firstly, an input signal x(n) is obtained; and the input signal x(n)x(n) is preprocessed to obtain an output signal y(n), an error signal e(n) and a desired signal d(n).
[0085] like Figure 6 as well as Figure 8 As shown, in step S3, the system error is the system expected signal d(n) minus the system output S which is only updated by the linear part iteratively. (0) (n) Obtain the preprocessing operating system error signal e*(n):
[0086] e*(n)=d(n)-S (0) (n);
[0087] In the nonlinear filtering process, the system error e(n) is obtained by subtracting the output signal y(n) from the system expected signal d(n):
[0088] Preferably, in step S4, the control points and weights in the nonlinear spline filter model and the weights in the preprocessing operation are iterated using the maximum relevant entropy criterion; Figure 6 As shown, the maximum relevant entropy criterion cost function is expressed as:
[0089]
[0090] Where σ>0 represents the kernel length parameter; in the preprocessing operation, the cost function is based on the system error signal e*(n), and in the filtering process, the loss function is represented by the system error e(n).
[0091] like Figure 6 As shown, in step S5, the random gradient algorithm is used to iterate the weights of the preprocessing process and the control points and weights of the filtering process; in the preprocessing priority optimization weight step, the maximum relevant entropy adaptive strategy with a simple structure is used to iteratively update the weights of the FIR filter, thereby obtaining the expression for iterative update of the preprocessing priority optimization FIR filter weights:
[0092] w (0) (n+1)=w (0) (n+1)+μe*(n)x(n);
[0093] The learning rate μ of the weights is about the FIR filter during preprocessing;
[0094] w(n+1)=w(n)+μ w e*(n)s(n);
[0095] p(n+1)=p(n)+μ p e*(n)C T u(n)w(n).
[0096] In the system identification process, the iterative cycle of the heuristic preprocessing nonlinear spline adaptive filtering method is as follows: Figure 7 As shown, keep adjusting the parameters until the system identification is completed. Fig. 9 System identification results using heuristic nonlinear spline adaptive filtering method.
[0097] The heuristic nonlinear filtering system identification method proposed in the present invention has a wider range of application scenarios. This is because compared with the linear system identification description, the present invention is closer to the nonlinear description of the actual operating environment; the present invention designs a preprocessing operation to improve the convergence speed of the identification system; the present invention adopts a robust cost function; and the actual existing pulse interference environment is taken into consideration; thereby, the application scope of the present invention is improved.
[0098] 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 heuristic preprocessing nonlinear spline adaptive filtering identification system, characterized by: It includes a preprocessing module, an optimization processing module and an output module. The preprocessing module first obtains an input signal x(n); and preprocesses the input signal x(n) to obtain an output signal y(n), an error signal e(n) and a desired signal d(n); a nonlinear spline filtering model is constructed inside the optimization processing module; A nonlinear spline filter model is used to adjust the control points of the filter identification system according to the input signal x(n) and the system error, and then the initial intermediate output is obtained. The system error and the system weight are combined, and the system error is calculated according to the expected signal d(n) and the output signal y(n). Finally, an adaptive optimization strategy is designed based on the system error. Finally, the adaptive optimization strategy is used to optimize the input signal x(n) to obtain the output signal y(n).
2. A heuristic preprocessing nonlinear spline adaptive filtering identification method, characterized by: The filter identification method is the operating method of the filter identification system in claim 1, and specifically comprises the following steps: S1: First, construct a nonlinear spline filter model; S101: first construct the nonlinear filtering part; S102: constructing the linear part; S2: preprocess the input signal x(n); S3: Calculate the system error; S4: Iteratively optimize the control points and weights in the filtering system using the maximum correlation entropy criterion; S5: Use the stochastic gradient algorithm to iterate the weights of the preprocessing process and the control points and weights of the filtering process; S6: Perform multiple iterations and optimizations to finally obtain an output signal y(n).
3. The heuristic preprocessing nonlinear spline adaptive filtering identification method according to claim 1, characterized in that: In step S101, at time n, the input signal of the nonlinear filtering system is x(n), which can be expanded to be: x(n)=[x(n),x(n-1),...,x(n-M+1)] T ; The number of taps of the filter is represented by M; the input signal x(n) is obtained by the lookup table module to obtain the local parameter U in the nonlinear part representation. n and index i: Where Δx is the uniform interval between two control points, Q represents the number of control points in nonlinear spline interpolation, and the nonlinear part output s(n) is obtained based on the local parameters and index: s(n)=u(n) T Cp i (n); In the formula, C is represented as the spline basis matrix, where p i =[p x,i ,p s,i ] T is the control point, p x,i Indicates the corresponding x-axis coordinate of the control point. The value range of i is 0-N, indicating N+1 control points.
4. The heuristic preprocessing nonlinear spline adaptive filtering identification method according to claim 1, characterized in that: In step S102, the nonlinear spline part in the nonlinear spline filtering model is connected to the linear filtering part; the nonlinear part output s(n) is used as the input of the linear filtering part, and the output y(n) of the linear part is obtained through the FIR filter, and the output y(n) of the linear part is used as the output of the nonlinear spline filtering model.
5. The heuristic preprocessing nonlinear spline adaptive filtering identification system according to claim 4, characterized in that: Regarding the linear part of the nonlinear spline filtering model, the weight of the constructed FIR filter is defined as: w(n)=[w(0),w(1),…,w(n-1)] T ; y(n)=w T (n)s(n)。 6. The heuristic preprocessing nonlinear spline adaptive filtering identification method according to claim 1, characterized in that: In step S2, firstly, an input signal x(n) is obtained; and the input signal x(n)x(n) is preprocessed to obtain an output signal y(n), an error signal e(n) and a desired signal d(n).
7. A heuristic preprocessing nonlinear spline adaptive filtering identification method according to claim 6, characterized in that: In step S3, the system error is the system expected signal d(n) minus the system output S which is only updated by the linear part iteratively. (0) (n) Obtain the preprocessing operating system error signal e*(n): e*(n)=d(n)-S (0) (n); In the nonlinear filtering process, the system error e(n) is obtained by subtracting the output signal y(n) from the system expected signal d(n): e*(n)=d(n)-y(n).
8. A heuristic preprocessing nonlinear spline adaptive filtering identification method according to claim 7, characterized in that: In step S4, the control points and weights in the nonlinear spline filter model and the weights in the preprocessing operation are iterated using the maximum relevant entropy criterion; the maximum relevant entropy criterion cost function is expressed as: Where σ>0 represents the kernel length parameter; in the preprocessing operation, the cost function is based on the system error signal e*(n), and in the filtering process, the loss function is represented by the system error e(n).
9. A heuristic preprocessing nonlinear spline adaptive filtering identification method according to claim 8, characterized in that: In step S5, the random gradient algorithm is used to iterate the weights of the preprocessing process and the control points and weights of the filtering process; in the preprocessing priority optimization weight step, the maximum relevant entropy adaptive strategy with a simple structure is used to iteratively update the weights of the FIR filter, thereby obtaining the expression for iterative update of the preprocessing priority optimization FIR filter weights: w (0) (n+1)=w (0) (n+1)+μe*(n)x(n); The learning rate μ of the weights is about the FIR filter during preprocessing; w(n+1)=w(n)+µ w e*(n)s(n): p(n+1)=p(n)+µ p and*(n)C T u(n)w(n)。
