A bias compensation-based hybrid correlation entropy algorithm anti-impact noise method

Abstract

The application discloses a kind of anti-impact noise methods of mixed correlation entropy algorithm based on deviation compensation, comprising the following steps: signal is collected, and the noisy input signal of filter is obtained;Filter calculates the weight coefficient of current time k, and input signal passes through filter, and the filter output signal of current time k is obtained;S30, error signal is obtained based on reference signal through subtractor;Optimization cost function, introduce mixed correlation entropy criterion, by the convex combination of two different kernel width Gaussian functions as kernel function;Compensate input noise, by introducing unbiased criterion, the deviation amount caused by input error is derived in complex domain to compensate;S60, weight coefficient update;S70, repeat, minimize error signal, gradually approach reference signal, so as to achieve the effect of noise suppression.The application can effectively suppress the interference of input noisy and impact noise, improve the flexibility of algorithm and further improve the performance of algorithm.

Description

Technical Field

[0001] The present invention belongs to the technical field of noise processing, and in particular relates to an anti-impact noise method based on a hybrid correlation entropy algorithm with deviation compensation. Background Art

[0002] With the rapid development of modern economy and science and technology, various types of noise are ubiquitous and seriously impact human health. The noise we experience in everyday life is no longer simply Gaussian, but increasingly impulsive. Adaptive filtering algorithms have been widely applied in many signal processing applications. Real-domain applications include echo cancellation and noise control. Complex-domain applications include direction-of-arrival estimation and frequency estimation. Common filtering algorithms such as the least mean square algorithm, affine projection, and least squares method all achieve optimal estimation in the presence of Gaussian noise, assuming the input is uncontaminated. However, in practical applications, the noise distribution is not limited to Gaussian; impulsive noise can also be present. The input is also often contaminated by noise due to sampling errors, instrument errors, and other factors. In such cases, the steady-state performance of these algorithms deteriorates dramatically, and their convergence also deteriorates.

[0003] To address these issues, the correlation entropy criterion is considered a robust statistic, and the maximum correlation entropy criterion is an optimal criterion for handling non-Gaussian noise and has been successfully applied to adaptive filtering. However, the maximum correlation entropy algorithm is limited by kernel width. When the kernel width of a single kernel changes, its performance also changes significantly, making it unable to effectively suppress noise interference. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes an anti-impact noise method based on a hybrid correlation entropy algorithm with deviation compensation, which can effectively suppress the interference of input noise and impact noise, improve the flexibility of the algorithm and further enhance the performance of the algorithm.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation, comprising the steps of:

[0006] S10, collecting the signal to obtain the noisy input signal x(k) of the filter;

[0007] S20, the filter calculates and generates the weight coefficient ω(k) at the current time k, and the input signal x(k) passes through the filter to obtain the filter output signal y(k) at the current time k;

[0008] S30, obtaining an error signal e(k) based on the difference between the reference signal d0(k) and the output signal y(k) through a subtractor;

[0009] S40, optimize the cost function, introduce the mixed correlation entropy criterion, and use the convex combination of two Gaussian functions with different kernel widths as the kernel function to establish the relationship between the weight vector and the error function based on this cost function;

[0010] S50, compensating for input noise by introducing an unbiased criterion and deriving the deviation h(k) caused by the input error in the complex domain for compensation;

[0011] S60, weight coefficient update, through the iterative calculation of the update formula, the filter tap weight vector ω(k+1) at the next moment k+1 is obtained;

[0012] S70, repeating to minimize the error signal e(k) and gradually approaching the reference signal, thereby achieving the effect of noise suppression.

[0013] Furthermore, the signal is collected to obtain the input signal x(k) of the filter, wherein the input signal x(k) includes a noise-free signal and a noise signal, and x(k)=u(k)+n(k).

[0014] Furthermore, the filter coefficients generate the weight coefficients ω(k) of the filter at the current moment k, ω(k) = [ω1(k), ω2(k), ..., ω L-1 (k)] T , its initial value is zero, that is, ω(0)=0.

[0015] Furthermore, when establishing the optimization cost function, based on the mixed correlation entropy criterion, a convex combination of two Gaussian functions with different kernel widths is used as the kernel function, and a relationship between the weight coefficient and the error signal is established as the cost function, including the following steps:

[0016] The cost function of the algorithm based on mixed correlation entropy is constructed as:

[0017]

[0018] Among them, ω(k) represents the weight coefficient, E is the expected operation, α m Represents the mixing coefficient, m = 1, 2, and satisfies α1 + α2 = 1, 0 ≤ α m ≤1,σ m (m=1, 2) are the kernel widths of the two Gaussian kernel functions.

[0019] Furthermore, by establishing the relationship between the error and the weight coefficient, the weight coefficient is iteratively calculated to obtain the updated weight coefficient; the feedback is fed back to the filter input to adjust the error signal to minimize the error signal and gradually approach the reference signal; the relationship between the error and the updated weight coefficient is established as follows:

[0020]

[0021] Where μ is the step size factor of the algorithm update, 0<μ<1.

[0022] Furthermore, the deviation caused by the input noise is compensated by adding the deviation compensation vector h(k) to the relationship between the weight coefficient update, that is,

[0023]

[0024] Furthermore, by adopting an unbiased criterion, the deviation compensation vector h(k) is calculated, including the steps of:

[0025] Step 1: Define the weight vector error Δω(k) = ω(k) - ω0, where ω0 is the optimal solution of the algorithm;

[0026] Step 2: Define a noise-free error θ(k) = u0(k) + n0(k) - ω H (k)u(k)=n0(k)-Δω H (k)u(k);

[0027] Step 3: Based on Step 1 and Step 2, construct a relationship between the error vector e(k) and the error-free vector θ(k), e(k) = θ(k) - ω H (k)n(k);

[0028] Step 4: Assuming there is a deviation compensation vector h(k), the update formula for the weight vector error is:

[0029]

[0030] Step 5: The unbiased criterion formula is satisfied when E[Δω(k)|x(k)]=0. Substitute the formula in Step 4 to obtain

[0031] Step 6: Based on the deviation compensation amount obtained above, the weight coefficient update formula for adding the bias is written as

[0032]

[0033] in, is the variance of the input noise signal.

[0034] Furthermore, the variance estimation method of the input noise signal is:

[0035]

[0036]

[0037]

[0038] Where: γ is the input-output noise ratio, and λ is the forgetting factor.

[0039] The beneficial effects of adopting this technical solution are:

[0040] The present invention can effectively suppress the interference of input noise and impact noise, improve the flexibility of the algorithm and further enhance the performance of the algorithm. In order to expand the application scope of the above algorithm, the method proposed in the present invention is mainly based on the noise suppression of the signal in the complex domain. The algorithm in the proposed method has lower steady-state error and faster convergence. In addition, in practical applications, the presence of sampling error or modeling error causes the input signal to be contaminated by noise. In this case, by introducing the unbiased criterion, the deviation caused by the input error is derived in the complex domain for compensation. Even if the input contains noise, the algorithm of the present invention can compensate to a certain extent and reduce the impact of the input noise. This method not only solves the problem of input noise, but also can filter the impact noise, and has the characteristics of low steady-state error and fast convergence.

[0041] The present invention uses the entropy of a convex combination of two kernel functions as the cost function. An adaptive filtering algorithm adjusts and updates the tap coefficients, gradually bringing the output signal closer to the reference signal. Unlike traditional adaptive filtering algorithms, the present invention offers greater flexibility, lower steady-state error, and faster convergence. Although the present invention utilizes a convex combination of two kernel functions, the computational complexity does not double, but only increases by N times. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of a flow chart of a method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation according to the present invention;

[0043] Figure 2 Schematic diagram of the principle of a method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation according to the present invention;

[0044] Figure 3 2 is a comparison chart of the detection results of the embodiment of the present invention and the traditional method. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings.

[0046] In this embodiment, see Figure 1 and Figure 2 As shown, the present invention proposes a method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation, comprising the steps of:

[0047] S10, collecting the signal to obtain the noisy input signal x(k) of the filter;

[0048] S20, the filter calculates and generates the weight coefficient ω(k) at the current time k, and the input signal x(k) passes through the filter to obtain the filter output signal y(k) at the current time k;

[0049] S30, obtaining an error signal e(k) based on the difference between the reference signal d0(k) and the output signal y(k) through a subtractor;

[0050] S40, optimize the cost function, introduce the mixed correlation entropy criterion, and use the convex combination of two Gaussian functions with different kernel widths as the kernel function to establish the relationship between the weight vector and the error function based on this cost function;

[0051] S50, compensating for input noise by introducing an unbiased criterion and deriving the deviation h(k) caused by the input error in the complex domain for compensation;

[0052] S60, weight coefficient update, through the iterative calculation of the update formula, the filter tap weight vector ω(k+1) at the next moment k+1 is obtained;

[0053] S70, repeating to minimize the error signal e(k) and gradually approaching the reference signal, thereby achieving the effect of noise suppression.

[0054] As an optimization solution for the above embodiment, in step S10, the signal is collected to obtain the filter input signal x(k). The input signal x(k) includes a noise-free signal and a noise signal, x(k) = u(k) + n(k). Correspondingly, d0(k) = u0(k) + n0(k), where u0(k) is a noise-free reference signal and n0(k) is a noise reference signal.

[0055] As an optimization solution of the above embodiment, in step S20, the filter coefficient generates the weight coefficient ω(k) of the filter at the current moment k, ω(k)=[ω1(k), ω2(k), ..., ω L-1 (k)] T , its initial value is zero, that is, ω(0) = 0, L represents the tap length of the adaptive filter, which can be 16, 32 or 64.

[0056] As an optimization solution of the above embodiment, in step S30, the error signal e(k)=d0(k)-y(k) is obtained;

[0057] As an optimization solution for the above embodiment, in step S40, the cost function is optimized and the hybrid correlation entropy criterion is introduced. The convex combination of two Gaussian functions with different kernel widths is used as the kernel function, and the relationship between the weight vector and the error function is established based on this as the cost function. The steps include:

[0058] The cost function of the algorithm based on mixed correlation entropy is constructed as:

[0059]

[0060] Among them, ω(k) represents the weight coefficient, E is the expected operation, α m Represents the mixing coefficient, m = 1, 2, and satisfies α1 + α2 = 1, 0 ≤ α m ≤1,σ m (m=1, 2) are the kernel widths of the two Gaussian kernel functions.

[0061] By establishing the relationship between the error and the weight coefficient, the weight coefficient is iterated to obtain the updated weight coefficient; the feedback is sent to the filter input to adjust the error signal to minimize the error signal and gradually approach the reference signal; the relationship between the error and the updated weight coefficient is established as follows:

[0062]

[0063] Where μ is the step size factor of the algorithm update, 0<μ<1.

[0064] As an optimization solution of the above embodiment, in step S50, in order to deal with the biased estimation generated when the input signal has noise, in order to solve this bias problem, the bias compensation vector h(k) is added to the relationship between the weight coefficient update to compensate for the bias caused by the input noise, that is,

[0065]

[0066] By adopting the unbiased criterion, the deviation compensation vector h(k) is calculated, including the steps:

[0067] Step 1: Define the weight vector error Δω(k) = ω(k) - ω0, where ω0 is the optimal solution of the algorithm;

[0068] Step 2: Define a noise-free error θ(k) = u0(k) + n0(k) - ω H (k)u(k)=n0(k)-Δω H (k)u(k);

[0069] Step 3: Based on Step 1 and Step 2, construct a relationship between the error vector e(k) and the error-free vector θ(k), e(k) = θ(k) - ω H (k)n(k);

[0070] Step 4: Assuming there is a deviation compensation vector h(k), the update formula for the weight vector error is:

[0071]

[0072] Step 5: The unbiased criterion formula is satisfied when E[Δω(k)|x(k)]=0. Substitute the formula in Step 4 to obtain

[0073] The specific derivation formula is as follows:

[0074]

[0075] in, is the variance of the input noise signal.

[0076] Preferably, the input noise variance of the bias compensation is added It is often unknown, so theoretical estimation is required. The variance estimation method of the input noise signal is:

[0077]

[0078]

[0079]

[0080] Where: γ is the input-output noise ratio, and λ is the forgetting factor.

[0081] As an optimization solution of the above embodiment, in step S60, the weight coefficient update formula can be expressed as:

[0082]

[0083] As attached Figure 3 The figure below shows a comparison of the MSD of the proposed method for impact noise mitigation using the hybrid correlation entropy algorithm with bias compensation. The top curve in the figure shows the results of the proposed method, while the two curves below represent the traditional MCC algorithm and the BCMCC algorithm, respectively. Due to the multi-core width and bias compensation terms incorporated into the proposed method, its steady-state performance significantly outperforms the other two algorithms.

[0084] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation, characterized in that: Including steps: S10, collecting the signal to obtain the noisy input signal χ(k) of the filter; S20, the filter calculates and generates the weight coefficient ω(k) at the current time k, and the input signal x(k) passes through the filter to obtain the filter output signal y(k) at the current time k; S30, obtaining an error signal e(k) based on the difference between the reference signal d0(k) and the output signal y(k) through a subtractor; S40, optimize the cost function, introduce the mixed correlation entropy criterion, and use the convex combination of two Gaussian functions with different kernel widths as the kernel function to establish the relationship between the weight vector and the error function based on this cost function, including the following steps: The cost function of the algorithm based on mixed correlation entropy is constructed as: Among them, ω(k) represents the weight coefficient, E is the expected operation, α m Represents the mixing coefficient, m=1,2, and satisfies α1+α2=1, 0≤α m ≤1,σ m (m=1,2) is the kernel width of the two Gaussian kernel functions; S50, compensating for input noise by introducing an unbiased criterion and deriving the deviation h(k) caused by the input error in the complex domain for compensation; By adopting the unbiased criterion, the deviation compensation vector h(k) is calculated, including the steps: Step 1: Define the weight vector error Δω(k) = ω(k) - ω0, where ω0 is the optimal solution of the algorithm; Step 2: Define a noise-free error θ(k) = u0(k) + n0(k) - ω H (k)u(k)=n0(k)-Δω H (k)u(k); Step 3: Based on Step 1 and Step 2, construct a relationship between the error vector e(k) and the error-free vector θ(k), e(k) = θ(k) - ω H (k)n(k); Step 4: Assuming there is a deviation compensation vector h(k), the update formula for the weight vector error is: Step 5: When E[Δω(k)|χ(k)]=0, E(Δω(k+1)|x(k)]=0; Substitute the formula in Step 4 to obtain Step 6: Based on the above deviation compensation vector, the weight coefficient update formula for adding the bias is written as in, is the variance of the input noise signal; S60, weight coefficient update, through the iterative calculation of the update formula, the filter tap weight vector ω(k+1) at the next moment k+1 is obtained; S70, repeating to minimize the error signal e(k) and gradually approaching the reference signal, thereby achieving the effect of noise suppression.

2. The method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation according to claim 1, characterized in that: The signal is collected to obtain the input signal x(k) of the filter, wherein the input signal x(k) includes a noise-free signal and a noise signal, and x(k)=u(k)+n(k).

3. The method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation according to claim 1, characterized in that: The filter coefficient generates the weight coefficient ω(k) of the filter at the current time k, ω(k)=[ω1(k),ω2(k),…,ω L-1 (k)] T , its initial value is zero, that is, ω(0)=0.

4. The method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation according to claim 1, characterized in that: By establishing the relationship between the error and the weight coefficient, the weight coefficient is iterated to obtain the updated weight coefficient; the feedback is sent to the filter input to adjust the error signal to minimize the error signal and gradually approach the reference signal; the relationship between the error and the updated weight coefficient is established as follows: Where μ is the step size factor of the algorithm update, 0<μ<1.

5. The method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation according to claim 4, characterized in that: The deviation caused by input noise is compensated by adding the deviation compensation vector h(k) to the relationship between weight coefficient update, that is, 6. The method for resisting impact noise based on a hybrid correlation entropy algorithm with deviation compensation according to claim 1, characterized in that: The variance estimation method of the input noise signal is: Where: γ is the input-output noise ratio, and λ is the forgetting factor.

Images