Noise suppression method in line spectrum feature extraction based on double singular value decomposition

Abstract

The invention discloses a noise suppression method in line spectrum feature extraction based on double singular value decomposition. Relates to a method for suppressing noise of a target in water. Strong background noise interference of a target in water can be suppressed, and line spectrum features can be extracted. The method comprises the following steps: carrying out segmentation processing on data received by a sonar sensor, solving a frequency spectrum matrix and a normalized frequency spectrum matrix for the data, respectively carrying out singular value decomposition on the frequency spectrum matrix and the normalized frequency spectrum matrix, constructing a weighting coefficient by utilizing a maximum singular value and a minimum singular value, and weighting a Welch method power spectrum to suppress noise. The method provided by the invention has the capability of resisting strong noise interference, and can extract the line spectrum characteristics under strong background noise interference; the operation is simple, and the method has good self-adaptability to noise suppression in line spectrum feature extraction; the calculation amount is small, and the engineering application value is high; the method is suitable for detection and identification of underwater targets.

Description

Technical Field

[0001] The invention relates to the field of underwater acoustic signal processing, and in particular to a method for suppressing underwater target noise. Background Art

[0002] In order to detect and identify underwater targets by passive sonar, it is often necessary to extract the line spectrum features of underwater targets. Commonly used line spectrum feature estimation methods include classical spectrum estimation method and modern spectrum estimation method. The classical spectrum estimation method is also called non-parametric spectrum estimation method. It does not rely on the calculation model. It is a spectrum estimation method based on Fourier transform and has a low computational complexity. Among them, the Welch method has good variance performance (Hu Guangshu. Digital Signal Processing [M]. Beijing: Tsinghua University Press, 2003) and is widely used as a spectrum estimation method for measured data. Modern spectrum estimation methods, such as MA model parameter spectrum estimation method, multiple signal classification spectrum estimation method and spectrum estimation method based on compressed sensing theory, have improved the spectrum estimation accuracy of classical spectrum estimation methods to varying degrees, but the computational complexity is large. The data received by passive sonar not only contains underwater target signals, but also environmental noise, such as ocean noise, seismic noise, rain noise, etc. These noises will interfere with the underwater target signals. In the spectrum graph, the spectrum value of the noise is large, which makes it difficult to extract the line spectrum features of underwater targets.

[0003] Existing methods for suppressing ocean environmental noise include high-order statistics, adaptive filtering, wavelet threshold denoising, and singular spectrum decomposition. Bao Zhonghua et al. used fourth-order cumulant diagonal slicing to suppress Gaussian color noise to extract ship target line spectrum signals (see: Bao Zhonghua, Gong Shenguang, Wu Zhengguo, Li Jun. Harmonic signal line spectrum extraction based on fourth-order cumulant slicing spectrum [J]. Journal of Naval University of Engineering, 2010, 22(2): 31-34). He Xiying et al. used adaptive line spectrum enhancement method to suppress broadband ocean noise to extract underwater target line spectrum signals (see: He Xiying, Li Chengxin, Zhang Lei et al. Application research of adaptive line spectrum enhancement of vector hydrophone [J]. Journal of Marine Technology, 2011, 30(3): 68-71). Zhou Hefeng et al. proposed an improved wavelet semi-soft threshold function, using the eddy current search algorithm and the generalized cross-validation criterion to determine the stratified threshold to reduce the noise of underwater target signals (see: Zhou Hefeng, Zhao Yun, Gong Changfu et al. Research on underwater acoustic signal denoising method based on improved wavelet threshold function [J]. Acoustic Technology, 2019, 38(5): 143-144). Li Yaan et al. obtained the singular spectrum and corresponding eigenvectors of the signal and noise by performing singular value decomposition on the underwater acoustic signal in the time domain, and reconstructed the state space using the eigenvectors corresponding to the larger singular values ​​to achieve the purpose of noise reduction (see: Li Yaan, Wang Hongchao, Chen Jing. Research on underwater acoustic signal denoising method based on singular spectrum decomposition [J]. Systems Engineering and Electronics, 2007, 29(4): 524-527).

[0004] However, the existing methods for denoising underwater targets have the following problems: the high-order statistics method can theoretically completely suppress Gaussian noise, but since the data actually processed is of finite length, the method has limited ability to suppress strong background noise interference, and the high-order statistics method has a large amount of calculation; the adaptive filtering method uses the difference in correlation radius between line spectrum signals and broadband noise to suppress noise, and requires a lower estimation error for the noise; the wavelet threshold denoising method requires setting a threshold, and the selection of the threshold function has always been a difficulty. If the threshold is selected in combination with other methods, the amount of calculation will increase; the singular spectrum decomposition denoising method has a difficulty in selecting larger singular values, and often requires the signal to have a higher signal-to-noise ratio. Summary of the invention

[0005] In order to suppress the strong background noise interference of underwater targets and extract line spectrum features, the present invention proposes a noise suppression method in line spectrum feature extraction based on double singular value decomposition. This method considers that the size of the singular value corresponds to the energy of the signal and noise, segments the data received by the sonar sensor, calculates the spectrum matrix and the normalized spectrum matrix for the data, and then performs singular value decomposition on the spectrum matrix and the normalized spectrum matrix in frequency bands, constructs weighting coefficients using singular values, and weights the Welch method power spectrum to suppress noise. This method does not need to select larger singular values, can suppress the strong background noise interference of underwater targets, and thus extract line spectrum features.

[0006] A noise suppression method in line spectrum feature extraction based on double singular value decomposition of the present invention comprises the following steps:

[0007] S1, segment the data x(n) received by the sonar sensor, perform Fourier transform on each segment of data, and obtain the spectrum matrix χ(f);

[0008] S2, perform singular value decomposition on χ(f) band by band to construct the singular value parameter η Δ ;

[0009] S3, spectrum z of each segment of data in χ(f) i (f) is normalized to obtain the normalized spectrum matrix K(f), and then K(f) is subjected to singular value decomposition band by band to construct the singular value parameters β of the normalized spectrum matrix Δ ;

[0010] S4, for each singular value parameter η in the analysis band Δ , β Δ and the maximum singular value d'1 interpolation to obtain the singular value parameters η(f), β(f) and the maximum singular value d'1(f) at each frequency point, and use η(f), β(f) and d'1(f) to construct the singular value weighting coefficient ψ(f) at each frequency point;

[0011] S5, using the singular value weighting coefficient ψ'(f) at each frequency point to weight the power spectrum of the data received by the sonar sensor to suppress noise.

[0012] Furthermore, the specific steps of obtaining the spectrum matrix χ(f) in step S1 are as follows:

[0013] S1.1 divides the data x(n) received by the sonar sensor into segments, each segment has N sampling points, N is a positive integer, and some overlap between segments is allowed; the i-th segment of x(n) is expressed as:

[0014] x i (n) = x[n+(i-1)N]d i[n+(i-1)N],0<n≤N-1,i=1,2,...,L

[0015] In the formula, n represents the number of sampling points, d i (n) is a window function of length N, L is the number of segments;

[0016] S1.2 to x i (n) Perform Fourier transform to obtain the spectrum z of each segment of data i (f);

[0017] S1.3 Using the spectrum z of each segment of data i (f) Construct a spectrum matrix and obtain the spectrum matrix χ(f), χ(f) = [z1(f), z2(f), ..., z L (f)] T .

[0018] Furthermore, the specific steps of step S2 are as follows:

[0019] S2.1 performs singular value decomposition on χ(f) band by band, and we have:

[0020]

[0021] In the formula, χ Δ (f) is the matrix corresponding to χ(f) in the analysis frequency band, Δ represents the analysis frequency band; U is The eigenvector matrix of The eigenvector matrix of min The matrix is ​​composed of d1>d2>...>d min , d1 is the maximum singular value, d min is the minimum singular value;

[0022] S2.2 Using the maximum singular value d1 and the minimum singular value d in the singular value matrix D min Construct the singular value parameter η within the analysis band Δ :

[0023] η Δ =d1-d min .

[0024] Furthermore, the analysis frequency band Δ is 1 Hz.

[0025] Furthermore, the specific steps of step S3 are as follows:

[0026] S3.1 Spectrum z of each segment of data in χ(f) i (f) Perform normalization processing, and the spectrum normalization result of each segment of data is recorded as κ i (f):

[0027]

[0028] Get the normalized spectrum matrix K(f), K(f) = [κ1(f),κ2(f),...,κ L (f)] T ;

[0029] S3.2 Using the singular value decomposition method described in step S2, perform singular value decomposition on the normalized spectrum matrix K(f) band by band, and obtain the maximum singular value and the minimum singular value corresponding to the normalized spectrum matrix in each analysis band, which are recorded as d'1 and d' respectively. min ;

[0030] S3.3 Using the maximum singular value d'1 and the minimum singular value d' in the analysis band min Construct singular value parameter β Δ :

[0031] β Δ =d'1-d' min .

[0032] Furthermore, the specific steps of step S4 are as follows:

[0033] S4.1 Use the cubic spline interpolation method to calculate the η of each analysis bandwidth Δ , β Δ Interpolate and d'1 at each frequency point to obtain the singular value parameters η(f), β(f) and the maximum singular value d'1(f) at each frequency point;

[0034] S4.2 Use η(f), β(f) and the maximum singular value d'1(f) to construct the parameter ψ(f): ψ(f) = d'1(f)[η(f)β(f)] p , where p is the exponent;

[0035] S4.3 normalizes ψ(f) to obtain the singular value weighting coefficient ψ'(f) at each frequency point

[0036]

[0037] Furthermore, the value range of p described in S4.2 is a positive integer from 1 to 5, and the specific value selection method is determined according to the size of the base, that is, if the maximum value and the minimum value of the product of the singular value parameters η(f)β(f) at the frequency point are significantly different, then the larger value is taken, otherwise, 1 is taken.

[0038] Furthermore, the specific steps of step S5 are as follows:

[0039] S5.1 Use the Welch method to calculate the power spectrum of the spectrum matrix χ(f), denoted as X(f);

[0040] S5.2 Use ψ'(f) to weight X(f) and obtain the weighted power spectrum X'(f):

[0041] X'(f) = ψ'(f)X(f).

[0042] The present invention has the following beneficial effects:

[0043] (1) The present invention proposes a noise suppression method in line spectrum feature extraction based on double singular value decomposition, which performs singular value decomposition on the spectrum matrix and normalized spectrum matrix of the data, constructs singular value weighting coefficients, and weights the Welch method power spectrum to suppress noise. The method has the ability to resist strong noise interference and can extract line spectrum features under strong background noise interference;

[0044] (2) The present invention does not need to select a threshold value or a large singular value, is simple to operate, and has good adaptability for noise suppression in line spectrum feature extraction;

[0045] (3) The present invention is based on Fourier transform and singular value decomposition, has a small amount of calculation, and has high engineering application value;

[0046] (4) The line spectrum feature extraction results obtained by the present invention can be further used for underwater target detection and recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall flow chart of the present invention;

[0048] Figure 2 It is the spectrum diagram after Fourier transformation of the measured data;

[0049] Figure 3 is the singular value parameter value and the maximum singular value at each frequency point obtained using the spectrum matrix and the normalized spectrum matrix;

[0050] Figure 4 The result graph of the singular value weighting coefficient value at each frequency point;

[0051] Figure 5 The figure shows the comparison of the noise suppression power spectrum results before and after weighting. DETAILED DESCRIPTION

[0052] In order to facilitate the understanding of those skilled in the art, the technical solution of the present invention is clearly and completely described below in conjunction with specific implementation. Obviously, the embodiments described below are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] The following is combined with Figure 1 The method of the present invention is described, and the noise suppression method in the line spectrum feature extraction described in the present invention is used to process the measured data received by the sonar sensor, and the sampling frequency is 4000Hz. The measured data processed is a single-frequency signal with a frequency of 275Hz emitted in the ocean, and the duration is 6 seconds. In the frequency domain, the single-frequency signal is reflected as a line spectrum. In order to facilitate the observation of noise suppression and line spectrum extraction, the received data is pre-processed by 255Hz-445Hz bandpass filtering to obtain pre-processed time domain data.

[0054] S1, segment the above time domain data x(n), perform Fourier transform on the data of each segment, and obtain the spectrum matrix χ(f); the specific steps are as follows:

[0055] S1.1 divides the data x(n) into segments, each segment has 12000 sampling points, and the number of segments is 2;

[0056] S1.2 For each segment of data x i (n) Perform Fourier transform. The number of Fourier transform points is 12000, and the spectrum z of each segment of data is obtained. i (f);

[0057] S1.3 Using the spectrum z of each segment of data i (f) Construct a spectrum matrix and obtain the spectrum matrix χ(f), χ(f) = [z1(f), z2(f), ..., z L (f)] T .

[0058] The purpose of segmenting the above time domain data x(n) is to construct a frequency domain data matrix and use the singular value decomposition method to solve the singular values ​​in the frequency domain data matrix in each analysis frequency band.

[0059] The above time domain data x(n) is segmented, and the data of each segment is Fourier transformed to obtain the spectrum matrix χ(f), in preparation for the subsequent band-by-band singular value decomposition of χ(f).

[0060] S2 performs singular value decomposition on χ(f) band by band to construct the singular value parameter η Δ ; The specific steps are as follows:

[0061] S2.1 performs singular value decomposition on χ(f) band by band, and we have:

[0062]

[0063] In the formula, χ Δ (f) is the matrix corresponding to the χ(f) analysis frequency band, Δ represents the analysis frequency band, which is 1Hz; U is The eigenvector matrix of The eigenvector matrix of min The matrix is ​​composed of d1>d2>...>d min , d1 is the maximum singular value, d min is the minimum singular value;

[0064] S2.2 Using the maximum singular value d1 and the minimum singular value d in the singular value matrix D min Construct the singular value parameter η within the analysis band Δ :

[0065] η Δ =d1-d min

[0066] Perform singular value decomposition on χ(f) band by band and construct the singular value parameter η Δ , in preparation for the subsequent construction of the singular value weighting coefficients at each frequency point.

[0067] S3, normalize the spectrum of each segment of data in χ(f) to obtain the normalized spectrum matrix K(f), then perform singular value decomposition on K(f) band by band to construct the singular value parameters β of the normalized spectrum matrix Δ ; The specific steps are as follows:

[0068] S3.1 normalizes the spectrum of each segment of data in χ(f), and the spectrum normalization result of each segment of data is recorded as κ i (f):

[0069]

[0070] Get the normalized spectrum matrix K(f), K(f) = [κ1(f),κ2(f),...,κ L (f)] T .

[0071] S3.2 Using the singular value decomposition method described in step S2, perform singular value decomposition on the normalized spectrum matrix K(f) band by band, and obtain the maximum singular value and the minimum singular value corresponding to the normalized spectrum matrix in each analysis band, which are recorded as d'1 and d' respectively. min .

[0072] S3.3 uses the maximum singular value d'1 and the minimum singular value d' in each analysis band min Construct the singular value parameters β within each analysis bandwidth Δ :

[0073] β Δ =d'1-d' min

[0074] For each segment of data in χ(f), the spectrum z i (f) is normalized to obtain the normalized spectrum matrix K(f), and then K(f) is subjected to singular value decomposition band by band to construct the singular value parameters β of the normalized spectrum matrix Δ , in preparation for the subsequent construction of the singular value weighting coefficients at each frequency point.

[0075] S4 for each singular value parameter η in the analysis band Δ , β Δ And the maximum singular value d'1 interpolation, get the singular value parameters η(f), β(f) and the maximum singular value d'1(f) at each frequency point, use η(f), β(f) and d'1(f) to construct the singular value weighting coefficient ψ(f) at each frequency point; the specific steps are as follows:

[0076] S4.1 Use the cubic spline interpolation method to calculate the η of each analysis bandwidth Δ , β Δ The and d'1 are interpolated at each frequency point to obtain the singular value parameters η(f), β(f) and the maximum singular value d'1(f) at each frequency point.

[0077] S4.2 Use η(f), β(f) and the maximum singular value d'1(f) to construct the parameter ψ(f): ψ(f) = d'1(f)[η(f)β(f)] p , where p is an exponent, and the value range of p is a positive integer from 1 to 5. The specific value selection method is determined according to the size of the base. That is, if the maximum and minimum values ​​of the product of the singular value parameters η(f)β(f) at the frequency point are significantly different, then the larger value is taken, otherwise, 1 is taken.

[0078] In this embodiment, the value of p is 2.

[0079] S4.3 normalizes ψ(f) to obtain the singular value weighting coefficient ψ'(f) at each frequency point:

[0080]

[0081] For each singular value parameter η in the analysis band Δ , β ΔThe singular value parameters η(f), β(f) and the maximum singular value d'1 are interpolated to obtain the singular value parameters η(f), β(f) and the maximum singular value d'1(f) at each frequency point. The singular value weighting coefficient ψ(f) at each frequency point is constructed using η(f), β(f) and d'1(f) to prepare for the subsequent weighting of the Welch method power spectrum.

[0082] S5, using the singular value weighting coefficient ψ'(f) at each frequency point to weight the power spectrum of the data received by the sonar sensor to suppress noise, thereby extracting line spectrum features; the specific steps are as follows:

[0083] S5.1 Use the Welch method to calculate the power spectrum of the spectrum matrix χ(f), denoted as X(f);

[0084] S5.2 Use ψ'(f) to weight X(f) and obtain the weighted power spectrum X'(f):

[0085] X'(f) = ψ'(f)X(f).

[0086] The present invention is based on the following principle: within the analysis band where the line spectrum signal is located, under a certain signal-to-noise ratio, the maximum singular value represents most of the energy of the signal, the minimum singular value represents the energy of the noise, and the difference between the maximum singular value and the minimum singular value is large; within the analysis band where the noise is located, the maximum singular value and the minimum singular value both represent the energy of the noise. Since the correlation between the noises is small, the maximum singular value and the minimum singular value are close, and the difference between the maximum singular value and the minimum singular value is small. Therefore, the difference between the maximum singular value and the minimum singular value in the analysis band can be used to distinguish the frequency band where the noise is located and the frequency band where the line spectrum signal is located. Considering that the size of the singular value is affected by the energy of the noise and the signal, the spectrum of each segment of the data is normalized to reduce the difference between the spectrum values ​​of each segment of the spectrum data. The method jointly uses the singular value difference of the spectrum matrix, the singular value difference of the normalized spectrum matrix and the maximum singular value to construct the singular value weighting coefficient, and suppresses the noise by weighting the Welch method power spectrum, thereby extracting the line spectrum. The characteristic of this method is that it is less affected by noise energy and can be used to suppress strong background noise interference of the signal.

[0087] The following is combined with Figure 2-5 The present invention will be further described.

[0088] Figure 2 The spectrum diagram of the measured data after Fourier transformation is given. Figure 3 The singular value parameter value and the maximum singular value at each frequency point obtained using the spectrum matrix and the normalized spectrum matrix are given. Figure 4 The singular value weighting coefficients at each frequency point are given. Figure 5 The comparison results of power spectra before and after weighting are given.

[0089] from Figure 2 It can be seen that in addition to the spectrum peak at the signal frequency of 275Hz, there is also a spectrum peak near 267Hz, which will interfere with the detection and identification of the signal. Figure 3 It can be seen that at the signal frequency of 275Hz, the singular value parameter value and the maximum singular value obtained by using the spectrum matrix and the normalized spectrum matrix are the largest, and the value is 1; at the spectrum peak near 267Hz, the singular value parameter value and the maximum singular value are both smaller than the values ​​at the line spectrum signal frequency, and the singular value parameter value obtained by using the normalized spectrum matrix is ​​smaller than the singular value parameter value obtained by using the spectrum matrix, and the values ​​at other noise frequencies are smaller. Figure 4 It can be seen that the singular value weighting coefficient constructed by using the singular value parameter values ​​and the maximum singular value obtained by the spectrum matrix and the normalized spectrum matrix is ​​1 at the signal frequency of 275Hz, while the singular value weighting coefficient value near 267Hz is about 0.15, and the singular value weighting coefficient at other frequencies is less than 0.15. It can be seen that the singular value weighting coefficient at the signal frequency is much larger than the singular value weighting coefficient at the noise frequency. The singular value weighting coefficient can be used to weight the power spectrum to suppress noise. Figure 5 It can be seen that before weighting, the noise spectrum value near 267Hz is relatively high, which is almost equal to the spectrum value of the line spectrum at 275Hz, indicating strong background noise interference; after weighting, the spectrum value of the line spectrum remains unchanged, the strong background noise interference near 267Hz is suppressed, the noise spectrum value is reduced, and the difference with the spectrum value of the line spectrum is about 11dB. The noise at other frequencies is also suppressed, which is beneficial to the subsequent detection and recognition of the target.

Claims

1. A noise suppression method in line spectrum feature extraction based on double singular value decomposition, characterized in that: The steps include: S1, segment the data x(n) received by the sonar sensor, perform Fourier transform on each segment of data, and obtain the spectrum matrix χ(f); S2, perform singular value decomposition on χ(f) band by band to construct the singular value parameter η Δ ; S3, spectrum z of each segment of data in χ(f) i (f) is normalized to obtain the normalized spectrum matrix K(f), and then K(f) is subjected to singular value decomposition band by band to construct the singular value parameters β of the normalized spectrum matrix Δ ; S4, for each singular value parameter η in the analysis band Δ , β Δ and the maximum singular value d′1 interpolation to obtain the singular value parameters η(f), β(f) and the maximum singular value d′1(f) at each frequency point, and use η(f), β(f) and d′1(f) to construct the singular value weighting coefficient ψ(f) at each frequency point; S5, using the singular value weighting coefficient ψ'(f) at each frequency point to weight the power spectrum of the data received by the sonar sensor to suppress noise.

2. The noise suppression method in line spectrum feature extraction based on double singular value decomposition according to claim 1 is characterized in that: The specific steps of obtaining the spectrum matrix χ(f) in step S1 are as follows: S1.1 divides the data x(n) received by the sonar sensor into segments, each segment has N sampling points, N is a positive integer, and some overlap between segments is allowed; the i-th segment of x(n) is expressed as: x i (n)=x[n+(i-1)N]d i [n+(i-1)N],0<n≤N-1,i=1,2,…,L In the formula, n represents the number of sampling points, d i (n) is a window function of length N, L is the number of segments; S1.2 to x i (n) Perform Fourier transform to obtain the spectrum z of each segment of data i (f); S1.3 Using the spectrum z of each segment of data i (f) Construct a spectrum matrix and obtain the spectrum matrix χ(f), χ(f) = [z1(f), z2(f), ..., z L (f)] T .

3. The noise suppression method in line spectrum feature extraction based on double singular value decomposition according to claim 1 is characterized in that: The specific steps of step S2 are as follows: S2.1 performs singular value decomposition on χ(f) band by band, and we have: In the formula, χ Δ (f) is the matrix corresponding to χ(f) in the analysis frequency band, Δ represents the analysis frequency band; U is The eigenvector matrix of The eigenvector matrix of min The matrix is ​​composed of d1>d2>...>d min , d1 is the maximum singular value, d min is the minimum singular value; S2.2 Using the maximum singular value d1 and the minimum singular value d in the singular value matrix D min Construct the singular value parameter η within the analysis band Δ : or Δ =d1-d min 。 4. The noise suppression method in line spectrum feature extraction based on double singular value decomposition according to claim 3 is characterized in that: The analysis frequency band Δ is 1 Hz.

5. The noise suppression method in line spectrum feature extraction based on double singular value decomposition according to claim 1 is characterized in that: The specific steps of step S3 are as follows: S3.1 Spectrum z of each segment of data in χ(f) i (f) Perform normalization processing, and the spectrum normalization result of each segment of data is recorded as κ i (f): Get the normalized spectrum matrix K(f), K(f) = [κ1(f),κ2(f),...,κ L (f)] T ; S3.2 Using the singular value decomposition method described in step S2, perform singular value decomposition on the normalized spectrum matrix K(f) band by band, and obtain the maximum singular value and the minimum singular value corresponding to the normalized spectrum matrix in each analysis band, which are recorded as d1' and d' respectively. min ; S3.3 Using the maximum singular value d1' and the minimum singular value d' in the analysis band min Construct singular value parameter β Δ : b Δ =d′1-d′ min 。 6. The noise suppression method in line spectrum feature extraction based on double singular value decomposition according to claim 1 is characterized in that: The specific steps of step S4 are as follows: S4.1 Use the cubic spline interpolation method to calculate the η of each analysis bandwidth Δ , β Δ Interpolate d′1 and d′1 at each frequency point to obtain the singular value parameters η(f), β(f) and the maximum singular value d′1(f) at each frequency point; S4.2 Use η(f), β(f) and the maximum singular value d′1(f) to construct the parameter ψ(f): ψ(f) = d′1(f)[η(f)β(f)] p , where p is the exponent; S4.3 normalizes ψ(f) to obtain the singular value weighting coefficient ψ'(f) at each frequency point 7. The noise suppression method in line spectrum feature extraction based on double singular value decomposition according to claim 6 is characterized in that: The value range of p described in S4.2 is a positive integer from 1 to 5. The specific value selection method is determined according to the size of the base, that is, if the maximum value and the minimum value of the product of the singular value parameters η(f)β(f) at the frequency point are significantly different, then the larger value is taken, otherwise, 1 is taken.

8. The noise suppression method in line spectrum feature extraction based on double singular value decomposition according to claim 1 is characterized in that: The specific steps of step S5 are as follows: S5.1 Use the Welch method to calculate the power spectrum of the spectrum matrix χ(f), denoted as X(f); S5.2 Use ψ'(f) to weight X(f) and obtain the weighted power spectrum X'(f): X'(f) = ψ'(f)X(f).

Images