CN101938317B - noise power spectral line spectrum detection method - Google Patents

Abstract

A noise power spectral line spectrum detection method comprises the following steps: setting a noise signal sequence as s , performing power spectrum estimation on the noise signal sequence to obtain a power spectrum p1 and a logarithmic power spectrum p2, wherein n is a non-negative integer; obtaining smooth spectrum by using sliding window orthogonal polynomial correlation fitting, and performing sliding window on the power spectrum p1 and the logarithmic power spectrum p2Orthogonal polynomial fitting is carried out to obtain a power smoothing spectrum ps1 of the noise signal and a log power smoothing spectrum ps2 of the noise signal, a difference spectrum of the power spectrum and the log power spectrum is calculated and normalized, and pd1 is [ p1 -ps1]/std{[p1(n)-ps1(n)]}pd2(n)＝[p2(n)-ps2(n)]/std{[p2(n)-ps2(n)]Pd1(n), pd2(n) are power spectrum specification difference spectrum and logarithm power spectrum specification difference spectrum respectively, an amplitude threshold G1 and a logarithm threshold G2 are set, and power spectrum and logarithm power spectrum line spectrum pl1(n), pl2(n) are extracted:

and (3) obtaining a target line spectrum pl (n) by integrating the power spectrum and the logarithmic power spectrum, wherein the spectrum values are given in a logarithmic form:

Description

Line-spectrum detection method for noise power spectra

First, technical field：

The invention belongs to signal processing technology field, it is related to a kind of method of Noise line spectra detection.

2nd, background technology：

Line-spectrum detection in noise has important meaning, and traditional extraction of line spectrum process still suffers from the shortcomings of complicated, computationally intensive computing, easy error extraction line spectrum or leakage are extracted.Its reason is that the method for trend extraction is complicated single with decision threshold.

The present invention is directed to the two shortcomings, proposes a kind of method that power spectrum Noise line spectra is extracted.Modal data sliding window is carried out using the orthogonal polynomial sequence precalculated and extracts trend term, is simplified the amount of calculation of trend extraction and is easily achieved.Target line spectrum is extracted using amplitude threshold and logarithm thresholding, the error extraction and leakage for effectively reducing line spectrum extract probability.

3rd, the content of the invention：

It is an object of the invention to provide a kind of line-spectrum detection method for noise power spectra, it can be composed in balanced ambient noise and the comprehensive of log power spectrum adjudicate the line spectrum component extracted in noise by amplitude power.

The object of the present invention is achieved like this：

A kind of line-spectrum detection method for noise power spectra, it is characterised in that for the power spectrum of noise signal, trend term is eliminated using fast background equalization methods, is integrated using difference spectrum and ratio spectrum thresholding and is extracted power spectrum line spectrum.Including following process：

A. noise signal sequence is set as s (n), and to its power Power estimation, it is nonnegative integer to obtain power spectrum p1 (n) and log power spectrum p2 (n), n,

B. smooth spectrum is obtained using sliding window orthogonal polynomial correlated fitting

If power spectrum data p (n) length N points, the window of an a length of M points, M ＜ N, step-length m points are taken, m ＜ M, if the initial starting point w=0 of window is overlapped with the starting point of power spectrum, N, M and m are positive integer, w nonnegative integers, w is the start position of window

A) 5 groups of canonical orthogonal sequences, computational methods are precalculated：

$x_i\left(n\right)={(\frac{2n}{M-1}-1)}^i,$ Wherein i=0,1,2,3,4, n=0,1 ..., M-1,

To x_i(n) Schmidt's Schimidt orthogonalizations are carried out, canonical orthogonal vector y is obtained_i(n), i=0,1,2,3,4,

B) w, w+1 ..., w+M-1 point are taken in power spectrum data p (n) as pending window power modal data s_M(n), i.e.,：

s_M(n)=p (n+w), n=0,1 ..., M-1

C) calculation window power spectrum data s_M(n) with canonical orthogonal vector y_i(n) correlation coefficient r_i,

D) calculation window fitting spectrumTake window fitting spectrum s '_M(n) m point datas are inserted in smooth modal data ss (n) before, i.e.,

Ss (n+w)=s '_M(n), n=0,1 ..., m-1

E) as w+m+M ＜ N-1, make w=w+m, return b), otherwise, into f),

F) as w+m+M ＞=N-1, remember w0=w, then make w=N-1-M, perform successively b), c) after, calculate

$s_M^{\′}\left(n\right)=\underset{i=0}{\overset{5}{\mathrm{\Σ}}}{r}_i{y}_i\left(n\right),$ Take ss (w0+m+n)=s '_M(w0+m-w+n),

Wherein n=0,1 ..., M-1+w-w0-m,

C. according to B methods, sliding window way of fitting is carried out to power spectrum p1 (n) and log power spectrum p2 (n), obtain the power smooth spectrum ps1 (n) of noise signal and the log power of noise signal smoothly composes ps2 (n)

D. calculate power spectrum and the difference of log power spectrum composed and standardized,

Pd1 (n)=[p1 (n)-ps1 (n)]/std { [p1 (n)-ps1 (n)] }

Pd2 (n)=[p2 (n)-ps2 (n)]/std { [p2 (n)-ps2 (n)] }

Pd1 (n), pd2 (n) are respectively that power spectrum specification difference spectrum and log power spectrum specification difference are composed,

E. amplitude threshold G1 and logarithm thresholding G2 is set, power spectrum and log power spectrum line spectrum pl1 (n), pl2 (n) is extracted：

F. comprehensive power spectrum and log power spectrum obtain target line spectrum pl (n), and spectrum is provided with logarithmic form：

Compared with prior art, the invention has the advantages that：

1) integrated noise power spectrum and noise log power spectrum extract line spectrum, and it is relatively reliable that more single power spectrum extracts line spectrum；

2) calculate the polynomial method of canonical orthogonal simple and easy to apply, smooth spectrum is obtained using sliding window orthogonal polynomial correlated fitting, preferable background trend can be obtained；

3) this method has preferable realizability.

4th, illustrate

Fig. 1 is the FB(flow block) of the present invention, wherein, 1. noise signals；2. calculate power spectrum；3. sliding window orthogonal polynomial correlated fitting；4. go trend term, standardization；5. extract amplitude line spectrum；6. calculate log power spectrum；7. sliding window orthogonal polynomial correlated fitting；8. go trend term, standardization；9. extract log-magnitude line spectrum；10. extract comprehensive logarithm line spectrum.

During Fig. 2 is sliding window, N, M, m correlation schematic diagram.

Fig. 3 is 16384 spot noise sequences.

Fig. 4 is noise power spectral sequence (above) and noise log power spectral sequence (figure below).

Fig. 5 is canonical orthogonal sequence.

Fig. 6 is noise power spectrum smoothing spectrum (above) and noise log power spectrum smoothing spectrum (figure below).

Fig. 7 is power spectrum specification difference spectrum (above) and log power spectrum specification difference spectrum (figure below).

Fig. 8 is noise power spectrum line spectrum.

5th, embodiment

For the power spectrum of noise signal, trend term is eliminated using fast background equalization methods, integrates using difference spectrum and ratio spectrum thresholding and extracts power spectrum line spectrum.Including following process：

Embodiment 1

A kind of line-spectrum detection method for noise power spectra,

A. noise signal sequence is set as s (n), and to its power Power estimation, it is nonnegative integer to obtain power spectrum p1 (n) and log power spectrum p2 (n), n,

B. smooth spectrum is obtained using sliding window orthogonal polynomial correlated fitting

If power spectrum data p (n) length N points, take the window of an a length of M points, M ＜ N, step-length m points, m ＜ M, if the initial starting point w=0 of window is overlapped with the starting point of power spectrum, N, M and m are positive integer, and w nonnegative integers, w is the start position of window, N, M, m relation are as shown in Figure 2

A) 5 groups of canonical orthogonal sequences, computational methods are precalculated：

$x_i\left(n\right)={(\frac{2n}{M-1}-1)}^i,$ Wherein i=0,1,2,3,4, n=0,1 ..., M-1,

To x_i(n) Schmidt's Schimidt orthogonalizations are carried out, canonical orthogonal vector y is obtained_i(n), i=0,1,2,3,4,

B) w, w+1 ..., w+M-1 point are taken in power spectrum data p (n) as pending window power modal data s_M(n), i.e.,：

s_M(n)=p (n+w), n=0,1 ..., M-1

C) calculation window power spectrum data s_M(n) with canonical orthogonal vector y_i(n) correlation coefficient r_i,

D) calculation window fitting spectrum

Take window fitting spectrum s '_M(n) m point datas are inserted in smooth modal data ss (n) before, i.e.,

Ss (n+w)=s '_M(n), n=0,1 ..., m-1

E) as w+m+M ＜ N-1, make w=w+m, return b), otherwise, into f),

F) as w+m+M ＞=N-1, remember w0=w, then make w=N-1-M, perform successively b), c) after, calculate

$s_M^{\′}\left(n\right)=\underset{i=0}{\overset{5}{\mathrm{\Σ}}}{r}_i{y}_i\left(n\right),$ Take ss (w0+m+n)=s '_M(w0+m-w+n),

Wherein n=0,1 ..., M-1+w-w0-m,

C. according to B methods, sliding window way of fitting is carried out to power spectrum p1 (n) and log power spectrum p2 (n), obtain the power smooth spectrum ps1 (n) of noise signal and the log power of noise signal smoothly composes ps2 (n)

D. calculate power spectrum and the difference of log power spectrum composed and standardized,

Pd1 (n)=[p1 (n)-ps1 (n)]/std { [p1 (n)-ps1 (n)] }

Pd2 (n)=[p2 (n)-ps2 (n)]/std { [p2 (n)-ps2 (n)] }

Pd1 (n), pd2 (n) are respectively that power spectrum specification difference spectrum and log power spectrum specification difference are composed,

E. amplitude threshold G1 and logarithm thresholding G2 is set, power spectrum and log power spectrum line spectrum pl1 (n), pl2 (n) is extracted：

F. comprehensive power spectrum and log power spectrum obtain target line spectrum pl (n), and spectrum is provided with logarithmic form：

Embodiment 2

Target noise signal sequence is gathered first for s (n), wherein n=0,1 ..., 16383, as shown in Figure 3.To noise signal power Power estimation, power spectrum p1 (n) and log power spectrum p2 (n) is obtained, as shown in figure 4, wherein n=0,1 .., 8191.Take the window of an a length of M=41 points, step-length m=5 points, if the initial starting point w=0 of window.

Smooth spectrum is obtained using sliding window orthogonal polynomial correlated fitting according to B methods, 5 groups of canonical orthogonal sequences ies are calculated_i(n), i=0,1,2,3,4, as shown in Figure 5.Sliding window way of fitting is carried out to power spectrum p1 (n) and log power spectrum p2 (n), obtain the power smooth spectrum ps1 (n) of noise signal and the log power of noise signal smoothly composes ps2 (n), as shown in Figure 6.

The difference for calculating power spectrum and log power spectrum is composed and standardized, and obtains specification difference spectrum pd1 (n), pd2 (n), as shown in Figure 7.

Amplitude threshold G1=3.0 and logarithm thresholding G2=3.0 is set, power spectrum and log power spectrum specification difference spectrum pl1 (n) is extracted, pl2 (n) obtains the line spectrum pl (n) of noise as shown in Figure 8 after integrating.In the figure 7 it can be seen that, in power spectrum specification difference spectrum, there is threshold value of many places range value more than 3.0 at 200 points at 3000 points to 4000 points between, in the specification difference spectrum of log power spectrum, 100 points, 200 points, 3000 points exceeded 3.0 threshold value, by the merging of two results, the value for having obtained 200 points and 3000 positions is line spectrum value, than the probability that the processing of single power spectrum or log power spectrum processing reduce error extraction line spectrum.

Claims (1)

1. a noise power spectrum line spectrum detection method, is characterized in that, A. Set the noise signal sequence as s(n), estimate its power spectrum, obtain power spectrum p1(n) and logarithmic power spectrum p2(n), n is a non-negative integer, B. Obtain a smooth spectrum by using a sliding window orthogonal polynomial correlation fitting Assuming that the power spectrum data p(n) has a length of N points, take a window with a length of M points, M<N, the step size is m points, m<M, and the initial starting point w=0 of the window coincides with the starting point of the power spectrum, N , M and m are positive integers, w is a non-negative integer, w is the starting position of the window, a) Pre-calculate 5 sets of orthonormal sequences, calculation method:

where i=0, 1, 2, 3, 4, n=0, 1, ..., M-1, Perform Schmidt-Schimidt orthogonalization on x _i (n) to obtain normal orthogonal vector y _i (n), i=0, 1, 2, 3, 4, b) Take w, w+1, ..., w+M-1 points in the power spectrum data p(n) as the window power spectrum data s _M (n) to be processed, namely: s _M (n) = p (n + w), n = 0, 1, ..., M-1 c) Calculating the correlation coefficient r _i of the window power spectrum data s _M (n) and the normative orthogonal vector y _i (n), d) Calculation window fitting spectrum

Take the data of m points before the window fitting spectrum s′ _M (n) and fill it in the smoothed spectrum data ss(n), that is ss(n+w)=s' _M (n), n=0, 1, ..., m-1 e) When w+m+M<N-1, make w=w+m, return to b), otherwise, enter f), f) When w+m+M>=N-1, record w0=w, and then set w=N-1-M, after performing b) and c) in sequence, calculate

Take ss(w0+m+n)=s′ _M (w0+m-w+n), where n=0, 1, ..., M-1+w-w0-m, C. According to method B, perform sliding window orthogonal polynomial fitting on the power spectrum p1(n) and logarithmic power spectrum p2(n), and obtain the power smoothing spectrum ps1(n) of the noise signal and the logarithmic power smoothing of the noise signal spectrum ps2(n), D. Calculate the difference spectrum of the power spectrum and the logarithmic power spectrum and normalize it, pd1(n)=[p1(n)-ps1(n)]/std{[p1(n)-ps1(n)]} pd2(n)=[p2(n)-ps2(n)]/std{[p2(n)-ps2(n)]} pd1(n), pd2(n) are the normalized difference spectrum of the power spectrum and the normalized difference spectrum of the logarithmic power spectrum, respectively, E. Set amplitude threshold G1 and logarithmic threshold G2, extract power spectrum and logarithmic power spectrum line spectrum pl1(n), pl2(n):

F. Combine the power spectrum and the logarithmic power spectrum to obtain the target line spectrum pl(n), and the spectral value is given in logarithmic form:

Images