CN103209036B - Based on the transient signal detection method of Hilbert-Huang Double-noise-reduction - Google Patents

Abstract

The present invention is to provide a kind of transient signal detection method based on Hilbert-Huang Double-noise-reduction.Comprise the following steps: initialization basic parameter: width threshold, empirical mode decomposition stop exponent number, wavelet detection thresholding multiple, energy density thresholding multiple, frequency discrimination rate coefficient, first order recursive coefficient; Determine effective intrinsic mode function by EMD wavelet detection, reject the IMF component of other Noises, realize the first heavy noise reduction; Utilize effective IMF component to ask Hilbert to compose square, then do Local Integral along frequency axis, obtain local instantaneous energy densimetric fraction, realize the second heavy noise reduction; Calculate signal local instantaneous energy densimetric fraction envelope, it can be used as detection statistic, do binary decision with presence or absence of signal, build local instantaneous energy density detector.The invention provides a kind of really adaptive, there is strong noise reduction capability, the signal detecting method that is applicable to low signal-to-noise ratio environment.

Description

Based on the transient signal detection method of Hilbert-Huang Double-noise-reduction

Technical field

What the present invention relates to is a kind of field of underwater acoustic signal processing, particularly relates to a kind of transient signal detection method based on Hilbert-Huang Double-noise-reduction.

Background technology

Underwater Acoustic Transient Signal can enter water, underwater sailing body ignition trigger by dropped object or the generation such as speed change turns to, ice sheet breaks, whale globefish pipes, contain abundant marine information, can be used for target identification, location, navigation etc., and the cpm signal that belongs to detects more, there is very strong researching value and application prospect widely.

The transient signal duration is short, usually only has several milliseconds, belongs to typical non-stationary signal, and classical signal processing method was lost efficacy to this.Current main stream approach has conventional energy to detect, relevant, Power-Law etc. in short-term.Energy measuring realizes simple, but it requires higher to signal to noise ratio (S/N ratio); Correlation method is by carrying out simple segmentation relevant treatment to signal data in short-term, realize high accuracy and detect, but " segmentation " destroys its temporal resolution according to certain statistical estimate; Power-Law detecting device is the detection transient signal test problems under Gaussian Background being summed up as any M point signal in N point DFT data, it does not need priori, can think complete adaptive method, but it is based on conventional DFT, adaptability is lacked to non-stationary signal.What more and more receive that scholars pay close attention to is variously have good adaptive ability, intuitively Time-Frequency Analysis Method or composite detection method.

Summary of the invention

The object of this invention is to provide a kind of really adaptive, there is strong noise reduction capability, the transient signal detection method based on Hilbert-Huang Double-noise-reduction that is applicable to low signal-to-noise ratio environment.

The object of the present invention is achieved like this, mainly comprises the steps:

(1) initialization basic parameter, mainly comprises: width threshold, empirical mode decomposition stop exponent number, wavelet detection thresholding multiple, energy density thresholding multiple, frequency discrimination rate coefficient, first order recursive coefficient;

(2) based on the adaptive noise reduction of empirical mode decomposition (EMD) wavelet detection, namely determine effective intrinsic mode function (IMF) by EMD wavelet detection, reject the IMF component of other Noises, realize the first heavy noise reduction;

(3) local instantaneous energy densimetric fraction noise reduction, namely utilize effective IMF component to ask Hilbert to compose square, then do Local Integral along frequency axis, obtain local instantaneous energy densimetric fraction, realize the second heavy noise reduction;

(4) calculate signal local instantaneous energy densimetric fraction envelope, it can be used as detection statistic, do binary decision with presence or absence of signal, build local instantaneous energy density detector.

The present invention can also comprise:

The step of the described adaptive noise reduction based on empirical mode decomposition wavelet detection is: first for the IMF component that EMD obtains, adopt running mean, the power of the every rank IMF of independent estimations, power average is multiplied by wavelet detection thresholding multiple as power threshold, then IMF screening is carried out by instantaneous power peak-value detection method, if IMF power peak is greater than power threshold, thinks that these rank IMF is effective, otherwise think invalid and reject.

Described local instantaneous energy densimetric fraction noise reduction is the priori of medium and low frequency section residing for Underwater Acoustic Transient Signal, Hilbert spectrum square is done to frequency Local Integral obtains, be equivalent to by a bandpass filter on frequency domain, thus the interference that left behind in the heavy noise reduction of filtering first.

Described local instantaneous energy density detector is using local instantaneous energy densimetric fraction envelope as detection statistic, does binary decision with presence or absence of signal.

The principal feature of method of the present invention is: (1) and conventional empirical mode decomposition pass through to extract high-order IMF and improve compared with the method for signal to noise ratio (S/N ratio), it is carry out IMF screening adaptively by instantaneous power peak-value detection method that neutron deficiency of the present invention detects, and can realize the first heavy noise reduction; (2), compared with the instantaneous energy densimetric fraction obtained with conventional frequency overall situation integration, the local instantaneous energy densimetric fraction in the present invention is that Hilbert spectrum square is done frequency Local Integral and obtained, and can realize the second heavy noise reduction; (3) compared with common transient signal detecting device, the present invention adopts local instantaneous energy densimetric fraction envelope as detection statistic, more can embody the local characteristics of transient signal.

Accompanying drawing explanation

Fig. 1 is the transient signal overhaul flow chart based on Hilbert-Huang Double-noise-reduction;

Fig. 2 (a) is Underwater Acoustic Transient Signal figure;

Fig. 2 (b) is for adding make an uproar transient signal and EMD decomposition result figure;

Fig. 3 is each rank IMF wavelet detection schematic diagram;

Fig. 4 is the survey statistic comparison diagram that local instantaneous energy Density Detection and conventional energy detect;

Fig. 5 is energy density-width combine detection process flow diagram;

Fig. 6 is ROC curve performance comparison diagram.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described:

Convert the transient signal testing process of Double-noise-reduction as shown in Figure 1 based on Hilbert-Huang (Hilbert-Huang), key step is:

(1) initialization basic parameter, described basic parameter comprises: width threshold, empirical mode decomposition stop exponent number, wavelet detection thresholding multiple, energy density thresholding multiple, frequency discrimination rate coefficient, first order recursive coefficient.

Width threshold: W=1ms;

Empirical mode decomposition stops exponent number: K=5;

Wavelet detection thresholding multiple: σ=4;

Energy density thresholding multiple: T_times=9;

HHT frequency discrimination rate coefficient: N=1000;

First order recursive coefficient: signal coefficient M1=50, threshold coefficient M2=2000.

(2) based on the adaptive noise reduction of empirical mode decomposition wavelet detection, namely determine effective IMF component by EMD wavelet detection, reject the IMF component of other Noises, realize the first heavy noise reduction.

1. empirical mode decomposition is carried out to input signal, concrete:

A) interpolation adopts cubic spline interpolation;

B) decompose end condition to be set to: the amplitude of surplus is less than 10 of original signal amplitude ^-10; The limit number of surplus is less than or equal to 2; IMF exponent number is greater than the decomposition upper limit;

C) component end condition is set to: have three threshold value sd1, sd2 and tol, wherein sd1 < sd2, and three values are all between 0 and 1.Calculate:

sx(k)＝2|men(k)/(uen(k)-den(k))| (1)

Wherein men (k), uen (k), den (k) are respectively the envelope average of signal, coenvelope and lower envelope.

Screen when three conditions meet one of them below and stop: the value of any point in sx (k) is all less than sd2, and the ratio more than sd1 in sx (k) is not more than tol, and the number of extreme point and zero crossing is more or less the same in 1; Be less than 3 limit residues; Screening number of times is greater than maximum screening number of times thresholding;

D) boundary treatment adopts image method.

Fig. 2 give Underwater Acoustic Transient Signal and add make an uproar after EMD decomposition result.

2. EMD wavelet detection adopts instantaneous power peak-value detection method:

From the angle analysis of power, if the 1st rank IMF does not have signal, if signal length is M, then its power can be expressed as:

N _i(k)＝P _ni(k),k＝0,1,...,M-1 (2)

If the i-th rank IMF has signal to exist, then its power can be expressed as:

N _i(k)＝P _ni(k)+P _si(k),k＝0,1,...,M-1 (3)

P _ni(k) be in the i-th rank IMF noise at the power in k moment, P _sik () is that in the i-th rank IMF, signal at the power in k moment, then has following judgement:

Wherein λ _ibeing the power threshold of the i-th rank IMF, is the multiple of the average power of the i-th rank IMF.The method of estimation of wavelet detection power threshold is as follows:

First every rank IMF is estimated _iinstantaneous power P _i(k):

P _i(k)＝(IMF _i(k)) ²，k＝0,1,...,M-1 (5)

Instantaneous power is done to the running mean N of L point _i(k):

$N_i\left(k\right)=\frac{\underset{i=0}{\overset{L.-1}{\mathrm{\&Sigma;}}}{P}_i(k-1)}{L},k=L-1,L,...,M-1---\left(6\right)$

Then power threshold λ _ifor

${\mathrm{\&lambda;}}_i=\mathrm{\&sigma;}\&times;\frac{\underset{k=L-1}{\overset{M-1}{\mathrm{\&Sigma;}}}{N}_i\left(k\right)}{M-L}---\left(7\right)$

In order to not drop-out, wavelet detection should get lower threshold coefficient σ.Work as N _iwhen the maximal value of () is greater than power threshold k, thinks effective IMF component, otherwise think invalid and reject.For noisy signal shown in Fig. 2, the 5th rank and surplus are detected as effectively by wavelet detection, and coincide with result shown in Fig. 2, each rank IMF wavelet detection schematic diagram as shown in Figure 3.

(3) local instantaneous energy densimetric fraction noise reduction, namely utilize effective IMF component to ask Hilbert to compose square, then do Local Integral along frequency axis, obtain local instantaneous energy densimetric fraction, realize the second heavy noise reduction.

Hilbert conversion is done to the i-th effective IMF in rank, argument θ _it (), as instantaneous phase, by formula (8) to time differentiate, obtains instantaneous frequency ω _i(t).

${\mathrm{\&omega;}}_i\left(t\right)=2\mathrm{\&pi;}\frac{{\mathrm{d\&theta;}}_i\left(t\right)}{\mathrm{dt}}---\left(8\right)$

Calculate Hilbert according to formula (9) again and compose H (ω, n), if C to D rank IMF is effective, and discretize ω _in () represents the instantaneous frequency of the i-th rank IMF in the n moment, and order

${{\mathrm{IMF}}_i}^{\&prime;}\left(n\right)=\left\{\begin{array}{cc}{\mathrm{IMF}}_i\left(n\right),& {\mathrm{\&omega;}}_i\left(n\right)=\mathrm{\&omega;}\\ 0,& \mathrm{else}\end{array}\right.---\left(9\right)$

Then:

$H(\mathrm{\&omega;},n)=\underset{i=C}{\overset{D}{\mathrm{\&Sigma;}}}\left|{{\mathrm{IMF}}_i}^{\&prime;}\left(n\right)\right|---\left(10\right)$

If ie is local instantaneous energy densimetric fraction, then:

$\mathrm{ie}\left(n\right)=\underset{\mathrm{\&omega;}}{\mathrm{\&Sigma;}}{\left|H(\mathrm{\&omega;},n)\right|}^2,n=\mathrm{0,1},...,M-1---\left(11\right)$

Wherein, the selection of limit of integration P and Q and sample frequency f _srelevant with Hilbert spectral frequency quality coefficient N.If limit of integration is [f ₀, f ₁], then:

expression rounds up, represent and round downwards.The local instantaneous energy densimetric fraction noise-reduction method that the present invention proposes is equivalent on frequency domain, added a bandpass filter, can further improve signal to noise ratio (S/N ratio).

(4) calculate signal local instantaneous energy densimetric fraction envelope, it can be used as detection statistic, do binary decision with presence or absence of signal, build local instantaneous energy density detector.

Calculate envelope iee (n) and energy density thresholding ν (n) of local instantaneous energy densimetric fraction ie (n) according to formula (14), the former makes filtering parameter m=M1, A=1, and the latter makes m=M2, A=T_times.

$y\left(n\right)=\frac{m-1}{m}y(n-1)+\frac{1}{m}(\mathrm{ie}\left(n\right)\&times;A)---\left(14\right)$

Using iee (n) as detection statistic, the detection statistic that the local instantaneous energy Density Detection of noisy signal shown in Fig. 2 and conventional energy detect contrasts as shown in Figure 4, and the detection method signal to noise ratio (S/N ratio) that the present invention proposes detects apparently higher than conventional energy.The local instantaneous energy density detector built adopts energy density-width combine detection logic, and idiographic flow as shown in Figure 5.

Fig. 6 gives the performance comparison ROC curve of transient signal detection method of the present invention and conventional energy detection method.

Finally it should be noted that, above embodiment is only in order to describe technical scheme of the present invention instead of to limit this technical method, the present invention can extend in application other amendment, change, application and embodiment, and therefore think that all such amendments, change, application, embodiment are all in spirit of the present invention and teachings.

Claims (2)

1., based on a transient signal detection method for Hilbert-Huang Double-noise-reduction, it is characterized in that comprising the following steps:

(1) based on the adaptive noise reduction of empirical mode decomposition EMD wavelet detection, first for the intrinsic mode function IMF component that EMD obtains, adopt running mean, the power of the every rank IMF of independent estimations, power average is multiplied by wavelet detection thresholding multiple as power threshold, then carries out IMF screening by instantaneous power peak-value detection method, if IMF power peak is greater than power threshold, thinks that these rank IMF is effective, otherwise think invalid and reject, realizing the first heavy noise reduction;

(2) local instantaneous energy densimetric fraction noise reduction, namely utilize effective IMF component to ask Hilbert to compose square, then do Local Integral along frequency axis, obtain local instantaneous energy densimetric fraction, realize the second heavy noise reduction;

(3) signal local instantaneous energy densimetric fraction envelope will be calculated as detection statistic, do binary decision with presence or absence of signal, build local instantaneous energy density detector;

EMD wavelet detection adopts instantaneous power peak-value detection method:

From the angle analysis of power, if the 1st rank IMF does not have signal, if signal length is M, then its power can be expressed as:

N _i(k)＝P _ni(k),k＝0,1,...,M-1 (2)

If the i-th rank IMF has signal to exist, then its power can be expressed as:

N _i(k)＝P _ni(k)+P _si(k),k＝0,1,...,M-1 (3)

P _ni(k) be in the i-th rank IMF noise at the power in k moment, P _sik () is that in the i-th rank IMF, signal at the power in k moment, then has following judgement:

Wherein λ _ibe the power threshold of the i-th rank IMF, be the multiple of the average power of the i-th rank IMF, the method for estimation of wavelet detection power threshold is as follows:

First every rank IMF is estimated _iinstantaneous power P _i(k):

P _i(k)＝(IMF _i(k)) ²，k＝0,1,...,M-1 (5)

Instantaneous power is done to the running mean N of L point _i(k):

$N_i\left(k\right)=\frac{\underset{i=0}{\overset{L-1}{\&Sigma;}}{P}_i(k-i)}{L},k=L-1,L,...,M-1---\left(6\right)$

Then power threshold λ _ifor:

${\mathrm{\&lambda;}}_i=\mathrm{\&sigma;}\&times;\frac{\underset{k=L-1}{\overset{M-1}{\&Sigma;}}{N}_i\left(k\right)}{M-L}---\left(7\right)$

In order to not drop-out, wavelet detection should get lower threshold coefficient σ, works as N _iwhen the maximal value of () is greater than power threshold k, thinks effective IMF component, otherwise think invalid and reject, for noisy signal, the 5th rank and surplus are detected as effectively by wavelet detection.

2. the transient signal detection method based on Hilbert-Huang Double-noise-reduction according to claims 1, it is characterized in that: described local instantaneous energy densimetric fraction noise reduction is the priori of medium and low frequency section residing for Underwater Acoustic Transient Signal and knows interfere information, Hilbert spectrum square is done to frequency local selective integration obtains, be equivalent to the bandpass filter fallen into time-frequency two-dimensional zero by, thus the interference that left behind in the heavy noise reduction of filtering first.