CN104732970A - Ship radiation noise recognition method based on comprehensive features - Google Patents

Abstract

The invention relates to a ship radiation noise recognition method based on comprehensive features. The method comprises the first step of establishing a known ship radiation noise signal training set, the second step of extracting features of ship radiation noise in the established training set, wherein the features include auditory features and statistical features, the auditory features include the spectrum flux, the maximum spectrum peak value and the center of time gravity, the statistical features include the power spectrum peak value, the average power, the frequency of the power spectrum peak value, the average frequency, the frequency bandwidth, the frequency kurtosis, the standard difference of the power, the power gradient, the frequency power gradient, the power kurtosis and the frequency power kurtosis, the third step of using the features of the ship radiation noise as a target recognition feature training classifier, the fourth step of reading signals to be recognized, the fifth step of the features of the signals to be recognized, and the sixth step of inputting the extracted features into the classifier so as to classify and recognize the signals to be recognized.

Description

A kind of ship-radiated noise recognition methods based on comprehensive characteristics

Technical field

The present invention relates to noise identification field, particularly a kind of ship-radiated noise recognition methods based on comprehensive characteristics.

Background technology

The fast development of the underwater sound and electronic information technology, make to utilize ship-radiated noise to carry out target classification identification and become an important research topic, it is an important component part of underwater information system, its research is received always to the very big concern of many scholars, engineering technical personnel and military service.

Ship-radiated noise is very complicated, with marine environment and the motion state of boats and ships own closely related.They are along with different marine sites, different time and constantly changing.These all produce considerable influence to the type of sonic propagation and ship-radiated noise, bring larger difficulty to the Classification and Identification of ship-radiated noise.But different naval vessel is due to the difference of the immanent structures such as Ship Structure, ship type, screw propeller size, the number of blade, propulsion system, and the noise of institute's radiation is also different.Therefore, can by identifying the type on naval vessel to the classification of ship-radiated noise.

High-order statistic in modern signal processing technology, wavelet transformation, fractal geometry, artificial neural network, information fusion and data mining scheduling theory and method have been widely used in ship-radiated noise identification.But because the limitation of various method, have impact on the correctness of target identification, the recognition result in actual environment is unsatisfactory.The mankind can reach very high level for the identification of ship-radiated noise, but it is qualitative and experimental for describing the aural signature of the mankind at present, lacks the description of image and determines quantitative analysis.

The auditory system of the mankind has very excellent natural scale and robustness to the decomposition of voice signal and process, has good selectivity to sound source characteristic, has good adaptability again to neighbourhood noise.Therefore, can from human auditory system feature, research is applicable to the new feature extractive technique of underwater sound signal, finds the effective feature volume in the subjective sense of hearing amount of people's ear, reaches the object improving Underwater Targets Recognition rate.Aural signature quantitatively or pictute to probing into the mankind, the mechanism of ship noise identification and Underwater Targets Recognition are all had great importance.

The method according to human auditory system mechanism identification ship-radiated noise is still lacked in prior art.

Summary of the invention

The object of the invention is to overcome in prior art the defect of the method lacked according to human auditory system mechanism identification ship-radiated noise, thus a kind of ship-radiated noise recognition methods based on comprehensive characteristics is provided.

To achieve these goals, the invention provides a kind of ship-radiated noise recognition methods based on comprehensive characteristics, comprising:

Step 1), build known ships radiated noise signal training set;

Step 2), in the training set constructed by step 1), extract the feature of ship-radiated noise; Wherein, described feature comprises aural signature and statistical nature, described aural signature comprises spectrum flux, the highest spectrum peak and Center of Time Gravity, and described statistical nature comprises spectrum peak, average power, the frequency of spectrum peak, average frequency, frequency bandwidth, frequency kurtosis, the standard deviation of power, power pitch, the gradient of power, power kurtosis, the kurtosis of power in frequency in frequency;

Step 3), utilizing step 2) feature of ship-radiated noise that obtains is as the features training sorter of target identification;

Step 4), read signal to be identified;

Step 5), extract the feature of signal to be identified, described feature comprises aural signature and statistical nature, described aural signature comprises spectrum flux, the highest spectrum peak and Center of Time Gravity, and described statistical nature comprises spectrum peak, average power, the frequency of spectrum peak, average frequency, frequency bandwidth, frequency kurtosis, the standard deviation of power, power pitch, the gradient of power, power kurtosis, the kurtosis of power in frequency in frequency;

Step 6), the feature extracted in step 5) is input to the sorter of step 3) training, thus Classification and Identification is done to signal to be identified.

In technique scheme, in described step 3), be also included in using the feature of ship-radiated noise as target identification features training sorter before, also comprise and carry out regular to the value of described feature, they are converted into the value between 0-1.

In technique scheme, the step 2 described) or step 4) in, extract described spectrum flux and comprise:

Step 2-1-1-1), first normalized is done to time domain waveform;

Step 2-1-1-2), determine counting of every frame sampling sequence, be discrete FFT after hamming window is added to every frame sequence, delivery square to obtain discrete power spectrum;

Step 2-1-1-3), calculate the spectral vectors of adjacent two frames according to the discrete power of adjacent two frames spectrum, and then calculate the Pearson correlation coefficients between adjacent two frame spectral vectors; Its computing formula is as follows:

$r=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(X_i-\stackrel{\‾}{X})(Y_i-\stackrel{\‾}{Y})}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(X_i-\stackrel{\‾}{X})}^2}\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(Y_i-\stackrel{\‾}{Y})}^2}}$

Wherein, X, Y are the spectral vectors that adjacent two frame n tie up, X _i, Y _ibe respectively i-th value in two vectors, be respectively two vectorial averages;

Step 2-1-1-4), according to the Pearson correlation coefficients between each adjacent two frame spectral vectors, the spectrum flux of calculating noise signal; Its computing formula is as follows:

$\mathrm{SF}=\frac{1}{M}\underset{k=1}{\overset{M}{\mathrm{\Σ}}}|r_k,r_{k-1}|$

Wherein, M be noise signal be divided into the quantity of time frame; r _k, r _k-1be adjacent two time frames spectral vectors between Pearson correlation coefficients.

In technique scheme, the step 2 described) or step 4) in, extract the highest spectrum peak and comprise:

Step 2-1-2-1), first time domain waveform is normalized;

Step 2-1-2-2), hamming window is added to time domain sequences after be discrete FFT, the discrete power spectrum square obtaining signal of delivery, and be decibel by the unit conversion of power;

Step 2-1-2-3), finally find out frequency value F corresponding to noise power spectrum peak-peak _m.

In technique scheme, the step 2 described) or step 4) in, extract described Center of Time Gravity and comprise:

Step 2-1-3-1), first time domain waveform is normalized;

Step 2-1-3-2), obtain the signal energy in each moment;

Step 2-1-3-3), calculate the time domain center of gravity of noise signal, its computing formula is as follows:

$\mathrm{TC}=\frac{\underset{t}{\mathrm{\Σ}}\mathrm{tE}\left(t\right)}{\underset{t}{\mathrm{\Σ}}E\left(t\right)};$

Wherein, E (t) is the energy value on time-domain diagram corresponding to t.

In technique scheme, described sorter adopts SVM support vector machine.

The invention has the advantages that:

The present invention has fully utilized the subjective characteristics of the simulating human sense of hearing and the objective characteristics of Corpus--based Method, the subjective feeling of the mankind for sound is stated quantitatively by three features, and the objective information of sound is reflected by statistical nature, fully utilize this two category feature, and Feature Combination is optimized, effectively can improve recognition performance, make identifying closer to mankind's identifying.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the inventive method;

Fig. 2 (a) composes the schematic diagram of the dissimilar ship-radiated noise of flux identification for utilizing;

The schematic diagram that Fig. 2 (b) is the ship-radiated noise that utilizes Center of Time Gravity identification dissimilar;

Fig. 3 (a) shows the distribution situation on eight class naval vessels under st2-st5 X-Y scheme;

Fig. 3 (b) shows the distribution situation on eight class naval vessels under st3-st5 X-Y scheme.

Embodiment

Now the invention will be further described by reference to the accompanying drawings.

Ship-radiated noise recognition methods of the present invention realizes based on passive sonar.After the signal of the radiated noise that in the water such as the passive reception naval vessel of passive sonar, target produces and underwater sound equipment transmitting, for these noises and signal, with reference to figure 1, method of the present invention adopts the identification of following steps realization to ship-radiated noise.

Step 1), build known ships radiated noise signal training set.

First the pre-service such as normalization are done to the ships radiated noise signal of known class, then the ship-radiated noise of the known class after normalization is configured to training set.

Step 2), in the training set constructed by step 1), extract the feature of ship-radiated noise.

In the present invention, the feature of the ship-radiated noise extracted comprises two large classes, is respectively aural signature and statistical nature.Be described with regard to the leaching process of this two large category feature respectively below.

Step 2-1), extract aural signature.

Feature extraction is exactly extract the parameter that some can characterize target physical properties, thus obtains the essential characteristic that can characterize target signature.The aural signature that will extract in this step comprises three classes: spectrum flux, the highest spectrum peak and Center of Time Gravity.

Step 2-1-1), extract spectrum flux

The spectrum flux of voice signal describes people's ear in time to the impression degree of sound, i.e. frequency transient characteristic on a timeline.It simulates the non-linear resolution characteristic of human auditory system, reflects the characteristic information that voice signal is a large amount of and important, have impact on subjective tone color consumingly, closer to the auditory perception of people.Pearson correlation coefficients reflects the degree of Two Variables linear correlation.According to the computing method of spectrum flux, the extraction of spectrum flux characteristics is carried out frame by frame, and concrete extraction step is as follows:

Step 2-1-1-1), first pre-service is done to time domain waveform (i.e. noise signal), namely waveform is normalized;

Step 2-1-1-2), determine count (as when sampling rate is 8000Hz, every frame 600 point, 300 overlaps) of every frame sampling sequence, be 2048 discrete FFT after hamming window is added to every frame sequence, delivery square obtain discrete power spectrum;

Step 2-1-1-3), calculate the spectral vectors of adjacent two frames according to the discrete power of adjacent two frames spectrum, and then calculate the Pearson correlation coefficients between adjacent two frame spectral vectors; Its computing formula is as follows:

$r=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(X_i-\stackrel{\‾}{X})(Y_i-\stackrel{\‾}{Y})}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(X_i-\stackrel{\‾}{X})}^2}\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(Y_i-\stackrel{\‾}{Y})}^2}}$

Wherein, X, Y are the spectral vectors that adjacent two frame n tie up, X _i, Y _ibe respectively i-th value in two vectors, be respectively two vectorial averages.

Step 2-1-1-4), according to the Pearson correlation coefficients between each adjacent two frame spectral vectors, the spectrum flux of calculating noise signal; Its computing formula is as follows:

$\mathrm{SF}=\frac{1}{M}\underset{k=1}{\overset{M}{\mathrm{\Σ}}}|r_k,r_{k-1}|$

Wherein, M be noise signal be divided into the quantity of time frame; r _k, r _k-1be adjacent two time frames spectral vectors between Pearson correlation coefficients.

Step 2-1-2), extract the highest spectrum peak.

Power spectrum signal reflects signal energy stochastic distribution situation, and noise power composes the maximum frequency values of frequency values representation signal energy corresponding to peak-peak.

The extraction step of the highest spectrum peak feature is as follows:

Step 2-1-2-1), first to time domain waveform pre-service, namely waveform is normalized;

Step 2-1-2-2), hamming window is added to time domain sequences after be 2048 discrete FFT, the discrete power spectrum square obtaining signal of delivery, and be decibel by the unit conversion of power;

Step 2-1-2-3), finally find out frequency value F corresponding to noise power spectrum peak-peak _m.

Step 2-1-3), extraction time center of gravity.

The Center of Time Gravity of noise signal i.e. the center of gravity of temporal envelope, reflect the time domain specification of signal, and its concrete extraction step is as follows:

Step 2-1-3-1), first to time domain waveform pre-service, namely waveform is normalized;

Step 2-1-3-2), obtain the signal energy in each moment;

Step 2-1-3-3), calculate the time domain center of gravity of noise signal, its computing formula is as follows:

$\mathrm{TC}=\frac{\underset{t}{\mathrm{\Σ}}\mathrm{tE}\left(t\right)}{\underset{t}{\mathrm{\Σ}}E\left(t\right)}.$

Wherein, E (t) is the energy value on time-domain diagram corresponding to t.

Step 2-2), extract statistical nature.

For the ease of the extraction of ship-radiated noise statistical nature, first noise signal is divided into T time frame, and calculates the short Fourier transform of the 2F point (STFT) of every frame.Like this, signal is by represented by F spectral coefficient of T frame.In formula below, p _t,frepresent the power of signal at moment t frequency f place.

The statistical nature extracted comprises:

Spectrum peak, the i.e. maximal value of all time frame general powers:

st ₂＝M＝max(p _t)

$p_t=\underset{f}{\mathrm{\Σ}}{p}_{t,f}$

Average power, the i.e. mean value of all time frame general powers:

${\mathrm{st}}_3=\frac{{\mathrm{\Σ}}_t{p}_t}{T}$

The frequency of spectrum peak, the frequency namely corresponding to spectrum peak:

$\begin{array}{c}{\mathrm{st}}_5=\arg \underset{f}{\max }\left(p_f\right)\\ p_f=\underset{t}{\mathrm{\Σ}}{p}_{t,f}\end{array}$

Average frequency, the i.e. average frequency of signal, wherein, P represents total power signal:

$\begin{array}{c}{\mathrm{st}}_6=\stackrel{\‾}{f}=\frac{{\mathrm{\Σ}}_{f}f\·p_f}{P}\\ P=\underset{f}{\mathrm{\Σ}}{p}_f\end{array}$

RMS bandwidth, i.e. frequency bandwidth:

${\mathrm{st}}_7=B=\sqrt{\frac{{\mathrm{\Σ}}_f{f}^2*p_f}{P}-{\stackrel{\‾}{f}}^2}$

Frequency kurtosis, i.e. average frequency kurtosis:

${\mathrm{st}}_9=\sqrt{\frac{{\mathrm{\Σ}}_f{(f-\stackrel{\‾}{f})}^4*p_f}{B^{4}P}}$

Power SD, the i.e. standard deviation of power, wherein, F is the number of spectral coefficient, and T is the number of time frame:

${\mathrm{st}}_{14}=\frac{\mathrm{FT}{\mathrm{\Σ}}_{t,f}{\left(p_{t,f}\right)}^2}{P^2}-1$

Power pitch, the i.e. gradient of power:

${\mathrm{st}}_{17}=\frac{\frac{1}{\mathrm{FT}}{\mathrm{\Σ}}_{t,f}{(p_{t,f}-\stackrel{\‾}{P})}^3}{{\stackrel{\‾}{P}}^3}$

$\stackrel{\‾}{P}=P/\mathrm{FT}$

Power pitch, the i.e. gradient of power in frequency:

${\mathrm{st}}_{19}=\frac{\frac{1}{F}{\mathrm{\Σ}}_f{(\frac{p_f}{T}-\stackrel{\‾}{P})}^3}{{\stackrel{\‾}{P}}^3}$

Power kurtosis, the i.e. kurtosis of power:

${\mathrm{st}}_{40}=\frac{\frac{1}{F}{\mathrm{\Σ}}_{t,f}{(P_{t,f}-\stackrel{\‾}{P})}^4}{{\stackrel{\‾}{P}}^4}$

Power kurtosis F, the i.e. kurtosis of power in frequency:

${\mathrm{st}}_{22}=\frac{\frac{1}{F}{\mathrm{\Σ}}_f{(\frac{p_f}{T}-\stackrel{\‾}{P})}^4}{{\stackrel{\‾}{P}}^4}$

Above-mentioned 11 statistical natures are the features that can reflect ship-radiated noise, therefore obtain following 11 dimension statistical nature (st by above-mentioned statistical nature ₂, st ₃, st ₅, st ₆, st ₇, st ₉, st ₁₄, st ₁₇, st ₁₉, st ₂₀, st ₂₂).

Step 2-3), by step 2-1) three aural signatures obtaining and step 2-2) statistical nature that obtains is constructed as follows the vector of the feature for representing ship-radiated noise:

v＝{SF,F _m,TC,st ₂,st ₃,st ₅,st ₆,st ₇,st ₉,st ₁₄,st ₁₇,st ₁₉,st ₂₀,st ₂₂}。

Step 3), the feature of ship-radiated noise that previous step is obtained as the feature input sorter of target identification, with training classifier.

As the preferred implementation of one, flooding the contribution of other features to prevent a certain feature value excessive, carrying out regular to eigenwert before by the feature of ship-radiated noise input sorter, they being converted into the value between 0-1.

Sorter in this step can adopt SVM support vector machine, SVM support vector machine develops from the optimal separating hyper plane linear separability situation, its mechanism and processing procedure are equivalent to the former input space to transform to a new feature space, and in new space, solve optimum linearity classification plane.

Wherein, the principal mode of kernel function has following four kinds: linear kernel function, Polynomial kernel function, radial kernel function and Sigmoid kernel function, what the application adopted is the Polynomial kernel function on 6 rank, in other embodiments, also can adopt the kernel function of other types.

The classification function that SVM training data is formed has lower surface properties: the linear combination of SVM to be one group with support vector be nonlinearity in parameters function, and therefore the expression formula of classification function is only relevant with the quantity of support vector, and independent of the dimension in space.When processing the classification of high-dimensional input spaces, SVM is particularly effective.

Step 4), read signal to be identified;

Step 5), extract signal characteristic to be identified; Described feature comprises aforesaid 3 aural signatures and 11 statistical natures.

Step 6), the feature extracted in step 5) is input to trained sorter, thus Classification and Identification is done to signal to be identified.

The effect of the inventive method is proved below by experiment.Detect the radiated noise on eight class naval vessels in an experiment, as shown in Figure 2, wherein, ' asterisk ' represents active Sonar signal; ' square ' represents ton naval vessel, and ' circle ' represents freighter; ' plus sige ' represents another kind of large ship; ' rhombus ' represents small fishing vessel; ' right triangle ' represents medium-sized surface vessel; ' lower triangle ' represents sea noise; ' pentagram ' represents small-sized aircraft.As seen from the figure, the spectrum flux (Fig. 2 (a)) in acoustic feature and Center of Time Gravity (Fig. 2 (b)) can be used for distinguishing dissimilar ship-radiated noise, and comparatively speaking, the distinction of spectrum flux to dissimilar ship-radiated noise is better than Center of Time Gravity.

Fig. 3 (a) shows the distribution situation on eight class naval vessels under st2-st5 X-Y scheme, and Fig. 3 (b) shows the distribution situation on eight class naval vessels under st3-st5 X-Y scheme.These two figure illustrate that the separability of these statistical natures is all better.

It should be noted last that, above embodiment is only in order to illustrate technical scheme of the present invention and unrestricted.Although with reference to embodiment to invention has been detailed description, those of ordinary skill in the art is to be understood that, modify to technical scheme of the present invention or equivalent replacement, do not depart from the spirit and scope of technical solution of the present invention, it all should be encompassed in the middle of right of the present invention.

Claims (6)

1., based on a ship-radiated noise recognition methods for comprehensive characteristics, comprising:

Step 1), build known ships radiated noise signal training set;

Step 2), in the training set constructed by step 1), extract the feature of ship-radiated noise; Wherein, described feature comprises aural signature and statistical nature, described aural signature comprises spectrum flux, the highest spectrum peak and Center of Time Gravity, and described statistical nature comprises spectrum peak, average power, the frequency of spectrum peak, average frequency, frequency bandwidth, frequency kurtosis, the standard deviation of power, power pitch, the gradient of power, power kurtosis, the kurtosis of power in frequency in frequency;

Step 3), utilizing step 2) feature of ship-radiated noise that obtains is as the features training sorter of target identification;

Step 4), read signal to be identified;

Step 5), extract the feature of signal to be identified, described feature comprises aural signature and statistical nature, described aural signature comprises spectrum flux, the highest spectrum peak and Center of Time Gravity, and described statistical nature comprises spectrum peak, average power, the frequency of spectrum peak, average frequency, frequency bandwidth, frequency kurtosis, the standard deviation of power, power pitch, the gradient of power, power kurtosis, the kurtosis of power in frequency in frequency;

Step 6), the feature extracted in step 5) is input to the sorter of step 3) training, thus Classification and Identification is done to signal to be identified.

2. the ship-radiated noise recognition methods based on comprehensive characteristics according to claim 1, it is characterized in that, in described step 3), also be included in using the feature of ship-radiated noise as target identification features training sorter before, also comprise and carry out regular to the value of described feature, they are converted into the value between 0-1.

3. the ship-radiated noise recognition methods based on comprehensive characteristics according to claim 1 and 2, is characterized in that, the step 2 described) or step 4) in, extract described spectrum flux and comprise:

Step 2-1-1-1), first normalized is done to time domain waveform;

Step 2-1-1-2), determine counting of every frame sampling sequence, be discrete FFT after hamming window is added to every frame sequence, delivery square to obtain discrete power spectrum;

Step 2-1-1-3), calculate the spectral vectors of adjacent two frames according to the discrete power of adjacent two frames spectrum, and then calculate the Pearson correlation coefficients between adjacent two frame spectral vectors; Its computing formula is as follows:

$r=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(X_i-\stackrel{\‾}{X})(Y_i-\stackrel{\‾}{Y})}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(X_i-\stackrel{\‾}{X})}^2}\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(Y_i-\stackrel{\‾}{Y})}^2}}$

Wherein, X, Y are the spectral vectors that adjacent two frame n tie up, X _i, Y _ibe respectively i-th value in two vectors, be respectively two vectorial averages;

Step 2-1-1-4), according to the Pearson correlation coefficients between each adjacent two frame spectral vectors, the spectrum flux of calculating noise signal; Its computing formula is as follows:

$\mathrm{SF}=\frac{1}{M}\underset{k=1}{\overset{M}{\mathrm{\Σ}}}|r_k,r_{k-1}|$

Wherein, M be noise signal be divided into the quantity of time frame; r _k, r _k-1be adjacent two time frames spectral vectors between Pearson correlation coefficients.

4. the ship-radiated noise recognition methods based on comprehensive characteristics according to claim 1 and 2, is characterized in that, the step 2 described) or step 4) in, extract the highest spectrum peak comprise:

Step 2-1-2-1), first time domain waveform is normalized;

Step 2-1-2-2), hamming window is added to time domain sequences after be discrete FFT, the discrete power spectrum square obtaining signal of delivery, and be decibel by the unit conversion of power;

Step 2-1-2-3), finally find out frequency value F corresponding to noise power spectrum peak-peak _m.

5. the ship-radiated noise recognition methods based on comprehensive characteristics according to claim 1 and 2, is characterized in that, the step 2 described) or step 4) in, extract described Center of Time Gravity and comprise:

Step 2-1-3-1), first time domain waveform is normalized;

Step 2-1-3-2), obtain the signal energy in each moment;

Step 2-1-3-3), calculate the time domain center of gravity of noise signal, its computing formula is as follows:

$\mathrm{TC}=\frac{\underset{t}{\mathrm{\Σ}}\mathrm{tE}\left(t\right)}{\underset{t}{\mathrm{\Σ}}E\left(t\right)};$

Wherein, E (t) is the energy value on time-domain diagram corresponding to t.

6. the ship-radiated noise recognition methods based on comprehensive characteristics according to claim 1 and 2, is characterized in that, described sorter adopts SVM support vector machine.