CN105488520A - Multi-resolution singular-spectrum entropy and SVM based leakage acoustic emission signal identification method - Google Patents

Abstract

The invention relates to a multi-resolution singular-spectrum entropy and SVM based leakage acoustic emission signal identification method, and belongs to the technical field of acoustic emission signal mode identification. The method comprises the steps of firstly, performing experimental data acquisition by adopting a digital acoustic emission system to acquire three simulated acoustic emission signals of knocking, abrasive paper and lead breaking; secondly, performing wavelet multi-scale decomposition on the acquired acoustic emission signals; thirdly, calculating singular-spectrum entropies of all layers and combining the singular-spectrum entropies into an eigenvector; and finally, dividing the eigenvector into a training set and a test set, wherein the training set is used for training to obtain SVM classifier parameters, and the trained classifier is used to perform classification testing on the test set, so that different acoustic emission signals can be identified by classification. According to the method, features of various samples can be well described by adopting a feature extraction method combining the wavelet multi-scale decomposition with the multi-resolution singular-spectrum entropy, and the identification method adopting the SVM has higher identification rate for the acoustic emission signals.

Description

Based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM more

Technical field

The present invention relates to based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM more, belong to acoustic emission signal mode identification technology.

Background technology

In today of rapid industrial development, all kinds of pressure pipeline, high-pressure boiler are seen everywhere, in use because the reasons such as burn into wearing and tearing may cause pipeline or furnace wall materials breakage to cause leaking, once leakage point can not get timely process, be easy to industrial accident occurs, thus bring serious economic loss and casualties.Acoustic emission testing technology be a kind of Elastic wave utilizing local material quick release of energy to produce as the detection technique of driving source, occupy an important position in Non-Destructive Testing.Current acoustic emission detection has achieved certain achievement in research in tool wear, Corrosion monitoring etc., so be applied to also be feasible in industrial pipeline Leak Detection.

At present, time domain parameter method, time domain waveform method, frequency domain method, wavelet analysis and empirical mode decomposition etc. are mainly contained to the disposal route of acoustic emission signal.These methods describe the feature of acoustic emission signal from different perspectives, provide diversified acoustic emission signal detection method.Although the document in acoustic emission detection field is a lot, the acoustic emission research of pipe leakage is mainly to the judgement having No leakage and leakage point position at present, few according to the document leaking acoustic emission signal research leakage signal parameter state.

Judge that leaking acoustic emission type is one of fundamental purpose of leaking acoustic emission detection, the determination of leaking acoustic emission type contributes to making assessment to damage of material situation further, to take corresponding remedial measures in advance.Therefore, study a kind of can in the sophisticated signal of many interference and low signal-to-noise ratio, fast and exactly extraction leakage information feature the method for identified leakage type is significant.

Wavelet analysis has good time-frequency locality, is particularly suitable for the analysis of non-stationary signal.Wavelet singular spectrum entropy merges the correlation theories such as wavelet transformation, the svd of signal and information entropy extraction, a kind of wavelet information entropy extracting method obtained.Signal can be mapped to independently linear space from wavelet space by wavelet singular spectrum entropy measure, directly reflects that the distribution of feature mode energy in analyzed signal time frequency space is uncertain.Analyzed signal is simpler, and energy is more concentrated, and wavelet singular spectrum entropy is less; Otherwise signal is more complicated, energy is overstepping the bounds of propriety loose, and wavelet singular spectrum entropy is larger.Therefore, wavelet singular composes complexity and the uncertainty that entropy measure can be used for weighing measured signal, provides the result of quantification and intuitive analysis for leaking acoustic emission research.Support vector machine (SVM) is a kind of new machine learning method under Corpus--based Method theory, can solve the small sample, the excessively problem such as study, high dimension, Local Minimum that occur in practical application, under tangible especially small sample, nonlinear situation, there is good generalization ability.

Based on this, herein first with the collection of laboratory equipment digital sound emission coefficient knock, sand paper and disconnected plumbous three kinds of simulated sound transmit; Then ask for wavelet singular spectrum entropy, be combined into characteristic parameter; Support vector machine is finally utilized to carry out training and testing to characteristic parameter.

Summary of the invention

The invention provides based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM more, leak acoustic emission signal for solution and detect classification problem, can effectively classify to leakage acoustic emission signal, promptly and accurately judging leakage failure type when leakage failure occurs, providing reference for taking corresponding remedial measures.

The present invention is based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM to be achieved in that more

First adopt digital sound emission coefficient to carry out experimental data collection, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit; Then respectively Multiscale Wavelet Decomposition is carried out to the acoustic emission signal gathered; Try to achieve the singular spectrum entropy of every one deck again, formed a proper vector; Finally proper vector is divided into training set and test set, training set is used for training and obtains support vector machine classifier parameter, and carry out class test with the sorter trained to test set, its result is exactly the Classification and Identification to different acoustic emission signal.

Described as follows based on the concrete steps differentiating the leakage acoustic emission signal recognition methods of singular spectrum entropy and SVM more:

A, acoustic emission signal collection: adopt digital sound emission coefficient to carry out experimental data collection, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit each m group, signal is designated as x (t);

B, carry out wavelet transformation to original signal x (t), if decomposition scale is j, then the wavelet decomposition signal of x (t) is d ₁, d ₂..., d _j, represent from high band to low-band signal successively;

C, respectively to decompose after each layer coefficients be reconstructed, if the reconstruction signal on jth layer is D _j={ d _j(k) }, suppose that will reconstruct 1 n ties up phase space, so n ties up phase space the 1st vector is D _jin d ₁(1), d ₂(2) ..., d _jn (), then move to right 1 step, and the 2nd vector is d ₁(2), d ₂(3) ..., d _j(n+1), the like, the matrix that (N-n+1) × n ties up can be constructed:

$A_j=\left[\begin{array}{cccc}{d}_j\left(1\right)& d_j\left(2\right)& \mathrm{...}& d_j\left(n\right)\\ d_j\left(2\right)& d_j\left(3\right)& \mathrm{...}& d_j\left(n+1\right)\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ d_j\left(N-n+1\right)& d_j\left(N-n+2\right)& \mathrm{...}& d_j\left(N\right)\end{array}\right]$

Wherein, N represents the reconstruction signal D on jth layer _j={ d _j(k) } span of k in k ∈ [1, N], from 1 to N;

D, every layer matrix A to one group of signal _jcarry out svd, calculate the singular spectrum entropy of signal, combined by the entropy of each layer, namely obtain proper vector, the singular spectrum entropy according to the method three types acoustic emission signal being total to 3*m group is obtained, and each group signal obtains a stack features vector;

E, respectively get every class signal two stack features vector as training set, two groups as test set, use their training classifiers, select gaussian radial basis function kernel function, gamma parameter is 1, and penalty factor is 200;

F, have selected classifier parameters after, get every class signal a stack features vector as training set, the proper vector of remaining set, as test set, carries out class test with the sorter trained to test set.

Described m value can select 20, a value to select 6.

Principle of work of the present invention is:

The first step: test the SAEU2S digital sound emission coefficient that Beijing Soundwel Technology Co., Ltd. can be adopted to produce and carry out Test Data Collecting.During experiment, keep sensor and accident point invariant position, simulate respectively at accident point place knock, sand paper and disconnected lead three kinds of operating mode sound sources, and record relevant experimental data.For verifying the validity of said method, in laboratory conditions, according to the regulation (U.S. ASTM976 file) of international Non-Destructive Testing circle, using 0.5mmHB pencil-lead to carry out disconnected lead test, can obtain simulating disconnected plumbous sound-source signal.And with sand paper and metal bar, simulate sand paper in same position and knock two kinds of different operating mode sound-source signals.Experiment is each gathers at 1200 o'clock as one group of signal, can gather respectively knock, sand paper and disconnected plumbous each 20 groups of three kinds of sound-source signals;

Second step: the basic skills of wavelet analysis uses the tower algorithm of Mallat, carries out depression of order decomposition to signal x (t).DMT modulation formula is: in formula: H and G is respectively low-pass filter and Hi-pass filter; c _j(k) and d _jk () is respectively approximation coefficient and the detail coefficients of signal decomposition yardstick.If decomposition scale is j, then the wavelet decomposition signal of x (t) is d ₁, d ₂..., d _j, c _j, represent from high band to low-band signal successively.

3rd step: be reconstructed each layer coefficients after decomposition respectively, reconstruction formula is c _j-1(k)=H ^*c _j(k)+G ^*d _j(k), H in formula ^*and G ^*be respectively the dual operator of H and G.If the reconstruction signal on jth layer is D _j={ d _j(k) }, suppose that will reconstruct 1 n ties up phase space, so n ties up phase space the 1st vector is D _jin d ₁(1), d ₂(2) ..., d _jn (), then move to right 1 step, and the 2nd vector is d ₁(2), d ₂(3) ..., d _j(n+1), the like, the matrix A that (N-n+1) × n ties up can be constructed _j.

$A_j=\left[\begin{array}{cccc}{d}_j\left(1\right)& d_j\left(2\right)& \mathrm{...}& d_j\left(n\right)\\ d_j\left(2\right)& d_j\left(3\right)& \mathrm{...}& d_j\left(n+1\right)\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ d_j\left(N-n+1\right)& d_j\left(N-n+2\right)& \mathrm{...}& d_j\left(N\right)\end{array}\right]$

4th step: to every layer matrix A _jcarry out svd (singularvaluedecomposition, SVD), calculate the singular spectrum entropy of signal.To every layer matrix A of one group of signal _jcarry out svd (singularvaluedecomposition, SVD), calculate the singular spectrum entropy of signal.The entropy of each layer is combined, namely obtains proper vector.Obtained by the singular spectrum entropy of three types acoustic emission signal totally 60 groups of signals according to the method, each group signal obtains a stack features vector.

5th step: get respectively every class signal two stack features vector as training set, two groups as test set, use their training classifiers, select gaussian radial basis function kernel function, gamma parameter is 1, and penalty factor is 200.Now SVM classifier is 100% to the recognition correct rate of sample to be tested.After have selected suitable classifier parameters, the 6 stack features vectors getting every class signal, as training set, remain 10 stack features vectors as test set.Result shows, SVM is very high to three kinds of classification results accuracy of leaking acoustic emission signal.

Fig. 7 expresses support for the architecture of vector machine, SVM is when carrying out linear classification, by selecting suitable kernel function that input vector is mapped to high-dimensional feature space, and constructing optimal hyperlane wherein, reaching classification object, x (1) wherein, x (2) ..., x (n) represents sample sequence to be identified, K (X, X ₁) represent kernel function, export Y and represent recognition result.The sample point that in Fig. 8 and Fig. 9, circle is irised out is called support vector (supportvectors), and it is positioned near optimal hyperlane, is distributed in the edge zone of respective sample areas, represented by dotted arrows optimal hyperlane.

The invention has the beneficial effects as follows:

1, by the type identification problem of leaking acoustic emission signal is converted into pattern classification problem, when penalty factor and gamma Selecting parameter suitable, SVM has stronger resolution characteristic to the acoustic emission information that test sample book comprises, can distinguish simultaneously knock, sand paper, disconnected plumbous signal, and recognition correct rate is very high, automatic classification can be realized;

2, support vector machine has good statistical learning ability to Small Sample Database;

3, by asking the wavelet singular spectrum entropy of acoustic emission signal, acoustic emission signal status information can be reflected preferably, the proper vector of acoustic emission signal can be it can be used as.

Accompanying drawing explanation

Fig. 1 is processing flow chart of the present invention;

Fig. 2 is the time domain beamformer that three types of the present invention leaks acoustic emission signal;

Fig. 3 is the frequency-domain waveform figure that three types of the present invention leaks acoustic emission signal;

Fig. 4 is the oscillogram after knocking of the present invention reconstruct;

Fig. 5 is the oscillogram after sand paper signal reconstruction of the present invention;

Fig. 6 is that the present invention is broken the oscillogram after plumbous signal reconstruction;

Fig. 7 is the system assumption diagram of support vector machine of the present invention;

SVM training sample visual image when Fig. 8 is training classifier parameter of the present invention;

Fig. 9 is SVM test sample book visual image of the present invention.

Embodiment

Embodiment 1: as shown in figs 1-9, based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM, first adopts digital sound emission coefficient to carry out experimental data collection more, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit; Then respectively Multiscale Wavelet Decomposition is carried out to the acoustic emission signal gathered; Try to achieve the singular spectrum entropy of every one deck again, formed a proper vector; Finally proper vector is divided into training set and test set, training set is used for training and obtains support vector machine classifier parameter, and carry out class test with the sorter trained to test set, its result is exactly the Classification and Identification to different acoustic emission signal.

Described as follows based on the concrete steps differentiating the leakage acoustic emission signal recognition methods of singular spectrum entropy and SVM more:

A, acoustic emission signal collection: adopt digital sound emission coefficient to carry out experimental data collection, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit each m group, signal is designated as x (t);

B, carry out wavelet transformation to original signal x (t), if decomposition scale is j, then the wavelet decomposition signal of x (t) is d ₁, d ₂..., d _j, represent from high band to low-band signal successively;

C, respectively to decompose after each layer coefficients be reconstructed, if the reconstruction signal on jth layer is D _j={ d _j(k) }, suppose that will reconstruct 1 n ties up phase space, so n ties up phase space the 1st vector is D _jin d ₁(1), d ₂(2) ..., d _jn (), then move to right 1 step, and the 2nd vector is d ₁(2), d ₂(3) ..., d _j(n+1), the like, the matrix that (N-n+1) × n ties up can be constructed:

$A_j=\left[\begin{array}{cccc}{d}_j\left(1\right)& d_j\left(2\right)& \mathrm{...}& d_j\left(n\right)\\ d_j\left(2\right)& d_j\left(3\right)& \mathrm{...}& d_j\left(n+1\right)\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ d_j\left(N-n+1\right)& d_j\left(N-n+2\right)& \mathrm{...}& d_j\left(N\right)\end{array}\right]$

Wherein, N represents the reconstruction signal D on jth layer _j={ d _j(k) } span of k in k ∈ [1, N], from 1 to N;

D, every layer matrix A to one group of signal _jcarry out svd, calculate the singular spectrum entropy of signal, combined by the entropy of each layer, namely obtain proper vector, the singular spectrum entropy according to the method three types acoustic emission signal being total to 3*m group is obtained, and each group signal obtains a stack features vector;

E, respectively get every class signal two stack features vector as training set, two groups as test set, use their training classifiers, select gaussian radial basis function kernel function, gamma parameter is 1, and penalty factor is 200;

F, have selected classifier parameters after, get every class signal a stack features vector as training set, the proper vector of remaining set, as test set, carries out class test with the sorter trained to test set.

Embodiment 2: as shown in figs 1-9, based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM, first adopts digital sound emission coefficient to carry out experimental data collection more, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit; Then respectively Multiscale Wavelet Decomposition is carried out to the acoustic emission signal gathered; Try to achieve the singular spectrum entropy of every one deck again, formed a proper vector; Finally proper vector is divided into training set and test set, training set is used for training and obtains support vector machine classifier parameter, and carry out class test with the sorter trained to test set, its result is exactly the Classification and Identification to different acoustic emission signal.

Described as follows based on the concrete steps differentiating the leakage acoustic emission signal recognition methods of singular spectrum entropy and SVM more:

A, acoustic emission signal collection: adopt digital sound emission coefficient to carry out experimental data collection, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit each m group, signal is designated as x (t);

B, carry out wavelet transformation to original signal x (t), if decomposition scale is j, then the wavelet decomposition signal of x (t) is d ₁, d ₂..., d _j, represent from high band to low-band signal successively;

C, respectively to decompose after each layer coefficients be reconstructed, if the reconstruction signal on jth layer is D _j={ d _j(k) }, suppose that will reconstruct 1 n ties up phase space, so n ties up phase space the 1st vector is D _jin d ₁(1), d ₂(2) ..., d _jn (), then move to right 1 step, and the 2nd vector is d ₁(2), d ₂(3) ..., d _j(n+1), the like, the matrix that (N-n+1) × n ties up can be constructed:

$A_j=\left[\begin{array}{cccc}{d}_j\left(1\right)& d_j\left(2\right)& \mathrm{...}& d_j\left(n\right)\\ d_j\left(2\right)& d_j\left(3\right)& \mathrm{...}& d_j\left(n+1\right)\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ d_j\left(N-n+1\right)& d_j\left(N-n+2\right)& \mathrm{...}& d_j\left(N\right)\end{array}\right]$

Wherein, N represents the reconstruction signal D on jth layer _j={ d _j(k) } span of k in k ∈ [1, N], from 1 to N;

D, every layer matrix A to one group of signal _jcarry out svd, calculate the singular spectrum entropy of signal, combined by the entropy of each layer, namely obtain proper vector, the singular spectrum entropy according to the method three types acoustic emission signal being total to 3*m group is obtained, and each group signal obtains a stack features vector;

E, respectively get every class signal two stack features vector as training set, two groups as test set, use their training classifiers, select gaussian radial basis function kernel function, gamma parameter is 1, and penalty factor is 200;

F, have selected classifier parameters after, get every class signal a stack features vector as training set, the proper vector of remaining set, as test set, carries out class test with the sorter trained to test set.

Described m value selects 20, a value to select 6.

Embodiment 3: as shown in figs 1-9, based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM more, first the SAEU2S digital sound emission coefficient adopting Beijing Soundwel Technology Co., Ltd. to produce carries out experimental data collection, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit each 20 groups; Then respectively Multiscale Wavelet Decomposition is carried out to the acoustic emission signal gathered; Try to achieve the singular spectrum entropy of every one deck again, formed a proper vector; Finally proper vector is divided into training set and test set, training set is used for training and obtains support vector machine classifier parameter, and carry out class test with the sorter trained to test set, its result is exactly the Classification and Identification to different acoustic emission signal.

Described as follows based on the concrete steps differentiating the leakage acoustic emission signal recognition methods of singular spectrum entropy and SVM more:

A, acoustic emission signal collection: the SAEU2S digital sound emission coefficient adopting Beijing Soundwel Technology Co., Ltd. to produce carries out experimental data collection, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit each 20 groups, signal is designated as x (t); Every class acoustic emission signal enumerates one, and its time domain waveform and spectrogram are respectively as shown in Figures 2 and 3;

B, carry out wavelet transformation to original signal x (t), if decomposition scale is 6, then the wavelet decomposition signal of x (t) is d ₁, d ₂..., d _j, represent from high band to low-band signal successively; Respectively each layer coefficients after decomposing is reconstructed, knocks, waveform after sand paper and disconnected plumbous three types acoustic emission signal reconstruct as shown in Figure 4, Figure 5 and Figure 6; Wherein transverse axis represents sampled point, and the longitudinal axis represents amplitude;

C, respectively to decompose after each layer coefficients be reconstructed, if the reconstruction signal on jth layer is D _j={ d _j(k) }, suppose that will reconstruct 1 n ties up phase space, so n ties up phase space the 1st vector is D _jin d ₁(1), d ₂(2) ..., d _jn (), then move to right 1 step, and the 2nd vector is d ₁(2), d ₂(3) ..., d _j(n+1), the like, the matrix that (N-n+1) × n ties up can be constructed:

$A_j=\left[\begin{array}{cccc}{d}_j\left(1\right)& d_j\left(2\right)& \mathrm{...}& d_j\left(n\right)\\ d_j\left(2\right)& d_j\left(3\right)& \mathrm{...}& d_j\left(n+1\right)\\ .& .& .& .\\ .& .& .& .\\ .& .& .& .\\ d_j\left(N-n+1\right)& d_j\left(N-n+2\right)& \mathrm{...}& d_j\left(N\right)\end{array}\right]$

Wherein, N represents the reconstruction signal D on jth layer _j={ d _j(k) } span of k in k ∈ [1, N], from 1 to N;

D, svd (singularvaluedecomposition, SVD) is carried out to the signal after each group reconstruct, calculate the singular spectrum entropy of signal.The entropy of each layer is combined, namely obtains proper vector.Obtained by the singular spectrum entropy of three types acoustic emission signal totally 60 groups of signals according to the method, each group signal obtains a stack features vector.Wherein Partial Feature vector parameter is as shown in table 1; Wherein y ₁, y ₂..., y ₆represent 6 layers of wavelet singular spectrum entropy respectively, represent from high band to low-band signal, classification 1,2,3 represents respectively knock, sand paper, break plumbous signal.

Table 1 Partial Feature vector parameter

E, respectively get every class signal two stack features vector as training set, two groups as test set, use their training classifiers, select gaussian radial basis function kernel function, gamma parameter is 1, and penalty factor is 200; Specifically as shown in Figure 8, totally 6 groups of training sets, 6 groups of test sets, in figure, 6 sample points represent training sample;

F, have selected classifier parameters after, get every class signal 6 stack features vector as training set, the proper vector of remaining set, as test set, carries out class test with the sorter trained to test set.Specifically as shown in Figure 9, totally 18 groups of training sets, 30 groups of test sets, in figure, 30 sample points represent test sample book.

Concrete, now the classification results of SVM classifier is as shown in Figure 8, and the recognition correct rate of sample to be tested is 100%.After have selected suitable classifier parameters in step e, the 6 stack features vectors getting every class signal, as training set, remain 10 stack features vectors as test set.Classification results as shown in Figure 9, concrete recognition correct rate is as shown in table 2, visible SVM is very high to three kinds of classification results accuracy of leaking acoustic emission signal, the comprehensive discrimination of SVM reaches 96.7%, only have a sand paper signal not identify correctly, the recognition correct rate of knocking and disconnected plumbous signal all reaches 100%.

Table 2 recognition correct rate

By reference to the accompanying drawings the specific embodiment of the present invention is explained in detail above, but the present invention is not limited to above-mentioned embodiment, in the ken that those of ordinary skill in the art possess, various change can also be made under the prerequisite not departing from present inventive concept.

Claims (3)

1. based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM, it is characterized in that: first adopt digital sound emission coefficient to carry out experimental data collection more, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit; Then respectively Multiscale Wavelet Decomposition is carried out to the acoustic emission signal gathered; Try to achieve the singular spectrum entropy of every one deck again, formed a proper vector; Finally proper vector is divided into training set and test set, training set is used for training and obtains support vector machine classifier parameter, and carry out class test with the sorter trained to test set, its result is exactly the Classification and Identification to different acoustic emission signal.

2. according to claim 1 based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM more, it is characterized in that: described as follows based on the concrete steps differentiating the leakage acoustic emission signal recognition methods of singular spectrum entropy and SVM more:

A, acoustic emission signal collection: adopt digital sound emission coefficient to carry out experimental data collection, gather respectively knock, sand paper and disconnected plumbous 3 kinds of simulated sound transmit each m group, signal is designated as x (t);

B, carry out wavelet transformation to original signal x (t), if decomposition scale is j, then the wavelet decomposition signal of x (t) is d ₁, d ₂..., d _j, represent from high band to low-band signal successively;

C, respectively to decompose after each layer coefficients be reconstructed, if the reconstruction signal on jth layer is D _j={ d _j(k) }, suppose that will reconstruct 1 n ties up phase space, so n ties up phase space the 1st vector is D _jin d ₁(1), d ₂(2) ..., d _jn (), then move to right 1 step, and the 2nd vector is d ₁(2), d ₂(3) ..., d _j(n+1), the like, the matrix that (N-n+1) × n ties up can be constructed:

$A_j=\left[\begin{array}{cccc}{d}_j\left(1\right)& d_j\left(2\right)& ...& d_j\left(n\right)\\ d_j\left(2\right)& d_j\left(3\right)& ...& d_j(n+1)\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ \·& \·& \·& \·\\ d_j(N-n+1)& d_j(N-n+2)& ...& d_j\left(N\right)\end{array}\right]$

Wherein, N represents the reconstruction signal D on jth layer _j={ d _j(k) } span of k in k ∈ [1, N], from 1 to N;

D, every layer matrix A to one group of signal _jcarry out svd, calculate the singular spectrum entropy of signal, combined by the entropy of each layer, namely obtain proper vector, the singular spectrum entropy according to the method three types acoustic emission signal being total to 3*m group is obtained, and each group signal obtains a stack features vector;

E, respectively get every class signal two stack features vector as training set, two groups as test set, use their training classifiers, select gaussian radial basis function kernel function, gamma parameter is 1, and penalty factor is 200;

F, have selected classifier parameters after, get every class signal a stack features vector as training set, the proper vector of remaining set, as test set, carries out class test with the sorter trained to test set.

3. according to claim 2 based on the leakage acoustic emission signal recognition methods of differentiating singular spectrum entropy and SVM more, it is characterized in that: described m value selects 20, a value to select 6.