CN101718581A - Alarming method of nuclear power station loose part based on support vector machine - Google Patents

Abstract

The invention discloses an alarming method of a nuclear power station loose part based on a support vector machine. The alarming method comprises the following steps of: installing a plurality of acceleration sensors on a circuit of a nuclear reactor to acquire environmental noises and impact signals; forming a vibration signal band by knocking signals which are generated by forcibly hammering a steel plate with different forces; extracting characteristics of the knocking signals and constructing an identification module of the support vector module (SVM) by the knocking signals; carrying out an AR modeling on an impact signal, whitening the impact signal and acquiring a root mean square of the whitened signal, wherein the root mean square is used for representing the amplitude of the whitened signal; judging whether to a primary alarm is generated, and recording the whitened signal; acquiring the whitened signal triggering the primary alarm, extracting target characteristic of the whitened signal, inputting the target characteristic to the SVM, and judging whether the target characteristic is an interference signal or not by the SVM according to computation. The invention has the advantages of high data processing speed, high alarming accuracy rate, and low false alarming rate and missed alarming rate.

Description

Alarming method of nuclear power station loose part based on support vector machine

Technical field

The present invention relates to a kind of method of sending warning when loosening member of nuclear power station falls that is used to detect.

Technical background

Loosening element monitoring system (LPMS) is one of the most basic security tool in nuclear power station one loop, and the nuclear power station security of operation is had vital role.The major function of LPMS is whether the monitoring reaction cooling in heap but has in the system (RCS) parts loosening or fall, if any then sending warning.Existing loosening element warning pertinent literature has:

[1] BECHTOLD B, KUNZ U.KUES ' 95-The modern diagnosticsystem for loose parts monitoring[J] .Progress in Nuclear Energy, 1999,34 (3): 221-230. (BECHTOLD B, KUNZ U.KUES ' 95-advanced person's loosening element monitoring diagnosis system [J]. the nuclear energy progress, 1999,34 (3): 221-230.)

[2] SZAPPANOS G, POR G.Basics ideas and realization ofcompletely digitized loose part detection system HELPS[J] .Progress inNuclear Energy, 1999,34 (3): 195-201. (SZAPPANOS G, key concept and the realization [J] of POR G. total digitalization loosening element monitoring system HELPS. the nuclear energy progress, 1999,34 (3): 195-201.)

[3] POR G, KISS J, SOROSANSZKY I, et al.Development of afalse alarm free advanced loose parts monitoring system (ALPS) [J] .Progress in Nuclear Energy, 2003,43 (3): 243-251. (POR G, KISS J, SOROSANSZKY I, the exploitation [J] of the low rate of false alarm loosening element monitoring system of et al.. the nuclear energy progress, 2003,43 (3): 243-251.)

[4] Fang Lixian, Lou Yongjian. the applied research [J] of wavelet transformation in the loosening element detection system is reported to the police. atomic energy science technology, 2004,38 (5): 432-435.

[5] Mao Hanling, Huang Zhenfeng, Chen Zhongyi. based on the neural network alarm method [J] of signal time-frequency characteristics. Nuclear Power Engineering, 1998,19 (3): 265-269.

Nuclear power station operating experience for many years shows, LPMS helps to find in early days and get rid of loosening element, prevents that loosening element and the collision of inner miscellaneous part from contacting nuclear power station cause serious harm [1].Warning is the key function of LPMS system, and the availability of system is played a decisive role.All there is the high problem of rate of false alarm in existing many LPMS system, causes operating personnel to distrust system alarm, some nuclear power station even close the LPMS system.So how to distinguish the true and false of collision alarm, the rate of false alarm that reduces system is the focus and the difficult point of research.Existing system is general to carry out filtering to the signal of real-time collection earlier, then the amplitude by judging signal or in short-term root mean square (RMS) whether trigger elementary warning above preset threshold, again signal is carried out the true and false that labor is judged warning at last.Existing more about how differentiating the research of warning true and false aspect, document [2-3] proposes by judging whether signal surpasses threshold value or repeatedly judge whether to be true warning above threshold value at single passage at a plurality of passages.Document [4] proposes earlier signal to be carried out wavelet decomposition and obtains the high-order wavelet coefficient, and then judges the true and false by neural network.The degree of correlation between each channel signal of document [5] proposition calculating and the judgement true and false recently of signal high-frequency energy and low frequency energy.

But existing filtering method all can't be with the whole filterings of the noise in the signal, produce false alarm when the amplitude of noise surpasses preset value, or useful signal are submerged in the noise, fails to report the police, exist the accuracy rate of reporting to the police low, rate of false alarm, the shortcoming that rate of failing to report is high.

Summary of the invention

For the accuracy rate that overcomes prior art is low, rate of false alarm, the shortcoming that rate of failing to report is high the invention provides a kind of warning accuracy rate height, rate of false alarm, the alarming method of nuclear power station loose part based on support vector machine that rate of failing to report is low.

Alarming method of nuclear power station loose part based on support vector machine may further comprise the steps:

1, in nuclear reactor one loop a plurality of acceleration transducers are installed, obtaining the neighbourhood noise in the nuclear reactor, and when loosening element falls, produce, mix the impact signal that neighbourhood noise is arranged;

2, firmly hammer knocks the impact signal of steel plate to produce during loosening element falls in the simulation nuclear reactor with different dynamics, obtains the power hammer and knocks the knocking that steel plate produces, and constitutes the vibration signal storehouse by these knockings;

3, extract the feature of knocking, with knocking feature construction support vector machine (SVM) model of cognition;

4, will carry out the AR modeling by the impact signal that obtains in the step 1),, and obtain the root mean square of whitened signal, characterize the amplitude of this whitened signal with described root mean square the impact signal albefaction; Default amplitude threshold judges whether the root mean square of whitened signal surpasses amplitude threshold, if then send primary alarm, and write down this whitened signal; If not, then continue monitoring;

5, obtain the whitened signal that triggers primary alarm, extract the target signature of this whitened signal, this target signature is imported in the SVM, judged according to calculating whether this target signature is undesired signal by SVM, if then return step 4); If not, then send secondary and report to the police, overhaul to notify the staff.

Further, step 2) in, the knocking in the vibration signal storehouse is that the power hammer knocks the vibration signal that different parts produced in nuclear reactor one loop.

Further, in the step 3), the feature of extracting signal may further comprise the steps:

3.1) current demand signal that collects is intercepted, get the preceding 0.1s of vibration beginning to beginning back 0.5s;

3.2) adopt Short Time Fourier Transform to carry out time frequency analysis to the signal that is truncated to, window width is 1024 points;

3.2) adopt pivot analysis (PCA) technology to compressing through the signal behind the time frequency analysis, extract feature.

Technical conceive of the present invention is: adopt the AR model to carry out albefaction to signal and handle, and calculate whitened signal RMS, compare with the amplitude threshold of presetting, judge whether the one-level warning.Suppose that the signal of gathering is x (n), then its whitened signal is:

$w\left(n\right)=x\left(n\right)+\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{a}_{k}x(n-k)$

Getting window width is 5ms, the RMS of signal calculated:

$\mathrm{RMS}=\sqrt{\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_i^2}$

If RMS surpasses threshold value, then write down this whitened signal, and trigger one-level and report to the police, the notification data process computer carries out next step analysis to signal.

Adopt Short Time Fourier Transform to carry out time frequency analysis to the whitened signal that triggers the one-level warning, obtain the time-frequency figure (seeing accompanying drawing) of signal, then time-frequency figure is compressed.

$\mathrm{STFT}=\underset{-\∞}{\overset{\∞}{\∫}}w\left(t\right)q(t-\mathrm{\τ})e^{-\mathrm{j\ωt}}\mathrm{dt}$

Q is a window function in the following formula, and the time-frequency figure resolution that what-if obtains is m * n, and each row of picture are joined end to end, and obtains the vectorial V of a m * n dimension, and this vector sum projection matrix is multiplied each other:

Y＝PV

Vectorial Y after can obtaining compressing, with respect to raw data, the data after the compression have only original about 1% can express the quantity of information more than 95% in the former data.

The SVM model that the input of data after the compression has trained is discerned, to judge whether that sending secondary reports to the police.

α wherein _iBe model parameter,

Be kernel function.

If f (y)＜0, then expression is once reported to the police and is spurious alarm, otherwise sends the secondary alerting signal, and notice nuclear power station operating personnel further check.

Whether the present invention determines whether that above amplitude threshold needs carry out one-level and report to the police according to the amplitude of vibration signal earlier, and only the signal that the triggering one-level is reported to the police carries out feature identification, sends the secondary warning, has accelerated the recognition efficiency of SVM, has reduced the rate of false alarm of system.

Adopt the AR model that signal has been carried out the albefaction processing, improved the signal to noise ratio (S/N ratio) of signal greatly, reduce the rate of failing to report of system.Employing has the SVM of fine generalization ability as the Classification and Identification device, makes that the discrimination of system is effectively improved.Adopt the PCA method that signal is compressed, removed the correlativity between the data, make data volume reduce, accelerated the efficient of SVM training and identification.

It is fast that the present invention has data processing speed, warning accuracy rate height, rate of false alarm, the advantage that rate of failing to report is low.

Description of drawings

Fig. 1 is a process flow diagram of the present invention

Fig. 2 is the synoptic diagram of the signal after Short Time Fourier Transform

Fig. 3 is the synoptic diagram of the signal of process AR filtering

Fig. 4 is the synoptic diagram of amplitude variable signal of the present invention through AR filtering

Embodiment

With reference to accompanying drawing, further specify the present invention:

Embodiment one

With reference to accompanying drawing, further specify the present invention:

Alarming method of nuclear power station loose part based on support vector machine may further comprise the steps:

1, in nuclear reactor one loop a plurality of acceleration transducers are installed, obtaining the neighbourhood noise in the nuclear reactor, and when loosening element falls, produce, mix the impact signal that neighbourhood noise is arranged;

2, firmly hammer knocks the impact signal of steel plate to produce during loosening element falls in the simulation nuclear reactor with different dynamics, obtains the power hammer and knocks the knocking that steel plate produces, and constitutes the vibration signal storehouse by these knockings;

3, extract the feature of knocking, with knocking feature construction support vector machine (SVM) model of cognition;

4, will carry out the AR modeling by the impact signal that obtains in the step 1),, and obtain the root mean square of whitened signal, characterize the amplitude of this whitened signal with described root mean square the impact signal albefaction; Default amplitude threshold judges whether the root mean square of whitened signal surpasses amplitude threshold, if then send primary alarm, and write down this whitened signal; If not, then continue monitoring;

5, obtain the whitened signal that triggers primary alarm, extract the target signature of this whitened signal, this target signature is imported in the SVM, judged according to calculating whether this target signature is undesired signal by SVM, if then return step 4); If not, then send secondary and report to the police, overhaul to notify the staff.

Further, step 2) in, the knocking in the vibration signal storehouse is that the power hammer knocks the vibration signal that different parts produced in nuclear reactor one loop.

Further, in the step 3), the feature of extracting signal may further comprise the steps:

3.1) current demand signal that collects is intercepted, get the preceding 0.1s of vibration beginning to beginning back 0.5s;

3.2) adopt Short Time Fourier Transform to carry out time frequency analysis to the signal that is truncated to, window width is 1024 points;

3.2) adopt pivot analysis (PCA) technology to compressing through the signal behind the video analysis, extract feature.

Basic theory of the present invention has:

Autoregressive model (AR model, Autoregressive Model)

Ambient noise signal x (n) can be regarded as the response by a certain definite system of white noise w (n) excitation, if represent this system with the AR model, then its system function is:

$H\left(z\right)=\frac{1}{1+\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{a}_k{z}^{-k}}$

In the formula: p is system's exponent number; a _kBe the coefficient of AR model, can get:

$x\left(n\right)=w\left(n\right)-\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{a}_{k}x(n-k)$

Under the situation of the p rank of known signal x (n) and signal AR model parameter, can try to achieve white noise w (n):

$w\left(n\right)=x\left(n\right)+\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{a}_{k}x(n-k)$

This model is called the albefaction model of signal.The exponent number p of AR model can adopt based on the final predicated error criterion (FPE) of square error minimum and determine that its computing formula is:

$\mathrm{FPE}\left(p\right)={\widehat{\mathrm{\σ}}}^2\left(\frac{N+p+1}{N-p-1}\right)$

In the formula:

Standard deviation for the model evaluated error; N is the signal length of using when calculating the AR parameter.

Because

Reduce along with the increase of p,

Then increase, therefore, can find the exponent number p an of the best along with the increase of p _OptMake the FPE minimum.

Accompanying drawing 3 has shown that signal carries out the effect that albefaction is handled again through after mixing, and the signal to noise ratio (S/N ratio) of collision alarm and ground unrest is 0db, as can be seen from this figure, easy to identifyly in the whitened signal goes out collision alarm.

Pivot analysis data compression principle (PCA, Principal components analysis)

Because the data volume of signal time-frequency figure own is very big, and the very big redundancy of existence between the data, data are directly imported SVM to be discerned, can cause arithmetic speed slow excessively, and redundant data can impact final judgement, so adopt pivot analysis (principalcomponents analysis) to come data are compressed,, extract data characteristics to eliminate data redundancy.Pivot analysis is called Karhunen-Loeve transformation again, its objective is a kind of conversion of design, with data map to a low bit space, represents raw data with less data volume.

Suppose one group of sample w _i∈ R ^{N * 1}(i=1,2 ..., be m the sample of signal of handling through albefaction of set W m), the sample average vector is Then the covariance matrix C of sample can be expressed as:

$C=\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}(w_i-\mathrm{\μ}){(w_i-\mathrm{\μ})}^T\∈R^{n\×n}$

Make λ _i(i=1,2 ..., r) be r nonzero eigenvalue of Matrix C, V _iFor Matrix C corresponding to λ _iProper vector, to eigenwert according to ordering from big to small after, only need get top several classification and just can enough express whole data accurately, it can say the dimension reduction of data.Can choose front p maximum proper vector according to the shared energy proportion of eigenwert and form transformation matrix P=[p ₁, p ₂... p _r] ∈ R ^{N * r}Raw data is by multiplying each other with transformation matrix: Y=PW obtains the data after the conversion.

Support vector machine (SVM, Support vector machine)

Support vector machine is called for short SVM to be had good recurrence estimated capacity and have very good popularization ability under limited training sample situation.

Suppose sample of signal Y be into the data after the PCA compression, corresponding alarming result is G={g ₁, g ₂..., g _m, wherein the value of G is-1 and 1, if be 1 then expression is true reports to the police, otherwise is spurious alarm, supposes that funtcional relationship between the two is:

ω∈R ^nl，b∈R

Wherein:

To import data and be mapped to high-dimensional feature space, so that the nonlinear function fitting problems in the former input space is converted into the linear function fit problem in the higher dimensional space from the n dimension.

Target is exactly to seek suitable ω and b, and according to Statistical Learning Theory, ω and b can be by minimizing the structure risk function R _RegObtain:

$R_{\mathrm{reg}}=\frac{1}{2}{\left|\right|\mathrm{\ω}\left|\right|}^2+\frac{C}{H}\underset{k=1}{\overset{H}{\mathrm{\Σ}}}{L}_{\mathrm{\ϵ}}$

Wherein: L _εBe ε insensitive loss function: L _ε=max{0, | g-f (y) |-ε }

Consider the situation that allows error of fitting, introduce slack variable ξ _i, ξ _i ^*, make:

Objective function becomes:

$\underset{\mathrm{\ω},b,\mathrm{\ξ},{\mathrm{\ξ}}^{*}}{\min }J(\mathrm{\ω},\mathrm{\ξ},{\mathrm{\ξ}}^{*})=\frac{1}{2}{\left|\right|\mathrm{\ω}\left|\right|}^2+C\underset{i=1}{\overset{H}{\mathrm{\Σ}}}({\mathrm{\ξ}}_i+{\mathrm{\ξ}}_i^{*})$

Wherein: ε return to allow maximum error, and constant C＞0 expression is to the punishment degree of the sample that exceeds error ε, is compromise between the complexity of function f and the sample fitting precision.

Utilize the principle of duality, method of Lagrange multipliers, the dual form of above-mentioned optimization problem is:

$s.t.\left\{\begin{array}{c}\underset{i=1}{\overset{H}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})=0\\ {\mathrm{\α}}_i,{\mathrm{\α}}_i^{*}\∈[0,C]\end{array}\right.$

Can obtain Lagrange multiplier α by separating top equation _i, α _i ^*, and then try to achieve ω, function expression can be written as:

The b that wherein setovers can calculate by KKT (Karush-Kuhn-Tucker) condition.Introduce kernel function

Realize the mapping of former input data to high-dimensional feature space, kernel function commonly used has polynomial kernel function, radially basic kernel function, Sigmoid kernel function.This paper adopts the most frequently used radially basic kernel function:

$K(y_i,y)=\exp (-\frac{{|y-y_i|}^2}{{\mathrm{\σ}}^2})$

So can get:

$f\left(y\right)=\underset{i=1}{\overset{H}{\mathrm{\Σ}}}({\mathrm{\α}}_i-{\mathrm{\α}}_i^{*})K(y_i,y)+b$

Obtain α by training data _i, α _i ^*, b can obtain the analytical expression of f (y) afterwards.Know promptly by this expression formula whether the signal of input sends secondary and report to the police.

Technical conceive of the present invention is: adopt the AR model to carry out albefaction to signal and handle, and calculate whitened signal RMS, compare with the amplitude threshold of presetting, judge whether the one-level warning.Suppose that the signal of gathering is x (n), then its whitened signal is:

$w\left(n\right)=x\left(n\right)+\underset{k=1}{\overset{p}{\mathrm{\Σ}}}{a}_{k}x(n-k)$

Getting window width is 5ms, the RMS of signal calculated:

$\mathrm{RMS}=\sqrt{\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{w}_i^2}$

If RMS surpasses threshold value, then write down this whitened signal, and trigger one-level and report to the police, the notification data process computer carries out next step analysis to signal.

Adopt Short Time Fourier Transform to carry out time frequency analysis to the whitened signal that triggers the one-level warning, obtain the time-frequency figure (seeing accompanying drawing) of signal, then time-frequency figure is compressed.

$\mathrm{STFT}=\underset{-\∞}{\overset{\∞}{\∫}}w\left(t\right)q(t-\mathrm{\τ})e^{-\mathrm{j\ωt}}\mathrm{dt}$

Q is a window function in the following formula, and the time-frequency figure resolution that what-if obtains is m * n, and each row of picture are joined end to end, and obtains the vectorial V of a m * n dimension, and this vector sum projection matrix is multiplied each other:

Y＝PV

Vectorial Y after can obtaining compressing, with respect to raw data, the data after the compression have only original about 1% can express the quantity of information more than 95% in the source data.

The SVM model that the input of data after the compression has trained is discerned, to judge whether that sending secondary reports to the police.

α wherein _iBe model parameter,

Be kernel function.

If f (y)＜0, then expression is once reported to the police and is spurious alarm, otherwise sends the secondary alerting signal, and notice nuclear power station operating personnel further check.

Whether the present invention determines whether that above amplitude threshold needs carry out one-level and report to the police according to the amplitude of vibration signal earlier, and only the signal that the triggering one-level is reported to the police carries out feature identification, sends the secondary warning, has accelerated the recognition efficiency of SVM, has reduced the rate of false alarm of system.

Adopt the AR model that signal has been carried out the albefaction processing, improved the signal to noise ratio (S/N ratio) of signal greatly, reduce the rate of failing to report of system.Employing has the SVM of fine generalization ability as the Classification and Identification device, makes that the discrimination of system is effectively improved.Adopt the PCA method that signal is compressed, removed the correlativity between the data, make data volume reduce, accelerated the efficient of SVM training and identification.

Embodiment two

In conjunction with concrete test, the present invention is described:

1. simulated conditions

Emulation experiment is tested on the steel plate of a 3200mm * 2000mm * 20mm, and steel plate respectively fills up a shockproof damping unit for four jiaos.Used 720g, 185g, 110g, 55g, 30g, 10g the steel ball of totally 6 kinds of different qualities carried out impact test, the position of at every turn falling and highly not limitting is with the different falling position of the different parts simulation loosening element of steel ball bump steel plate.The sensor distribution mode is a triangle in the test.Ground unrest adopts Qinshan nuclear power plant actual measurement noise (comprising the hot noise of pressure vessel container, the hot noise of steam generator, the hot noise of main pump low head, pressure vessel container cold conditions noise, steam generator cold conditions noise, main pump low head cold conditions noise), tests the warning effect at algorithm under different signal to noise ratio (S/N ratio)s, the dynamic change noise conditions and under different in width impulsive noise is disturbed respectively.

2. simulation result

The collision alarm that the different quality loosening element produces is listed in table 1 in following lowest signal-to-noise that can detect of different noise effects.

The lowest signal-to-noise that can detect during table 1 different quality different background noise

Table?1Lowest?SNR?can?be?detected?of?different?loose?part?mass?and?noise

Can find out that from table 1 data the frequency spectrum of detected lowest signal-to-noise of actual energy and noise and the frequency spectrum of collision alarm all have relation, because the quality of loosening element is big more, the signal low-frequency component that collision produces is many more, serious more with the noise aliasing, so for the loosening element of big quality, the lowest signal-to-noise that can detect can be than higher.But the energy of the loosening element of big quality collision is big, and signal to noise ratio (S/N ratio) generally can be very not low, so use influence little to reality.

Figure 4 shows that ground unrest changes the collision alarm treatment effect afterwards that superposes afterwards with sinusoidal rule, can find out that the variation of noise amplitude is little to the influence that detects, signal has been eliminated the fluctuation that changes in amplitude is brought substantially after albefaction.

Algorithm speed is tested, and calculating used CPU is that AMD Turion64x2 dominant frequency is 1.6GHz, internal memory 1GB.Adopt this method processing time length to be that the signal of 16.72s is consuming time and only be 0.52s, as seen can satisfy the requirement that real-time online is handled fully.

Under the situation that noise and impulse disturbances are arranged, the SVM model is tested, sample size is 120,60 normal impingement signals, the pulse signal of 60 different in width, recognition result: false alarm receptance 0%, correct warning reject rate is 0.83%, and total false rate is 0.83%, only has a sample to be misjudged and is false alarm.

The described content of this instructions embodiment only is enumerating the way of realization of inventive concept; protection scope of the present invention should not be regarded as only limiting to the concrete form that embodiment states, protection scope of the present invention also reach in those skilled in the art conceive according to the present invention the equivalent technologies means that can expect.

Claims (3)

1. based on the alarming method of nuclear power station loose part of support vector machine, may further comprise the steps:

1), a plurality of acceleration transducers are installed in nuclear reactor one loop, obtaining the neighbourhood noise in the nuclear reactor, and when loosening element falls, produce, mix the impact signal that neighbourhood noise is arranged;

2), firmly hammer knocks the impact signal of steel plate to produce during loosening element falls in the simulation nuclear reactor with different dynamics, obtains the power hammer and knocks the knocking that steel plate produces, and constitutes the vibration signal storehouse by these knockings;

3), extract the feature of knocking, with knocking feature construction support vector machine (SVM) model of cognition;

4), will carry out the AR modeling by the impact signal that obtains in the step 1), with the impact signal albefaction, and obtain the root mean square of whitened signal, characterize the amplitude of this whitened signal with described root mean square; Default amplitude threshold judges whether the root mean square of whitened signal surpasses amplitude threshold, if then send primary alarm, and write down this whitened signal; If not, then continue monitoring;

5), obtain the whitened signal that triggers primary alarm, extract the target signature of this whitened signal, with in this target signature input SVM, by SVM according to calculating to judge whether this target signature is undesired signal, if then return step 4); If not, then send secondary and report to the police, overhaul to notify the staff.

2. the alarming method of nuclear power station loose part based on support vector machine as claimed in claim 1 is characterized in that: step 2) in, the knocking in the vibration signal storehouse is that the power hammer knocks the vibration signal that different parts produced in nuclear reactor one loop.

3. the alarming method of nuclear power station loose part based on support vector machine as claimed in claim 2 is characterized in that: in the step 3), the feature of extracting signal may further comprise the steps:

3.1) current demand signal that collects is intercepted, get the preceding 0.1s of vibration beginning to beginning back 0.5s;

3.2) adopt Short Time Fourier Transform to carry out time frequency analysis to the signal that is truncated to, window width is 1024 points;

3.2) adopt pivot analysis (PCA) technology to compressing through the signal behind the time frequency analysis, extract feature.

Images