CN104951763A - Power generator set subsynchronous risk evaluating method based on wave recording big data abnormal detection - Google Patents

Abstract

The invention provides a power generator set subsynchronous risk evaluating method based on wave recording big data abnormal detection. The method comprises the following contents of collecting all torsional vibration wave recording files of the same machine unit in a period of time; performing Fourier transformation on the machine end three-phase current signals; extracting industrial frequency components, subsynchronous and supersynchronous harmonic component coefficients; constructing an original feature data set Feature; performing decorrelation and data cleaning on the original feature data set by a main component analysis method; obtaining a dimension reduction feature data set Feature_scomp; building the mahalanobis distance distribution between each sample and the statistics features on the basis of the obtained dimension reduction feature data set; using the value of the mahalanobis distance for measuring the subsynchronous risk; performing the system warning when the risk value reaches 90 percent or higher than 90 percent. The power generator set subsynchronous risk evaluating method has the advantages that the SSO (single sign-on) risk safe evaluation is fast realized, and an auxiliary decision measure is provided for the SSO defense and inhibition in electric network operation.

Description

Based on the subsynchronous methods of risk assessment of genset that the large data exception of record ripple detects

Technical field

The invention belongs to power system stability and control technical field, be specifically related to a kind of subsynchronous methods of risk assessment of genset detected based on the large data exception of record ripple, for power system stability and safe operation provide aid decision making and safe early warning.

Technical background

Under the background of long-distance and large-capacity power transmission demand, owing to adopting string to mend transmission of electricity, alternating current-direct current mixing transmission of electricity and FACTS equipment in a large number, make sub-synchronous oscillation (SSO) serious threat caused the thus safe and stable operation of electrical network and genset.At present, torsional oscillation of the unit protection and vibrational control aspect are laid particular emphasis on to SSO research, vibration mean of defense is lacked to grid operating monitoring.Therefore, further investigate subsynchronous risk assessment and preventive control technology thereof in machine network operation to be significant.

At present; shaft system of unit delivering polarization monitoring/control/protective device creates a large amount of recorded wave file; shaft system of unit tach signal and set end voltage/current signal in real time record operation of power networks; but; to the processing mode of recorded wave file often artificial, manual, analyze targetedly after fault; therefore, accumulate a large amount of history recorded wave files all the year round, do not formed except putting on record for history and operations staff can be helped to understand the knowledge of operation of power networks state.In addition, analyze main from highly sensitive axle system tach signal about SSO, the algorithm loaded down with trivial details by some and technology (as filtering, Damping calculating) process, and electric parameters signal is only used as assistant analysis, and its utilization factor is very little.If the mass data that can make full use of these accumulation excavates the abundant information that recorded wave file contains behind, realize system SSO Fast Identification and safety assessment early warning, greatly will promote the intelligent level of operation.

Summary of the invention

The object of the invention is to provide a kind of subsynchronous risk identification model by a large amount of torsional oscillation recorded wave file, thus realize the safety assessment of SSO risk fast, in operation of power networks, SSO defence and suppression provide aid decision making means.

A kind of subsynchronous methods of risk assessment of genset detected based on the large data exception of record ripple disclosed in this invention, rely on the current signal in a large amount of history recorded wave files of torsional oscillation equipment series accumulation, achieved rapid evaluation and the safe early warning of SSO risk by links such as feature extraction, dimensionality reduction compression, model foundation and risk identification.The present invention is concrete by the following technical solutions:

Based on the subsynchronous methods of risk assessment of genset that the large data exception of record ripple detects, it is characterized in that, described method comprises the steps:

(1) collect same genset run in the torsional oscillation recorded wave file of accumulation, by therefrom extractor end three-phase current signal ia, ib, ic, set up the database that dimension is N × 3, N is the number of recorded wave file;

(2) for object each in database and each phase current signal corresponding to each recorded wave file, extract pattern feature by Fourier transform, structure dimension is the eigenmatrix Feature of N × P, and P is each recorded wave file characteristic of correspondence index number;

(3) eigenmatrix of acquisition is projected to K mutually orthogonal coordinate axis, K<P, form the dimensionality reduction eigenmatrix Feature_comp that dimension is N × K, wherein K mutually orthogonal coordinate direction is decided by the covariance matrix feature decomposition result of Feature;

(4) based on above-mentioned dimensionality reduction eigenmatrix, be a sample depending on each row vector, repeatedly extracted by sample, reliable in parameters estimates and mahalanobis distance adds up the subsynchronous risk identification model set up based on data-driven;

(5) for the new recorded wave file of Real-time Collection, repeated characteristic extracts, dimensionality reduction compression and distance calculate, and according to the subsynchronous risk assessment of unit under the subsynchronous risk identification model realization current record ripple method of operation set up.

The present invention also comprises following preferred version further:

In step (2), the structure of described eigenmatrix Feature specifically comprises:

2.1 by three corresponding for recorded wave file each in the dimensional database of N × 3 objects and three-phase current signal i _a, i _b, i _ccarry out discrete Fourier transformation respectively, obtain the transform sequence I of three-phase current _a, I _b, I _c;

2.2 extract amplitude corresponding to power frequency, subsynchronous and supersynchronous oscillation frequency as pattern feature amount from each transform sequence;

When the torsion frequency of certain genset is respectively f ₁, f ₂..., f _m, m is torsional oscillation mode exponent number, then extract and frequency vector [50-f from each transform sequence _m..., 50-f ₂, 50-f ₁, 50,50+f ₁, 50+f ₂..., 50+f _m] corresponding amplitude vector, obtain each recorded wave file characteristic of correspondence vector:

Feature _j＝[A _j,1… A _j,2m+1B _j,1… B _j,2m+1C _j,1… C _j,2m+1] (1)

In formula, j=1,2 ... N, subscript j represent a jth recorded wave file, and A, B, C represent and I respectively _a, I _b, I _ccorresponding pattern feature, corresponding 2m+1 the characteristic index of each object in database;

2.3 comprehensive all recorded wave file characteristic of correspondence vectors, form the eigenmatrix Feature that dimension is N × P, wherein P=3 × (2m+1);

$\mathrm{Feature}={\left[\begin{array}{ccccc}{\mathrm{Feature}}_1^T& {\mathrm{Feature}}_2^T& ...& {\mathrm{Feature}}_{N-1}^T& {\mathrm{Feature}}_N^T\end{array}\right]}^T---\left(2\right).$

In step (3), obtain dimensionality reduction eigenmatrix specific implementation by rectangular projection and comprise:

3.1 standardized feature matrix F eature, then calculate its covariance matrix Cov;

3.2 couples of covariance matrix Cov carry out Eigenvalues Decomposition, by eigenwert according to order arrangement from big to small, if λ ₁>=λ ₂>=...>=λ _p, normal orthogonal proper vector corresponding is with it designated as γ respectively ₁, γ ₂..., γ _p;

3.3 ask and satisfy condition minimum K value, by the eigenmatrix after standardization to proper vector subspace [γ ₁, γ ₂..., γ _k] projection, obtain dimensionality reduction eigenmatrix Feaure_comp, store substrate and aforesaid proper vector subspace [γ simultaneously ₁, γ ₂..., γ _k].

In step (4), the subsynchronous risk identification model process of establishing based on dimensionality reduction eigenmatrix specifically comprises:

4.1 randomly draw H sample, wherein N/2≤H≤3N/4 from dimensionality reduction eigenmatrix Feaure_comp, calculate its sample average T ₁with covariance matrix S ₁;

4.2 according to sample average T ₁with covariance matrix S ₁, obtain the mahalanobis distance d of all N number of samples _j:

$d_j=\sqrt{(\mathrm{Feature}\_{\mathrm{comp}}_j{-T}_1)S_1^{-1}{(\mathrm{Feature}\_{\mathrm{comp}}_j-T_1)}^T};---\left(3\right)$

Wherein, j=1,2 ... N, Feature_comp _jfor the jth row vector of dimensionality reduction eigenmatrix.

4.3 H the samples selecting corresponding mahalanobis distance minimum from dimensionality reduction eigenmatrix, calculate its sample average T ₂with covariance matrix S ₂, when meeting det (S ₂)=det (S ₁) or det (S ₂during)=0, wherein, det represents determinant computing, by T ₁and S ₁the reliable estimation of T and variance S is expected respectively as dimensionality reduction eigenmatrix population distribution; Otherwise, based on T ₂and S ₂recalculate the mahalanobis distance d of all samples _j, and H the sample selecting corresponding mahalanobis distance minimum, calculate its sample average T ₃with covariance matrix S ₃so repeatedly, until det (S _n+1)=det (S _n) or det (S _n+1iteration is stopped during)=0, and by T _nand S _nexpect that the reliable estimation of T and variance S stores respectively as dimensionality reduction eigenmatrix population distribution;

4.4 based on the expectation T of storage in step 4.3 and the reliable estimator of variance S, and sample mahalanobis distance d obeys card side's distribution that degree of freedom is K, when the mahalanobis distance d of sample meets in time, is considered as exceptional sample and there is subsynchronous risk, and α is level of significance.

In step (5), under the described current method of operation, subsynchronous risk assessment processes specifically comprises:

5.1 for the new recorded wave file of Real-time Collection, by carrying out discrete Fourier transformation to machine end three-phase current signal wherein, extracts and power frequency, subsynchronous and supersynchronous corresponding pattern character vector U;

5.2 substrate [the γ that U is stored in step (3) ₁, γ ₂..., γ _k] projection, obtain compressing vectorial V;

5.3 based on the average T stored in step (4) and covariance S, calculate the mahalanobis distance D of the vectorial V of compression, as P (x≤D) >=95%, think that unit exists subsynchronous risk under the current method of operation, and this value more Risks is larger, wherein card side's distribution variables value of x to be degree of freedom be K.

The present invention has following Advantageous Effects: the present invention is a kind of Risk Identification Method of data-driven, overcomes modeling complexity, the difficult shortcoming determined and can only be used for ex-post analysis of Study first of classic method.Electric parameters signal in the history recorder data that the method accumulates based on unit, by the analysis to data, can set up torsional oscillation risk evaluation model.In the application, according to recording ripple in real time in conjunction with risk evaluation model, effectively can identify the subsynchronous risk situation under the current method of operation of unit fast, genset health status is grasped in time, to take accident prevention in time, there is very important realistic price.

Accompanying drawing explanation

Fig. 1 is the subsynchronous methods of risk assessment process flow diagram of genset disclosed in the present application; Fig. 2 is device recorder data access block diagram (for 3 mode) that the present invention realizes;

Fig. 3 is the eigenmatrix dimensionality reduction process flow diagram that the present invention realizes;

Fig. 4 is the subsynchronous risk identification model Establishing process figure that the present invention realizes;

Fig. 5 is the instantiation effect (wherein, scheming the SSO risk assessment that (a) is dimensionality reduction eigenmatrix dimension selection reference curve, figure (b) is mahalanobis distance distribution plan, figure (c) is part sample to illustrate) that the present invention realizes.

Embodiment

Below in conjunction with Figure of description and embodiment, technical scheme of the present invention is further elaborated.

Be illustrated in figure 1 the subsynchronous methods of risk assessment process flow diagram of genset disclosed in the present application, specific implementation technical scheme of the present invention is as follows:

Step 1: collect recorded wave file, building database

The present embodiment is in conjunction with in a certain thermal power generation unit operational process of certain power plant domestic, and the recorded wave file of CSC-812 shafting torsional oscillation protecting equipment of steam turbo-generator set record is described.As shown in Figure 2, device passes through at shaft system of unit head arranged for redundancy speed probe, and tail place arranges PT/CT, and for the SSO research caused because of a variety of causes provides analysis data, wherein, tach signal sampling rate is 1kHz, and voltage/current signals sampling rate is 600Hz.Device comprises two kinds of record ripple modes, and one is that device starts record ripple, and one records ripple manually.

This device a certain period, have recorded 2917 recorded wave files altogether, start recorded wave files and 48 recorded wave file manually comprising 2869 devices.By extraction machine end three-phase current signal i from each recorded wave file _a, i _b, i _c, set up the database that dimension is 2917 × 3:

Recorded wave file name	Current i a	Current i b	Current i c
				20131016023828TSRLB-0.dat
20131228135153TSRLB-0.dat
				HMI20140413174313.dat
HMI20140412145423.dat
				……

Step 2: extract feature, structural attitude matrix

According to Fig. 2, the corresponding shaft system of unit of the present embodiment comprises high intermediate pressure cylinder, low pressure (LP) cylinder A, low pressure (LP) cylinder B and generator 4 lumped mass blocks, its corresponding 3 shafting torsional oscillation modes (frequency is respectively 12.5Hz, 21.8Hz and 25.5Hz).According to dynamo-electric torsional oscillation interaction mechanism, there is subsynchronous and supersynchronous component corresponding with it in stator current, it is as follows that this realization character extracts concrete steps:

1. for a jth recorded wave file in above-mentioned database, to its three-phase current i _a, i _b, i _ccarry out discrete Fourier transformation respectively, obtain transform sequence I _a, I _b, I _c;

2. from I _a, I _b, I _cthe amplitude vector that middle extraction is respectively corresponding with frequency vector [24.5,28.2,37.5,50,62.5,71.8,75.5], merges all extraction of values and forms this recorded wave file characteristic of correspondence vector Feature _j

Feature _j=[A _{j, 1}a _{j, 7}b _{j, 1}b _{j, 7}c _{j, 1}c _{j, 7}] in (4) formula, j=1,2 ..., 2917, A, B, C represent and i respectively _a, i _b, i _ccorresponding pattern feature;

3. whole 2917 recorded wave file characteristics of correspondence vector in integrated data base, structural attitude matrix F eature, its dimension is 2917 × 21.

$\mathrm{Feature}=\left[\begin{array}{ccccccccc}{A}_{\mathrm{1,1}}& ...& A_{\mathrm{1,7}}& B_{\mathrm{1,1}}& ...& B_{\mathrm{1,7}}& C_{\mathrm{1,1}}& ...& C_{\mathrm{1,7}}\\ A_{\mathrm{2,1}}& ...& A_{\mathrm{2,7}}& B_{\mathrm{2,1}}& ...& B_{\mathrm{2,7}}& C_{\mathrm{2,1}}& ...& C_{\mathrm{2,7}}\\ .& & & & & & & & .\\ .& & & & & & & & .\\ .& & & & & & & & .\\ A_{\mathrm{2917,1}}& ...& A_{2917,7}& B_{\mathrm{2917,1}}& ...& B_{\mathrm{2917,7}}& C_{\mathrm{2917,1}}& ...& C_{\mathrm{2917,7}}\end{array}\right]---\left(5\right)$

Step 3: obtain dimensionality reduction eigenmatrix, stores major bases

The ultimate principle being realized data compression by orthogonal transformation is as follows: be provided with N number of sample, and each sample has P item test index, obtains raw data matrix

$X=\left[\begin{array}{c}{X}_1\\ X_2\\ .\\ .\\ .\\ X_N\end{array}\right]=\left[\begin{array}{cccc}{x}_{11}& x_{12}& ...& x_{1P}\\ x_{21}& x_{22}& ...& x_{2P}\\ .& & & .\\ .& & & .\\ .& & & .\\ x_{N1}& x_{N2}& ...& x_{\mathrm{NP}}\end{array}\right]---\left(6\right)$

For P dimension space one group of Complete Orthogonal base { ω ₁, ω ₂..., ω _pmeet

${\mathrm{\&omega;}}_i^T{\mathrm{\&omega;}}_j=\left\{\begin{array}{cc}0,& i\&NotEqual;j\\ 1,& i=j\end{array}\right.(i,j=\mathrm{1,2},...,P)---\left(7\right)$

Then raw data set X is at P orthogonal dimension base { ω ₁, ω ₂..., ω _punder be projected as

$Y=\left[\begin{array}{c}{X}_1\\ X_2\\ .\\ .\\ .\\ X_N\end{array}\right]\left[\begin{array}{cccc}{\mathrm{\&omega;}}_1& {\mathrm{\&omega;}}_2& ...& {\mathrm{\&omega;}}_P\end{array}\right]---\left(8\right)$

Statistics extensively adopts variance or standard deviation to represent uncertain, variance or standard deviation larger, uncertainty is larger, and quantity of information is larger.Therefore, raw data set X is along a direction ω _ivariance after projection can be expressed as

${\mathrm{Var}}_i=\frac{1}{N-1}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(X_j{\mathrm{\&omega;}}_i-\stackrel{\&OverBar;}{X}{\mathrm{\&omega;}}_i)}^2={\mathrm{\&omega;}}_i^{T}S{\mathrm{\&omega;}}_i---\left(9\right)$

In formula, expectation and the covariance matrix of raw data set is represented respectively with S

$\stackrel{\&OverBar;}{X}=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{X}_j,S=\frac{1}{N-1}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{(X_j-\stackrel{\&OverBar;}{X})}^T(X_j-\stackrel{\&OverBar;}{X})---\left(10\right)$

In order to find direction vector ω _iformula (9) is maximized, introduces Lagrange multiplier λ _i, structure unconstrained optimization problem

$\max ({\mathrm{\&omega;}}_i^{T}S{\mathrm{\&omega;}}_i+{\mathrm{\&lambda;}}_i(1-{\mathrm{\&omega;}}_i^T{\mathrm{\&omega;}}_i))---\left(11\right)$

The condition that this optimization problem obtains extreme point is

Sω _i＝λ _iω _i(12)

Now, raw data set is along ω _ivariance (i.e. quantity of information) after projection is

Var _i＝λ _i(13)

Therefore, in order to the quantity of information of data set is comparatively large after making to project, it is vectorial that projecting direction vector should get the larger eigenwert characteristic of correspondence with covariance.In addition, consider that raw data concentrates measurement unit and the dimension impact of P index, before dimensionality reduction, usually raw data matrix is done standardization.

Finally, the Data Dimensionality Reduction flow process that provides based on orthogonal transformation of composition graphs 3 is as follows:

1. standardized feature matrix F eature, then calculates its covariance matrix Cov;

2. Eigenvalues Decomposition is carried out to covariance matrix Cov, by eigenwert according to order arrangement from big to small, if λ ₁>=λ ₂>=...>=λ _p, normal orthogonal proper vector corresponding is with it designated as γ respectively ₁, γ ₂..., γ _p;

3. ask and satisfy condition minimum K value, by the eigenmatrix after standardization to proper vector subspace [γ ₁, γ ₂..., γ _k] projection, obtain dimensionality reduction eigenmatrix Feaure_comp, store substrate [γ simultaneously ₁, γ ₂..., γ _k].

Based on this realize step 2 obtain 2917 × 21 dimensional feature matrix F eature, Fig. 5 (a) give front k (k=1,2 ..., 21) and tie up the cumulative information amount of data for projection and the ratio curve figure of gross information content.The quantity of information that in figure, front 5 (K=5) of result display tie up data for projection reaches more than 95%, therefore, by Feature to [γ ₁, γ ₂..., γ ₅] projecting obtains the dimensionality reduction eigenmatrix Feature_comp that dimension is 2917 × 5.

Step 4: set up abnormality detection model, storage and distribution center and variance

By sample drawn and the reliable METHOD FOR ESTIMATING POPULATION DISTRIBUTION parameter of interative computation repeatedly, the ultimate principle setting up abnormality detection model according to mahalanobis distance distribution is as follows: the sample matrix Z one being tieed up to observed reading containing K

$Z=\left[\begin{array}{c}{Z}_1\\ Z_2\\ .\\ .\\ .\\ Z_N\end{array}\right]=\left[\begin{array}{cccc}{z}_{11}& z_{12}& ...& z_{1K}\\ z_{21}& z_{22}& ...& z_{2K}\\ .& & & .\\ .& & & .\\ .& & & .\\ z_{N1}& z_{N2}& ...& z_{\mathrm{NK}}\end{array}\right]---\left(14\right)$

Suppose that sample is obeyed K and tieed up normal distribution, order and | H ₁|=H, if det is (S ₁) ≠ 0, calculates mahalanobis distance wherein i=1,2 ..., N.Make H now ₂for H the subsample that mahalanobis distance in the first step is minimum, namely

$H_2:\{Z_{q_1},Z_{q_2},...,Z_{\mathrm{qH}}|d_1\left(q_1\right)\&le;d_1\left(q_2\right)\&le;...\&le;d_1\left(q_H\right)\&le;\underset{i\&NotElement;\{q_1,q_2,...,q_H\}}{d_1\left(i\right)}\}$

Again according to H ₂the average T calculated ₂with covariance S ₂repeat said process, each iteration all can cause det (S) minimization.Number H due to subset is limited, therefore, necessarily there is n and computing is restrained, be i.e. det (S _n+1)=det (S _n), think S _nand T _nfor K ties up the reliable estimation of normal distribution center T and variance S.Obey because the mahalanobis distance d of stochastic variable under the normal distribution of K dimension is similar to card side's distribution that degree of freedom is K, therefore, each sample point is to the corresponding card side's quantile of mahalanobis distance of distribution center make wherein α is the level of significance of corresponding sample.If α is very little, as being less than 0.05, then can think that this sample peels off, and is considered as abnormity point, can infer that this sample point is because special reason for fluctuation a certain in structural system causes.

Finally, based on the dimensionality reduction eigenmatrix Feature_comp that step 3 obtains, composition graphs 4 provides the concrete steps that abnormality detection model is set up:

1. from dimensionality reduction eigenmatrix Feaure_comp, randomly draw H sample, wherein N/2≤H≤3N/4, calculate its sample average T ₁with covariance matrix S ₁;

2. according to sample average T ₁with covariance matrix S ₁, obtain the mahalanobis distance d of all N number of samples _j:

$d_j=\sqrt{(\mathrm{Feature}\_{\mathrm{comp}}_j{-T}_1)S_1^{-1}{(\mathrm{Feature}\_{\mathrm{comp}}_j-T_1)}^T}---\left(15\right)$

Wherein, j=1,2 ... N.

3. from dimensionality reduction eigenmatrix, select H the sample that corresponding mahalanobis distance is minimum, calculate its sample average T ₂with covariance matrix S ₂, when meeting det (S ₂)=det (S ₁) or det (S ₂during)=0, by T ₁and S ₁the reliable estimation of T and variance S is expected respectively as dimensionality reduction eigenmatrix population distribution.Otherwise, based on T ₂and S ₂recalculate the mahalanobis distance d of all samples _j, and H the sample selecting corresponding mahalanobis distance minimum, calculate its sample average T ₃with covariance matrix S ₃so repeatedly, until det (S _n+1)=det (S _n) or det (S _n+1iteration is stopped during)=0, and by T _nand S _nexpect that the reliable estimation of T and variance S stores respectively as dimensionality reduction eigenmatrix population distribution;

4. based on the step 3. middle expectation T of storage and the reliable estimator of variance S, sample mahalanobis distance d obeys card side's distribution that degree of freedom is K, when meeting in time, is considered as exceptional sample and there is subsynchronous risk, and α is level of significance.

Dimension for the acquisition of the present embodiment step 3 is the dimensionality reduction eigenmatrix Feature_comp of 2917 × 5, therefrom repeatedly extract 2187 samples through interative computation reliably estimate dimension be 1 × 5 average T and dimension be the variance S of 5 × 5, and store this two estimators.

T＝[-0.7568 0.1210 0.2137 0.0805 -0.0507]

$S=\left[\begin{array}{ccccc}7.7675& 1.8784& -0.4536& 0.2755& -0.2164\\ 1.8784& 2.1930& -0.3403& 0.0854& -0.0618\\ -0.4536& -0.3403& 1.9062& -0.0316& 0.0191\\ 0.2755& 0.0854& -0.0316& 0.5737& 0.0113\\ -0.2164& -0.0618& 0.0191& 0.0113& 0.5694\end{array}\right]---\left(16\right)$

Fig. 5 (b) gives the mahalanobis distance distribution plan of all 2917 samples of the present embodiment.

Step 5:SSO risk assessment, for the new recorded wave file of Real-time Collection, repeated characteristic extracts, dimensionality reduction compression and distance calculate, and according to the subsynchronous risk assessment of unit under the subsynchronous risk identification model realization current record ripple method of operation set up.

For the present embodiment, mahalanobis distance obeys card side's distribution that degree of freedom is 5, and contrast chi-square distribution table, the mahalanobis distance corresponding when level of significance is 0.05 is 11.07, therefore, provides security alarm and aid decision making when new samples mahalanobis distance is greater than 11.07.Fig. 5 (c) gives the SSO risk evaluation result of part sample.

Above embodiment only understands core concept of the present invention for helping; the present invention can not be limited with this, for those skilled in the art, every according to thought of the present invention; any change done in specific embodiments and applications, all should be included within protection scope of the present invention.

Claims (5)

1., based on the subsynchronous methods of risk assessment of genset that the large data exception of record ripple detects, it is characterized in that, described method comprises the steps:

(1) collect same genset run in the torsional oscillation recorded wave file of accumulation, by therefrom extractor end three-phase current signal ia, ib, ic, set up the database that dimension is N × 3, N is the number of recorded wave file;

(2) for object each in database and each phase current signal of machine end corresponding to each recorded wave file, pattern feature is extracted by Fourier transform, structure dimension is the eigenmatrix Feature of N × P, and P is each recorded wave file characteristic of correspondence index number;

(3) eigenmatrix of acquisition is projected to K mutually orthogonal coordinate axis, K<P, form the dimensionality reduction eigenmatrix Feature_comp that dimension is N × K, wherein K mutually orthogonal coordinate direction is decided by the covariance matrix feature decomposition result of Feature.

(4) based on above-mentioned dimensionality reduction eigenmatrix, be a sample depending on each row vector, repeatedly extracted by sample, reliable in parameters estimates and mahalanobis distance adds up the subsynchronous risk identification model set up based on data-driven.

(5) for the new recorded wave file of Real-time Collection, repeated characteristic extracts, dimensionality reduction compression and distance calculate, and according to the subsynchronous risk assessment of unit under the subsynchronous risk identification model realization current record ripple method of operation set up.

2. a kind of subsynchronous methods of risk assessment of genset detected based on the large data exception of record ripple according to claim 1, it is characterized in that, in step (2), the structure of described eigenmatrix Feature specifically comprises:

2.1 by three corresponding for recorded wave file each in the dimensional database of N × 3 objects and three-phase current signal i _a, i _b, i _ccarry out discrete Fourier transformation respectively, obtain the transform sequence I of three-phase current _a, I _b, I _c;

2.2 extract amplitude corresponding to power frequency, subsynchronous and supersynchronous oscillation frequency as pattern feature amount from each transform sequence;

When the torsion frequency of certain genset is respectively f ₁, f ₂..., f _m, m is torsional oscillation mode exponent number, then extract and frequency vector [50-f from each transform sequence _m..., 50-f ₂, 50-f ₁, 50,50+f ₁, 50+f ₂..., 50+f _m] corresponding amplitude vector, obtain each recorded wave file characteristic of correspondence vector:

Feature _j＝[A _j,1… A _j,2m+1B _j,1… B _j,2m+1C _j,1… C _j,2m+1] (1)

In formula, j=1,2 ... N, subscript j represent a jth recorded wave file, and A, B, C represent and I respectively _a, I _b, I _ccorresponding pattern feature;

2.3 comprehensive all recorded wave file characteristic of correspondence vectors, form the eigenmatrix Feature that dimension is N × P, wherein P=3 × (2m+1);

$\mathrm{Feature}={\left[\begin{array}{ccccc}{\mathrm{Feature}}_1^T& {\mathrm{Feature}}_2^T& ...& {\mathrm{Feature}}_{N-1}^T& {\mathrm{Featur}}_N^T\end{array}\right]}^T---\left(2\right).$

3. a kind of subsynchronous methods of risk assessment of genset detected based on the large data exception of record ripple according to claim 1, is characterized in that, in step (3), obtains dimensionality reduction eigenmatrix specific implementation comprise by rectangular projection:

3.1 standardized feature matrix F eature, then calculate its covariance matrix Cov;

3.2 couples of covariance matrix Cov carry out Eigenvalues Decomposition, by eigenwert according to order arrangement from big to small, if λ ₁>=λ ₂>=...>=λ _p, normal orthogonal proper vector corresponding is with it designated as γ respectively ₁, γ ₂..., γ _p;

3.3 ask and satisfy condition minimum K value, by the eigenmatrix after standardization to proper vector subspace [γ ₁, γ ₂..., γ _k] projection, obtain dimensionality reduction eigenmatrix Feaure_comp, store substrate and aforesaid proper vector subspace [γ simultaneously ₁, γ ₂..., γ _k].

4. a kind of subsynchronous methods of risk assessment of genset detected based on the large data exception of record ripple according to claim 1, it is characterized in that, in step (4), the subsynchronous risk identification model process of establishing based on dimensionality reduction eigenmatrix specifically comprises:

4.1 randomly draw H sample, wherein N/2≤H≤3N/4 from dimensionality reduction eigenmatrix Feaure_comp, calculate its sample average T ₁with covariance matrix S ₁;

4.2 according to sample average T ₁with covariance matrix S ₁, obtain the mahalanobis distance d of all N number of samples _j:

$d_j=\sqrt{(\mathrm{Feature}\_{\mathrm{comp}}_j-T_1)S_1^{-1}{(\mathrm{Feature}\_{\mathrm{comp}}_j-T_1)}^T};---\left(3\right)$

Wherein, Feaure_comp _jrepresent the jth row vector of dimensionality reduction eigenmatrix, j=1,2 ... N;

4.3 H the samples selecting corresponding mahalanobis distance minimum from dimensionality reduction eigenmatrix, calculate its sample average T ₂with covariance matrix S ₂, when meeting det (S ₂)=det (S ₁) or det (S ₂during)=0, wherein, det represents determinant computing, by T ₁and S ₁the reliable estimation of T and variance S is expected respectively as dimensionality reduction eigenmatrix population distribution; Otherwise, based on T ₂and S ₂recalculate the mahalanobis distance d of all N number of samples _j, and H the sample selecting corresponding mahalanobis distance minimum, calculate its sample average T ₃with covariance matrix S ₃so repeatedly, until det (S _n+1)=det (S _n) or det (S _n+1iteration is stopped during)=0, and by T _nand S _nexpect that the reliable estimation of T and variance S stores respectively as dimensionality reduction eigenmatrix population distribution;

4.4 based on the expectation T of storage in step 4.3 and the reliable estimator of variance S, and sample mahalanobis distance d obeys card side's distribution that degree of freedom is K and dimensionality reduction eigenmatrix columns, when the mahalanobis distance d of sample meets in time, is considered as exceptional sample and there is subsynchronous risk, and α is level of significance.

5. a kind of subsynchronous methods of risk assessment of genset detected based on the large data exception of record ripple according to claim 1, it is characterized in that, in step (5), under the described current method of operation, subsynchronous risk assessment processes specifically comprises:

5.1 for the new recorded wave file of Real-time Collection, by carrying out discrete Fourier transformation to machine end three-phase current signal wherein, extracts and power frequency, subsynchronous and supersynchronous corresponding pattern character vector U;

5.2 substrate [the γ that U is stored in step (3) ₁, γ ₂..., γ _k] projection, obtain compressing vectorial V;

5.3 based on the average T stored in step (4) and covariance S, calculate the mahalanobis distance D of the vectorial V of compression, as P (x≤D) >=95%, think that unit exists subsynchronous risk under the current method of operation, and this value more Risks is larger, wherein card side's distribution variables value of x to be degree of freedom be K.