CN104462846A - Intelligent device failure diagnosis method based on support vector machine - Google Patents

Abstract

The invention provides an intelligent device failure diagnosis method based on a support vector machine. The method includes the steps that preprocessing operation is conducted on device data; a failure diagnosis case knowledge base is built; failure diagnosis is conducted on the support vector machine; the failure information is obtained, and troubleshooting guide is conducted. By means of the intelligent device failure diagnosis method based on the support vector machine, the failure feature of a device is highlighted to the maximum degree, the situations that the device data are incomplete and imprecise are reduced, the method provides the possibility for building a precise and reliable failure diagnosis model, the problem of aging of the diagnosis model along with the runtime of the device is solved, the misdiagnosis rate of the failure diagnosis model is reduced, and the correct rate and speed of the device failure diagnosis are increased to the maximum degree.

Description

A kind of equipment failure intelligent diagnosing method based on support vector machine

Technical field

The present invention is mainly concerned with Diagnosis Technique field, is more particularly to and refers in particular to a kind of diagnosis research art processes based on support vector machine.

Background technology

In important events such as some such as power plant, iron-smelter, satellite launch sites, factory is the safe operation guaranteeing its key equipment, often to drop into a large amount of maintainers to ensure the safe operation state of equipment.But staff, due to the problem such as diagnostic techniques or notice of self, unavoidably ignores the sign of some unit exceptions, once unit exception develops into fault in production, will bring huge economic loss to enterprise.Therefore the researchist of related fields drops into great effort for producing enterprise equipment intelligent diagnosis system, guarantees the safe operation of visual plant in factory.

Under general situation, the feature of equipment failure state is seldom that single signal shows, often comprehensively embodied by a lot of characteristic signal, there is falling the fault of vacuum in the condenser for example in power plant steam turbine, not only show in the precipitate trend of condenser vacuum value, also in condenser absolute pressure value, exhaust temperature, there is chain reaction in the characteristic signals such as condensing water temperature.Traditional fault detection method takes corresponding mono signal fault detection method for different characteristic signals, investigate one by one, but this method inefficiency, single fault-signal has limitation on performance fault characteristic, and cannot probe into out the relevance between other fault-signals.

For whole useful informations of total score desorption device fault, the synthetic fault diagnosis technology based on artificial intelligence technology becomes the study hotspot instantly exploring equipment fault diagnosis.Artificial intelligence technology with the Nonlinear Mapping model of fault type by mathematics method for digging structure equipment running status, thus is realized the whole feature of Cooperative Analysis equipment and jointly to feel the pulse the target of equipment failure.But these Nonlinear Modeling diagnostic techniquess still have area for improvement: first, its diagnosis effect has very large relation with device data sample process, and the feature that data sample more can embody corresponding failure type more can improve the efficiency of diagnosis.But existing intellectual technology is too coarse in treatment facility data, be confined on single dimensional normalization and Feature Dimension Reduction, cause the incomplete state of data sample out of true.Secondly, the disaggregated model of current intelligent fault diagnosis technique construction is substantially all unalterable, do not consider continuing along with operation hours, the fault signature of original extraction can be degenerated the decline caused accuracy of identification, and equipment run duration occurs that the novel fault had no is the problem that most diagnostic techniques is instantly ignored, the self how utilizing equipment real time data to realize diagnostic model is also current device fault diagnosis technology urgent problem.Finally, the diagnostic result of existing Intelligent Diagnosis Technology is incomplete to failure message, often lacks supplementing of the necessary informations such as failure-frequency, trouble location information, breakdown maintenance guidance.

For example, at paper " comparative studies of power transformer BP neural network failure diagnosis " (High-Voltage Electrical Appliances, 40th volume, 3rd phase) in, BP Neural Network Diagnosis method is applied in the middle of diagnosing fault of power transformer, and several dissolved gas analysis standards commonly used by improving transformer insulation oil promote the efficiency of Neural Network Diagnosis.Although notice that neural network has the advantageous properties such as parallel processing learning and memory Nonlinear Mapping adaptive ability and robustness, but the initial value of node in hidden layer size, node weights seriously restricts the speed of convergence of neural network, to such an extent as to the Fault Identification efficiency of neural network expectation cannot be reached.

In order to avoid neural network convergence problem and mistake problem concerning study, algorithm of support vector machine is with under the theoretical foundation of structural risk minimization, classification problem in solution small sample limitation situation has the advantage of self, and the classification results of global optimum can be obtained.National patent document " the GIS partial discharge fault type mode identification method based on support vector machine " (number of patent application: CN201310025822.0) proposes the technology utilizing support vector cassification algorithm to carry out the pattern-recognition of GIS partial discharge fault type.Document is told about after the data processing method such as linear normalization, Feature Dimension Reduction, uses the multiple SVM classifier of structure, realizes the identification to multiple discharge fault type.But above method still has weak point, by using the method for 1VS 1 to solve the two or more classification of SVM, but have ignored fault data and belonging to the identical situation of different faults classification poll.Document is very few by failure message simultaneously, cannot eliminate fail operation pass through guidance instruction to lower step.

For above phenomenon, need new Intelligent Fault Diagnosis Technique badly and solve above-mentioned produced problem, realize the collection to the rapid location of equipment state rapid identification, fault type and more specifically failure message, so just amplitude peak may improve accuracy and the speed of equipment fault diagnosis.

Summary of the invention

The object of the invention is to overcome the deficiencies in the prior art, providing a kind of by building fault diagnosis model accurately and reliably, improving the situation that current manual's intelligent failure diagnosis method is not high to equipment failure recognition efficiency.The present invention farthest can highlight the fault signature of equipment by the mode of regression filtering, reducing incomplete, the coarse situation of device data, providing possibility for building fault diagnosis model accurately and reliably; The present invention, on the basis using support vector machine, uses the function of simple and convenient mode implementation model incremental learning, to solve diagnostic model along with the aging problem of operation hours, reduces the misdiagnosis rate of fault diagnosis model; The present invention, under the basis using support vector machine, by identifying novel fault, realizes the constantly perfect of fault knowledge storehouse; The fault diagnosis model that the present invention realizes can provide more detailed failure diagnosis information, it is not only simple fault type, and comprising the breakdown maintenance guiding opinion that this fault belongs to the confidence level of often kind of fault, trouble location and extract from expert knowledge library, amplitude peak improves accuracy and the speed of equipment fault diagnosis.

Based on the equipment failure intelligent diagnosing method of support vector machine, comprise the following steps successively:

(1) for device data carries out pretreatment operation;

(2) fault diagnosis case base is built:

(3) fault diagnosis is carried out to support vector machine;

(4) obtain failure message and carry out maintenance guidance;

Wherein said step (1) comprises the following steps:

Step 1.1: regression filtering process is carried out to device data:

Step 1.2: eliminate dimension to residual error data, utilizes residual error data to follow the ratio of training residual error bound to turn fractional form into, eliminates the dimension impact between each equipment measuring point, fault measuring point feature highlighted, obtain fault fractional data;

Step 1.3: PCA dimension-reduction treatment is carried out to fault data, thus prevent the generation of the overfitting in Nonlinear Modeling process and the little measuring point of weight from upsetting normal mapping relations, Feature Dimension Reduction process is carried out to equipment fractional data.

Preferably, described step 1.1 comprises the following steps:

Step 1.1.1: the history data reading certain hour from device databases, wherein device data comprises fault state data and the nominal situation data of some;

Step 1.1.2: utilize nominal situation data to set up the recurrence filtering model of support vector machine;

Step 1.1.3: all device history service datas obtain residual error data through regression filtering process.

Preferably, described step (2) comprises the following steps:

Step 2.1: extract fault signature, with the main feature measuring point of the mode determination fault of the fault contribution degree of measuring point;

Step 2.2: analyze the degree of association between all kinds of fault, using fault signature similarity as distinguishing standard, obtains the degree of association progression of all kinds of fault:

Step 2.3: for all kinds of fault adds fault type label, the fault type that the degree of association is low is designated single failure type, the fault type that the degree of association is higher can merge and is designated a resultant fault type, and wherein resultant fault type proceeds the Fault Identification of particular type in the Fault Identification of next stage.

Preferably, described step (3) comprises the following steps:

Step 3.1: the disaggregated model of Training Support Vector Machines, utilizes fault data to carry out the training of disaggregated model, comprises the structure of preliminary classification model and the structure of concrete disaggregated model;

Step 3.2: the bandwidth σ parameter and the error penalty factor that utilize cross validation collection optimization gauss kernel function, training data is divided equally 3 equal portions at random, at every turn using 1 wherein part as test set, train as training data for remaining 2 parts, one group of parameter value that discrimination is the highest will as optimum bandwidth σ parameter factors and error penalty factor;

Step 3.3: utilize svm classifier Model Identification fault type, elementary model of cognition is used for determining apparatus data whether single failure type, and concrete model of cognition is for judging the concrete fault type of resultant fault data and exporting failure message;

The incremental learning of step 3.4:SVM disaggregated model, carries out the training of new fault type and the training again of old fault type.

Preferably, described step (4) comprises the following steps:

Step 4.1, obtains fault type and recognition credibility;

Step 4.2, obtain fault location information and with corresponding failure characteristic matching degree;

Step 4.3, according to expertise library inquiry also for maintainer gives maintenance direction.

Preferably, described step 1.1.2 concrete steps are: be that the nominal situation data M of [m × n] equidistantly compresses process by form, after determining extracted data spacing distance d, the data in all moment are asked for the vector that Euclid norm obtains a m dimension, from Euclid's vector that m ties up, l state moment is extracted for extracting distance, wherein with distance d floor data is reduced into packed data T;

Then, packed data T is carried out linear normalization process, eliminate each measuring point data dimension, obtain standard exercise data R, wherein normalization formula is as follows:

Wherein x is the numerical value of equipment measuring point, subscripting be " mark " be the form after normalization, subscripting be " former " be raw data form, subscripting be " max " be the maximal value of this measuring point, subscripting be " min " be the minimum value of this measuring point.

Finally, using standard exercise data R as the training data setting up SVM regression model, using all measuring point parameters as input parameter, and measuring point parameter builds a series of multiple-input and multiple-output SVM regression model as the mode exporting target component successively, with reference to historical data, an assessed value vector is simulated to test data vector.

Preferably, described step 1.2 concrete steps are:

According to healthy residual error data bound T, center normalized is carried out to fault windowed data W, eliminates each measuring point dimension, be specially:

<1> fault windowed data is normalized from small to large successively according to the sequence number of measuring point;

<2> is in measuring point i, and the residual error data being more than or equal to 0 processes in such a way:

Wherein, s _pfor the fault fractional data of measuring point i, T _onfor the residual error data upper limit of measuring point i, ω is the fault windowed data of measuring point i;

<3> is in measuring point i, and the residual error data being less than 0 processes in such a way:

Wherein, s _pfor the fault fractional data of measuring point i, T _onfor the residual error data upper limit of measuring point i, ω is the fault windowed data of measuring point i;

<4> repeats the operation of above-mentioned <2>-<3GreatT.Gre aT.GT two step, until complete the normalized of all measuring points.

Preferably, described step 1.3 concrete steps are:

Utilize PCA method of descent that the fault fractional data S obtained is mapped to low n-dimensional subspace n by orthogonal transformation, be transformed into the feature correlation behind subspace and drop to minimum point, be specially:

Fault fractional data S is solved fault covariance matrix Σ by <1>:

$\mathrm{\&Sigma;}=\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}(S_i-\stackrel{\&OverBar;}{S}){(S_i-\stackrel{\&OverBar;}{S})}^{*}=\left(s_{\mathrm{ij}}\right)$

Wherein, s _ijfor the numerical value of the i-th row, jth row in covariance matrix, n is total number of numerical value in scores vector, S _ifor the i-th fault fractional data in scores vector, for the average of numerical value in scores vector;

<2> solves correlation matrix according to fault covariance matrix Σ

$\hat{R}=\left(r_{\mathrm{ij}}\right),r_{\mathrm{ij}}=\frac{s_{\mathrm{ij}}}{\sqrt{s_{\mathrm{ii}}}\sqrt{s_{\mathrm{jj}}}}$

Wherein, r _ijfor correlation matrix in the i-th row, jth row numerical value, s _ijfor the numerical value of the i-th row, jth row in covariance matrix;

<3> solves ffault matrix major component from fault covariance matrix Σ, obtains eigenvalue λ _iwith proper vector t _i, then its n principal ingredient computing method are:

$\widehat{y_i}=\widehat{t_i^{*}}x,i=\mathrm{1,2},...,n$

Wherein, x is the fault fractional data vector of the i-th measuring point, be the major component numerical value of the i-th measuring point.

<4> asks for the major component feature that front p accumulation contribution rate is greater than 90%, reaches dimensionality reduction effect, and wherein the contribution rate of major component yi is: the accumulation contribution rate of front p is as follows: wherein p≤n, if the accumulation contribution rate of this P feature chosen is greater than 90%, just achieves fault fractional data S and is tieed up the dimensionality reduction object changing into P and tie up by n;

Train SVM fault grader based on the fault data matrix J after dimensionality reduction, carry out fault diagnosis.

The invention has the beneficial effects as follows:

(1) in treatment facility data, adopt the mode of regression filtering and coordinate the normalized method in center, while elimination data dimension, greatly highlight the feature of fault data, more can be beneficial to the speed and precision of training from data aspect lift scheme than the mode of simple employing linear normalization;

(2) adopt in the time period processing mode of device data the method adding Hamming time window, and the signal having the Duplication of half as far as possible to prevent some lofty between two time windows upset by peeling away of isolating, be beneficial to below disaggregated model for the abundant learning and training of fault progression process;

(3) the classification ballot form polytypic problem of SVM adopting 1VS 1 is being solved, avoid the unbalanced problem of training data ratio under one-to-many manner, the defect of error accumulation under directed acyclic mode can be eliminated again simultaneously, improve the reliability of category of model;

(4) under the classification form of 1VS 1, there is the identical problem that cannot draw fault type of poll for avoiding, adopt the pattern of double-deck diagnosis, regard failure collection high for fault correlation degree as resultant fault, resultant fault is by after the preliminary differentiation of preliminary classification model, again through the careful identification of concrete disaggregated model, successfully avoid the problem that cannot draw fault type, effectively improve the degree of accuracy of fault diagnosis system;

(5) incremental learning that disaggregated model is horizontal and longitudinal is realized, new fault type data can be added in fault knowledge storehouse in time, extend the diagnostic area of disaggregated model, again old fault type is trained again simultaneously, strengthen the memory of disaggregated model to old fault type, prevent the degeneration of disaggregated model recognition performance;

(6) more how useful failure message is obtained, such as belong to the reliability of all kinds of fault, there is position in fault, expert gives maintenance and instructs etc., there is provided fault type mode quantity of information abundanter than merely, contribute to maintenance personal's rationally arrangement of next step troubleshooting effort and expansion.

Accompanying drawing explanation

Fig. 1 equipment fault diagnosis method process flow diagram

Fig. 2 support vector machine diagnostic model structural drawing

Fig. 3 SVM is many classification ballot decision machine schematic diagram one to one.

Fig. 4 fault diagnosis system is to the recognition effect figure of test data

The main measuring point trend map of feature of all kinds of fault of Fig. 5 #1 steam turbine equipment

Embodiment

The following detailed description of specific embodiment of the invention; what be necessary to herein means out is; below implement just to further illustrate for of the present invention; limiting the scope of the invention can not be interpreted as; some nonessential improvement and adjustment that this art skilled person makes the present invention according to the invention described above content, still belong to protection scope of the present invention.

The present invention includes four main process, be the preprocessing process of device data respectively, the building process of fault diagnosis case base, support vector machine failure diagnostic process, failure message obtain and maintenance instruction course.The one or two part wherein belongs to device data process category; Part III belongs to Fault Identification category; Part IV belongs to failure message and obtains category.

Based on the equipment failure intelligent diagnosing method of support vector machine, as shown in Figure 1, the whole design of the present invention comprises following process:

Step 1 is the pretreatment operation of device history data, and fundamental purpose is showed as much as possible by the fault signature of process of data preprocessing by equipment failure data, and this process mainly comprises three key steps.

The regression filtering process of step 1.1 device data

Step 1.1.1 obtains device history service data

Selected device data should be the historical data of multiple faults type multi-measuring point exemplary apparatus, roughly operating process is: after obtaining selected measuring point from PI database, uses ACS software from power plant's database, read the history run status data of this relevant device enough time; Then utilize certain screening rule to carry out the division of fault data G and floor data M to all data, numerical value transfinites if occur in data trend, the phenomenon such as fluctuation is violent, ripple disable all can be divided into fault data; Finally utilize expertise knowledge and equipment failure historical record to mate all historical failure data, be divided into dissimilar fault by careful for fault data.

The device data of device databases importing after nan data screening with n measuring point contains m moment, then can regard the vector of a n dimension at the data strip of each observation point in j moment as, can be expressed as:

ω(t _j)＝[ω _j1,ω _j2,ω _j3,...ω _jn] ^T

Then this device data file should be the matrix form of m × n.Concrete form is as follows:

$\left[\begin{array}{c}{\mathrm{\&omega;}}_{11},{\mathrm{\&omega;}}_{12},...{\mathrm{\&omega;}}_{1n}\\ {\mathrm{\&omega;}}_{21},{\mathrm{\&omega;}}_{22},...{\mathrm{\&omega;}}_{2n}\\ ..................\\ ..................\\ ..................\\ {\mathrm{\&omega;}}_{m1},{\mathrm{\&omega;}}_{m2},...{\mathrm{\&omega;}}_{\mathrm{mn}}\end{array}\right]$

Step 1.1.2 sets up SVM and returns filtering model

First, be that the nominal situation data M of [m × n] equidistantly compresses process by form.Concrete grammar is: after determining extracted data spacing distance d, the data in all moment is asked for the vector that Euclid norm can obtain a m dimension, then from Euclid's vector that m ties up, extracts l state moment for extracting distance with distance d, wherein so just floor data can be reduced into packed data T;

Then, packed data T is carried out linear normalization process, to eliminate the skimble-scamble impact of each measuring point data dimension, obtain standard exercise data R, wherein normalization formula is as follows:

Wherein x is the numerical value of equipment measuring point, subscripting be " mark " be the form after normalization, subscripting be " former " be raw data form, subscripting be " max " be the maximal value of this measuring point, subscripting be " min " be the minimum value of this measuring point.

Finally, standard exercise data R can as the training data setting up SVM regression model.SVM Regression is in the feature space of higher-dimension by multiple measuring point Parameter Mapping, then construct strongly-convex problem with target component for exporting, by the feature space Margin Vector (namely support vector machine) that catches constantly Tuning function system go to approach one can solve the problem while but also comprise the complete function expression of input parameter and output parameter complex nonlinear mapping relations.For realizing the output of multi-measuring point parameter, the present invention all builds a series of multiple-input and multiple-output SVM regression model as the mode exporting target component as these measuring point parameters while input parameter successively by using all measuring point parameters.

Nonlinear problem f (X)=W ψ (X)+b is changed into the quadratic programming problem of following introducing Lagrange factor by SVM regression modeling principle exactly:

$\left\{\begin{array}{c}\min \frac{{\left|\right|W\left|\right|}^2}{2}+C\&CenterDot;\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&xi;}}_i,C\&GreaterEqual;0\\ s.t.W\&CenterDot;\mathrm{\&psi;}\left(X\right)+b\&GreaterEqual;y_i-\mathrm{\&epsiv;}-{\mathrm{\&xi;}}_i\\ W\&CenterDot;\mathrm{\&psi;}\left(X\right)+b\&le;y_i+\mathrm{\&epsiv;}+{\mathrm{\&xi;}}_i\\ {\mathrm{\&xi;}}_i\&GreaterEqual;0,i=1...l\end{array}\right.---\left(2\right)$

$\left\{\begin{array}{c}\min \frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{y}_j{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_{j}K(X_i,X_j)-\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i\\ s.t.\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_j{\mathrm{\&alpha;}}_j=\mathrm{0,0}\&le;{\mathrm{\&alpha;}}_j\&le;C(j=\mathrm{1,2}...,n)\end{array}\right.---\left(3\right)$

Wherein optimized Lagrange factor l is the number of support vector machine, the in store Lagrange factor α of SVM model ^*, support vector machine l, slack variable ξ, the important parameter such as error penalty factor.

Utilize standard exercise data R can build n SVM regression model, just can simulate an assessed value vector to test data vector with reference to historical data in real time.

Step 1.1.3 fault data changes into residual error data

Residual error is defined as the difference of instantaneous value and assessed value, and its formula is as follows:

residual＝realTimeValue-forecastValue (4)

Wherein, realTimeValue is instantaneous value, and forecastValue is predicted value, and residual is residual values.

This step mainly comprises two parts, and one is obtain healthy residual error bound and fault residual data, and two is for residual error data section carries out windowing process.

First, after n SVM regression model has built, inputted successively in SVM regression model by the normal training data M of shape as [m × n] and all carry out returning assessment and can obtain the residual matrix of [m × n], each measuring point of residual matrix is worth can be used as healthy residual error bound T most.

Then, according to mode identical above, all kinds of fault data G is inputted SVM regression model respectively successively, after obtaining the assessed value of fault data, turns to all kinds of fault residual data U.

Finally, carry out load time window process respectively to all kinds of residual error data U, reason is the lead-time that each fault is necessary, the data of single time point comprehensively cannot embody the characteristic attribute of fault.

The size of time window is defined as the cycle length that in all fault types of this equipment, the lead-time is the longest, and the residual error data after so just can ensureing elapsed time window process can cover out of order development time, and time window size defining method is:

t _w＝max[t ₁，t ₂，...，t _i，...t _n]

i∈[1，n]

(5)

Wherein, t _ibe the lead-time of i-th fault type, n is fault type number.

The formal character of time window is Hamming window (Hamming).Hamming window (Hamming) can develop the feature in mid-term by retention fault as far as possible, and slacken the unstable feature in fault late period morning, operational failure data can possess periodic feature, are beneficial to expressed intact fault signature attribute.With 1/2 of the velocity magnitude time window of positioning size of time shift window, two windows connected so just have the Duplication of 50%, can ensure the stationarity of similar fault data feature, avoid lofty data mode and be extracted.Here is the expression formula of Hamming window in time domain:

$\begin{array}{c}w\left(n\right)=0.54-0.46\cos \left(\frac{2\mathrm{n\&pi;}}{N-1}\right)\\ n=0,\mathrm{1,2},...,N-1\end{array}---\left(6\right)$

Wherein, w (n) is the window coefficient of the n-th data in window, and N is the size of time window.

Fault data through the concrete grammar of Hamming window process is, after each windowing, data after windowing process are averaged the global feature level of the fault signature representing whole time window, and such fault residual data U changes into fault windowed data W, and the data volume m of each measuring point is reduced to fault windowed data form is as follows:

$\begin{array}{c}\left[\begin{array}{c}{\mathrm{\&omega;}}_{11},{\mathrm{\&omega;}}_{12},...{\mathrm{\&omega;}}_{1n}\\ {\mathrm{\&omega;}}_{21},{\mathrm{\&omega;}}_{22},...{\mathrm{\&omega;}}_{2n}\\ ..................\\ ..................\\ ..................\\ {\mathrm{\&omega;}}_{\stackrel{\&OverBar;}{m}1},{\mathrm{\&omega;}}_{\stackrel{\&OverBar;}{m}2},...{\mathrm{\&omega;}}_{\stackrel{\&OverBar;}{m}n}\end{array}\right]\\ {\mathrm{\&omega;}}_{\mathrm{ij}}=\frac{1}{t_w}\underset{i=1}{\overset{t_w}{\mathrm{\&Sigma;}}}D\left(i\right)*w\left(i\right)\end{array}---\left(7\right)$

Wherein, w _ijbe the numerical value of the i-th row, jth row, D (i) is i-th fault data in window, and w (i) is the window coefficient of i-th fault data.

Step 1.2, residual error data eliminates dimension process

According to healthy residual error data bound T, the fault windowed data W of step 1.1.3 gained is carried out center normalized, eliminate the skimble-scamble problem of each measuring point dimension.Concrete steps are as follows:

(1) fault windowed data is normalized from small to large successively according to the sequence number of measuring point;

(2) in measuring point i, the residual error data being more than or equal to 0 processes in such a way:

Wherein, s _pfor the fault fractional data of measuring point i, T _onfor the residual error data upper limit of measuring point i, ω is the fault windowed data of measuring point i.

(3) in measuring point i, the residual error data being less than 0 processes in such a way:

Wherein, s _pfor the fault fractional data of measuring point i, T _underfor the residual error data upper limit of measuring point i, ω is the fault windowed data of measuring point i.

(4) above-mentioned two step operations are repeated, until complete the normalized of all measuring points.

Such fault windowed data W will be normalized to fault fractional data S.If the advantage done like this is exactly equipment state occurs in this moment occur abnormal alarm, the mark at least one measuring point in the fault fractional data vector in this moment is negative, otherwise whole measuring point will occur without negative fractional value.

Step 1.3, fault data PCA dimension-reduction treatment

The fault fractional data S that step 1.2 can be obtained by principal component analysis (PCA) (PCA) method of descent is mapped to the process of low n-dimensional subspace n by orthogonal transformation.The optimum correlativity considering data in mapping process, so be transformed into the feature correlation behind subspace to drop to minimum point.

The concrete steps of principal component analysis (PCA) (PCA) are as follows:

(1) fault fractional data S is solved fault covariance matrix Σ:

$\mathrm{\&Sigma;}=\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}(S_i-\stackrel{\&OverBar;}{S}){(S_i-\stackrel{\&OverBar;}{S})}^{*}=\left(s_{\mathrm{ij}}\right)---\left(10\right)$

Wherein, s _ijfor the numerical value of the i-th row, jth row in covariance matrix, n is total number of numerical value in scores vector, S _ifor the i-th fault fractional data in scores vector, for the average of numerical value in scores vector.

(2) correlation matrix is solved according to fault covariance matrix Σ

$\hat{R}=\left(r_{\mathrm{ij}}\right),r_{\mathrm{ij}}=\frac{s_{\mathrm{ij}}}{\sqrt{s_{\mathrm{ii}}}\sqrt{s_{\mathrm{jj}}}}---\left(11\right)$

Wherein, r _ijfor correlation matrix in the i-th row, jth row numerical value, sij be the i-th row in covariance matrix, jth row numerical value.

(3) ffault matrix major component is solved from fault covariance matrix Σ,

Mainly obtain eigenvalue λ _iwith proper vector t _i, then its n principal ingredient computing method are:

$\widehat{y_i}=\widehat{t_i^{*}}x,i=\mathrm{1,2},...,n---\left(12\right)$

Wherein, x is the fault fractional data vector of the i-th measuring point, be the major component numerical value of the i-th measuring point.

(4) ask for the major component feature that front p accumulation contribution rate is greater than 90%, reach dimensionality reduction effect

Method is: major component y _icontribution rate be: and the accumulation contribution rate of front p (p≤n) is as follows: if the accumulation contribution rate of this P feature chosen is greater than 90%, just achieves fault fractional data S and tieed up the dimensionality reduction object changing into P and tie up by n.

Fault data matrix J after final dimensionality reduction will be used for training SVM fault grader to carry out fault diagnosis.

Step 2 is the building process of fault diagnosis case base, and this part comprises three necessary steps:

Step 2.1, extracts fault signature

The object of this step determines the performance characteristic of fault, and a class fault is different from another kind of fault and is certain to show on the intensity of anomaly of equipment measuring point.The present invention with the main feature measuring point of the mode determination fault of the fault contribution degree of measuring point,

(1) all fault fractional datas will carry out the determination of the main feature measuring point of fault from small to large according to arrangement sequence number;

(2) the fault score matrix J of the i-th class is supposed _isize is then the fault scores vector form in each moment is as follows:

J _ij＝[j _ij1,j _ij2,...,j _ijt，...,j _ijp] (15)

Wherein, j _ijtbe the fault mark numerical value of the i-th class fault score matrix jth moment, t measuring point.

Scores vector for per moment asks each measuring point contribution degree and accumulation contribution degree successively according to formula below.Choose front K measuring point that accumulation contribution degree the is greater than 70% main measuring point of fault signature as this moment.

When all measuring points obtain the main measuring point of the fault signature in corresponding moment, add up the probability that each measuring point is selected as the main measuring point of fault signature, the measuring point choosing probability to be greater than 70% is picked out as the final main measuring point of the i-th class fault signature.

The formula of measuring point t fault contribution rate is as follows:

$G_{\mathrm{ijt}}=j_{\mathrm{ijt}}/\underset{t=1}{\overset{P}{\mathrm{\&Sigma;}}}{j}_{\mathrm{ijt}}---\left(16\right)$

Wherein, j _ijtbe the fault mark numerical value of the i-th class fault score matrix jth moment, t measuring point, G _ijtbe the fault contribution numerical value of the i-th class fault score matrix jth moment, t measuring point.

The formula of front k measuring point cumulative failure contribution rate is as follows:

$\mathrm{\&Omega;}=\underset{t=1}{\overset{k}{\mathrm{\&Sigma;}}}{j}_{\mathrm{ijt}}/\underset{t=1}{\overset{n}{\mathrm{\&Sigma;}}}{j}_{\mathrm{ijt}}---\left(17\right)$

Wherein, j _ijtbe the fault mark numerical value of the i-th class fault score matrix jth moment, t measuring point, G _ijtfor front K measuring point cumulative failure contribution numerical value.

(3) above-mentioned two step operations are repeated, until the fault signature completing all fault types is decided.

Finally, main for the fault signature of all fault types measuring point is extracted, is kept in fault signature data frame Π.

Step 2.2, the analysis of failure degree of association

This step, using fault signature similarity as distinguishing standard, determines all kinds of fault degree of association rank between any two; If there is k kind fault type, then need to mate k (k+1)/2 group fault type pair.

Degree of association rank is divided into 2 grades: rudimentary and senior.Concrete matching process is as follows:

First, when fault type i and fault type j carry out asking for fault correlation spend time, extract the main measuring point Π of respective fault signature from fault signature data frame Π _iand Π _j

Then, more than ask for common factor H, when the measuring point number of common factor H all accounts for Π _iand Π _jnumber 50% time, just think that the degree of association rank of fault type i and fault type j is for high, otherwise be low.

Finally, degree of association rank right for k (k+1)/2 group fault type is all asked for out, leave in fault type degree of association matrix K for subsequent use.

Step 2.3, adds fault type label

Fault type low for the degree of association is designated single failure type by this step, and the fault type that the degree of association is higher can merge and is designated a resultant fault type, and it will proceed the Fault Identification of particular type in the Fault Identification of next stage.

For the k kind fault type existed, first integer 1 is used to start single failure types all successively as its fault type label, and then successively remaining resultant fault type is identified, identification means is that the sub-fault type of resultant fault type is all designated same integer type label, be kept in fault type label vector, form is as follows:

MT＝[1,2,3,...,K-1,K-1,K-1,K,K] (18)

Wherein, label numeral 0 represents emerging fault type, there will not be label 0 in training data.

Step 3 is training and the failure diagnostic process of svm classifier model, belongs to the fault type recognition part of fault diagnosis, mainly comprises four steps:

Step 3.1, training svm classifier model

This part utilizes fault data to carry out the training of disaggregated model, comprises the training of preliminary classification model and the training of concrete disaggregated model;

The principle of classification summary of SVM: SVM can make point in training set as much as possible away from two classification problems of this plane for seeking optimum segmentation face.

This double optimization problem can be expressed as:

$\left\{\begin{array}{c}\min \mathrm{\&phi;}\left(w\right)=\frac{1}{2}{\left|\right|w\left|\right|}^2\\ s.t.y_i(w\&CenterDot;x_i+b)\&GreaterEqual;1,i=\mathrm{1,2,3},...,n\end{array}\right.---\left(19\right)$

Be translated into dual problem again, form is as follows:

$\left\{\begin{array}{c}\max W\left(a\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i-\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i{a}_j{y}_i{y}_j(x_i\&CenterDot;x_j)\\ w^{*}=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_i{y}_i{x}_i\\ b^{*}=y_i-w\&CenterDot;x_i\\ s.t.\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{a}_i=0,a_i\&GreaterEqual;0,i=\mathrm{1,2,3},...,n\end{array}\right.---\left(20\right)$

Classification function is orientated as the most at last:

$f\left(x\right)=\mathrm{sgn}(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{y}_i{a}_{i}K(x,x_i)+b^{*})---\left(21\right)$

Wherein, x is the data vector of independent variable measuring point, and y is the data vector of target measuring point, and w is the distance value of the divisional plane between two classes, and b is constant vector, a _ifor support vector machine.

The present invention, on the basis of svm classifier principle, sets up 1vs 1 classification ballot decision-making mechanism (as shown in Figure 3).To solve the polytypic problem of SVM.

The construction method of 1vs 1 classification ballot decision-making mechanism is: k kind fault type matches between two, will build k (k+1)/2 support vector disaggregated model.For example there are A, B, C, D tetra-kinds of fault types, so the present invention will set up [AB], [AC], [AD], [BC], [BD], [CD] six svm classifier models altogether.Suppose in four kinds of above fault types simultaneously, A, B, C are all single failure types, and D is resultant fault type comprises sub-fault type E, F, G, [EF], [FG], [EG] three sub-disaggregated models can be derived under the D root of mother stock class model [AD], [BD], [CD].Fault data matrix J in step 1.3 as training data according to svm classifier principle, first trained six mother stock class models with four large class failure modes, and then take the fault data of three groups to train three sub-disaggregated models, the situation of final svm classifier model is as follows:

Single failure	【AB】
					Single failure	【AC】

Resultant fault	【AD】	[EF]	[FG]	[EG]
					Single failure	【BC】
Resultant fault	【BD】	[EF]	[FG]	[EG]
					Resultant fault	【CD】	[EF]	[FG]	[EG]

The distribution situation of table 1 svm classifier model

The bandwidth σ parameter of step 3.2 optimization gauss kernel function

SVM is the approach realizing searching out an optimum segmentation face under Arbitrary Dimensions is ask the method for inner product that feature space is mapped to more high-dimensional reciprocal of duty cycle by kernel function to go to find optimum segmentation face (i.e. support vector machine), and the bandwidth σ of gaussian kernel function and error penalty factor play vital effect for the classifying quality of SVM.Therefore the present invention adopts cross validation collection to come bandwidth σ parameter and the penalty factor of optimization gauss kernel function.

The present invention runs cross validation collection method according to following method:

First, after determining late mother's disaggregated model, the order of subclassification model carries out two important parameter optimizations successively;

Then, for disaggregated model i, each class fault training data is divided equally 3 equal portions at random, at every turn using 1 wherein part as test set, train as training data for remaining 2 parts, use grid data service within the limits prescribed, carry out optimizing operation to the bandwidth parameter σ of SVM and penalty factor, a pair parameter value that discrimination is the highest will as optimum bandwidth σ parameter and penalty factor;

Finally, optimize all mother and sons' disaggregated models in order, be saved in relevant parameter vector P.

Step 3.3, svm classifier model carries out fault diagnosis

The present invention is applicable to as online real time data carries out fault diagnosis, and the idiographic flow of fault diagnosis is as follows:

(1) real time data rt enters in the SVM regression model that process 1 sets up, and carries out regression filtering process successively, normalized, PCA dimensionality reduction data processing be converted into real-time fractional data J;

(2) first, if the words that real-time fractional data J exists negative value turn to fault fractional data J, if namely the words that there is not negative value are judged to trouble-free nominal situation; Then, the mother stock class model collection after fault fractional data J enters the parameter optimization built in step 3.1 is successively classified; Next carries out ballot statistics, judge whether the preliminary type identified is resultant fault type, if, fault fractional data J also needs to enter successively in step 3.2 the subclassification Models Sets after building corresponding parameter optimization and carries out ballot statistics, identify its concrete fault type, if not, the preliminary type identified is exactly concrete fault type, and idiographic flow as shown in Figure 2.

The method of 1vs 1 classification ballot decision-making mechanism is after real time fail fractional data J enters whole mother (son) disaggregated model successively, add up the poll of often kind of fault type, the fault type that number of votes obtained is maximum is the fault type finally determined, if when occurring that the fault type of maximum poll is multiple, then be judged to be new fault type, label is decided to be 0.

Step 3.4, the incremental learning of svm classifier model

The present invention can realize the training again of training to new fault type and old fault type.

In real-time fault diagnosis process, be judged as fault type label 0 if occur, namely occur new fault type, this real time data can be preserved a label is in the data frame of new, and data frame size is [1000 × p]; Be judged as that if occur fault type label is non-zero, namely occur old fault type, this real time data can be preserved corresponding label is that the size of data frame is [1000 × p] in the data frame of old_n (n is fault type label);

(2) if after the data frame of new is filled with data, new fault type data added in original fault training data, start the incremental learning carrying out svm classifier according to the method for step 3.1, the data frame of new can be cleared simultaneously; After if the data frame of old_n is filled with data, old acquaintance is hindered categorical data and add in original fault training data, start the incremental learning carrying out svm classifier according to the method for step 3.1, the data frame of old_n can be cleared simultaneously;

(3) last, the parameter of method to the svm classifier model upgraded according to step 3.2 is optimized, undated parameter vector P.

Process 4 obtains for failure message and overhauls instruction course

This process and process 3 are all the fault diagnosis category of real time data, and belong to the failure message integrated part of real time fail data, this process comprises 3 key steps:

Step 4.1, obtains fault type and recognition credibility

In the present invention, after the identification of each svm classifier model of real time data rt under 1vs 1 form of process 3, can obtain the confidence level belonging to each fault type, the fault type finally determined belongs to the highest fault type of confidence level.

Specifically describe the recognition credibility algorithm how obtaining real time data rt fault diagnosis below:

(1) real time data rt is when using svm classifier model to carry out type identification, fault type can not only be obtained, and the decision value of a signed can be obtained, decision value can regard the vertical range in this data distance optimum segmentation face as, the numerical symbol of this decision value determines identification types obviously, and the size of this decision value absolute value represents the power that it possesses this category attribute.The present invention carries out asking for of recognition credibility ep according to following formula to real time data:

$\mathrm{ep}=100-|\mathrm{rtD}-\max D|*\frac{40}{|\max D-\min D|}---\left(22\right)$

MaxD wherein represents the decision value of maximum absolute value in the training data of this fault type, and minD represents the decision value that in the training data of this fault type, absolute value is minimum, and rtD represents the categorised decision value of real time data.

Be 100% according to the training data recognition credibility ep of formula (22) known decision value maximum absolute value, and the minimum training data recognition credibility ep of decision value absolute value is 60%.Therefore, real time data rt can obtain the confidence level of two classes.

Real time data rt through all svm classifier models, can obtain the confidence level of all fault types successively, shown in the following formula of specific algorithm:

$\begin{array}{c}\mathrm{ep}\_n=\frac{\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{ep}}_i}{N}*100\%\\ i=1,...,N\end{array}---\left(23\right)$

Wherein, n is fault type, and the fault training data of N to be type be n builds the number of svm classifier model, and epi is the confidence score of real time data to fault type i.

(3) fault type that real time data rt finally determines is the maximum type of ep_n value.

Step 4.2, obtain fault location information and with corresponding failure characteristic matching degree

Determine the fault type of real time data rt in step 4.1 after, real time data rt is converted into fault fractional data J vector form, wherein vectorial J mileage value is that negative measuring point is the doubtful position M broken down.

When if the fault type of real time data rt is n type fault, from fault signature data frame Π, consult fault signature Π _ n that corresponding types n is corresponding, and fault signature matching degree solves according to formula below:

$\begin{array}{c}\mathrm{matchDG}=\frac{\mathrm{len}\left(x\right)}{\mathrm{len}(\mathrm{\&Pi;}\_n)}*100\%\\ x=M\&cap;\mathrm{\&Pi;}\_n\end{array}---\left(24\right)$

Wherein, function len () is for asking for length function, and x is the common factor of real time fail measuring point and fault signature, and Π _ n is the fault signature of fault type n.

Step 4.3, according to expertise library inquiry also for maintainer gives maintenance direction

The guiding opinion that the present invention is undertaken keeping in repair by breaking down about this equipment imports in the middle of software, in order to calling at any time.

When equipment make a definite diagnosis there is certain known fault time, equipment failure intelligent diagnosis system starts to read the maintenance suggestion in expert knowledge library under corresponding failure catalogue, show maintainer successively according to the weight size of suggestion, the size of the matchDG value simultaneously obtained according to step 4.2 calculates prepares against maintainer's reasonable arrangement servicing time servicing time.

In order to further illustrate implementation process of the present invention, the present invention obtains the important measuring point data from No. 1 power generator turbine body equipment of certain thermal power plant, to verify the useful subsidy of the present invention to equipment fault diagnosis.

As follows based on support vector machine equipment failure intelligent diagnosing method key step to #1 steam turbine local equipment of the present invention:

One, data prediction operation is carried out to the historical data of #1 steam turbine equipment

Selected #1 steam turbine equipment is one have multiple faults type multi-measuring point exemplary apparatus, this example obtains 15 selected relevant measuring points from power plant PI database, they are main oil pump outlet oil pressure (kpa) respectively, bearing metal temperature (DEG C), turbine speed (r/s), returning-oil temperature (DEG C), bearing amplitude (mm/s), atmospheric pressure (kpa), axial displacement (mm/s), Thrust-vectoring nozzle (DEG C), main steam temperature (DEG C), reheat steam temperature (DEG C), thrust bearing shoe valve metal temperature (DEG C), steam flow (t/h), exhaust temperature (DEG C).

Use ACS software from power plant's database, to read #1 steam turbine equipment from whole historical data on Dec 31,3 years 1 day to 2013 January in 2011, data volume is about 1,300,000; The data of two-and-a-half years on the 31st in 1 day to 2013 May of January in 2011 are used for building svm classifier model as training data by we, use the data on June 1st, 2013 to Dec 31 as the effect of test data inspection svm classifier Symbolic fault diagnosis.

We need to utilize existing screening rule from #1 steam turbine equipment training data, isolate nominal situation data and abnormal data two class data.Nominal situation data represent the normal operating condition of #1 steam turbine equipment, and this partial data just can be used as the training data T building SVM regression model _return.Nominal situation data represent all fault types that #1 steam turbine equipment occurred in this period of history operational process, expertise knowledge and this equipment failure historical record is utilized to mate all historical failure data, dissimilar fault is divided into by careful for whole abnormal data, through fault data types of nuclear to after, contain four kinds of typical faults of steam turbine equipment in the abnormal data of two-and-a-half years: Steam Turbine Over-speed Accident, Steam Turbine Vibration be excessive, run Leaf damages or fracture and the water entering of steam turbine.

Training data T _returndata be mainly used for build regression model: utilize linear normalization method each measuring point dimension of training data to be zoomed between [0,1], then compress training data T in the mode of step 1.2 _returnto about 3000; Finally, 3000 training datas after compressing are used to build the SVM regression model of 10 multiple input single output according to SVM Regression.

Training data T _returnafter being input to the above-mentioned SVM regression model of 10 successively, can be training data and export the identical healthy residual matrix of size; Simultaneously by all kinds of fault data T _faultafter being input to the SVM regression model of 10 respectively successively, also fault data can be changed into fault residual data.It is the characteristic attribute comprehensively excavating fault residual data simultaneously, according to the mode of formula (5), formula (6), it is carried out to the fault windowed data of hamming window process, because four kinds of Steam Turbine Over-speed Accidents, Steam Turbine Vibration are excessive, run Leaf and to damage or fracture and water entering of steam turbine inaction interval are respectively 48.1 seconds, 13.3 seconds, 35.6 seconds, 20.5 seconds, be therefore defined as 50 seconds by unified for the time window size of hamming window.

#1 steam turbine equipment fault windowed data carries out center normalized according to the way of formula (7), (8) again, eliminate the dimension impact of each measuring point, and make a call to the mark of an identical standard situation for the state in each moment, the dimension of the #1 steam turbine equipment fault fractional data after PCA dimensionality reduction reduces to 9 from 13, has screened out high with other feature correlations and that major component contribution degree is low 4 features: reheat steam temperature (DEG C), thrust bearing shoe valve metal temperature (DEG C), steam flow (t/h), exhaust temperature (DEG C).

Two, #1 steam turbine equipment fault diagnosis case base is built

In order to embody the feature of all kinds of fault data of #1 steam turbine equipment better, essential attribute and the degree of association each other of each fault are still disclosed in this part from data plane, therefore still belong to the pretreatment stage of device data.

First, from #1 steam turbine equipment fault fractional data, fault signature is extracted.According to the account form of fault contribution degree and fault accumulation contribution rate, the essential characteristic of often kind of fault type can be extracted better, following table is for the fault fractional data of Steam Turbine Over-speed Accident, the fault contribution degree of each feature measuring point of exposition moment, can find out in this fault of Steam Turbine Over-speed Accident, the feature measuring point of cumulative failure contribution rate more than 70% is primary outlet oil pressure, bearing metal temperature, the main feature measuring point of turbine speed three.

Table 2 Steam Turbine Over-speed Accident fault fractional data measuring point fault contribution degree

According to above mode, the feature of four kinds of major failures of #1 steam turbine equipment can be extracted successively from all kinds of fault:

(1) Steam Turbine Over-speed Accident

Principal character measuring point: main oil pump outlet oil pressure, bearing metal temperature, turbine speed.

(2) Steam Turbine Vibration is excessive

Principal character measuring point: bearing metal temperature, returning-oil temperature, bearing amplitude.

(3) run Leaf to damage or fracture

Principal character measuring point: pressure, axial displacement, bearing metal temperature, returning-oil temperature, Thrust-vectoring nozzle.

(4) water entering of steam turbine

Principal character measuring point: main steam temperature, axial displacement, bearing metal temperature, bearing return oil temperature.

Work under connecing is the height analyzing all kinds of fault correlation degree, can judge that operation Leaf damages or between fracture defect and water entering of steam turbine fault, similarity reaches more than 50% relatively easily according to the decision rule of fault correlation degree in step 2.2, these two kinds of faults first should be merged into a kind of resultant fault to improve the accuracy of fault diagnosis.

According to above analysis, in #1 steam turbine four kinds of fault types, Steam Turbine Over-speed Accident, Steam Turbine Vibration is excessive belongs to single failure type, runs Leaf and to damage or fracture defect and water entering of steam turbine fault merge into a kind of resultant fault type.

Therefore, in this example, we are Steam Turbine Over-speed Accident, Steam Turbine Vibration is excessive adds fault type label 1,2 respectively, and resultant fault type adds label 3, and the sub-failure operation Leaf under resultant fault type damages or rupture and the water entering of steam turbine adds label 4,5.

Three, disaggregated model and the fault diagnosis of #1 steam turbine is trained

For solving four kinds of faults of svm classifier #1 steam turbine, according to the 1vs1 thinking of step 3.1, we are according to svm classifier principle, construct four svm classifier models to carry out the identification of four kinds of fault types.

(1) pattern master A is for identifying that major break down is crossed in Steam Turbine Over-speed Accident fault and Steam Turbine Vibration, adds label 1,2 respectively;

(2) pattern master B is for identifying Steam Turbine Over-speed Accident fault and resultant fault, adds label 1,2 respectively;

(3) pattern master C is for identifying that major break down and resultant fault are crossed in Steam Turbine Vibration, adds label 1,2 respectively;

(4) submodel D is for identifying that running Leaf damages or fracture defect and water entering of steam turbine fault, interpolation label 1,2 respectively.

Above all disaggregated models all adopt the method for cross validation collection to optimize gaussian kernel function bandwidth σ parameter in SVM model and error penalty factor according to step 3.2, and parameter optimization method adopts grid data service to carry out optimizing, wherein gaussian kernel function bandwidth σ scope is [0.1,10], step-size in search is 0.1, error penalty factor scope is [1,100], and step-size in search is 1.Training recognition effect after optimization is as follows:

	Optimal parameter [σ/C]	Test set A	Test set B	Test set C
					Pattern master A	【0.58，10.61】	91％	85％	93％
Pattern master B	【1.06，5.72】	92％	89％	90％
					Pattern master C	【0.78，20.51】	82％	94％	93％
Submodel D	【2.15，7.09】	88％	95％	92％

The recognition effect of table 3 four disaggregated models under optimal parameter

On June 1st, 2013 is carried out fault diagnosis to the data on Dec 31 as test data by this example.

In order to intactly capture all fault types of #1 steam turbine equipment, in this example set device Fault Intelligent Diagnosis System operationally between i when being 25 seconds integral multiples, read i-th-49 to i of #1 steam turbine equipment and amount to 50 test datas and carry out fault diagnosis.First these 50 test datas are carried out data prediction operation, the SVM regression filtering namely in process 1, center normalization, load time window, Feature Dimension Reduction four kinds of preprocess methods, be finally converted into a real time fail fractional data J; Then, whether comprise negative value in inspection real time fail fractional data J, if having, for abnormal data needs further fault diagnosis, if without, for nominal situation data are without the need to carrying out fault diagnosis.After being judged to abnormal data, real time fail fractional data J enters pattern master A, B, C successively, if be identified as the words of single failure type, fault type recognition completes, if be identified as resultant fault type, need continue to enter submodel D and continue to identify, obtain final fault type.If there is the situation that maximum poll is identical, then sentenced by fault type and make new fault type, type is 0.

Citing, after this section of real time data access arrangement Fault Intelligent Diagnosis System on October 9,11: 25 10 o'clock on the 9th 35 minutes to 2013 October in 2013, changes into real time fail fractional data J=[89.5 ,-5.3,65.1 ,-25.2,78.5 ,-38.4,64.3,93.0,25.2].Check out to there is negative value phenomenon, should sentence and make abnormal data, the result entering three pattern masters A, B, C is successively followed successively by 1,2,2.Resultant fault type obtains two tickets, and continuing to enter submodel D judged result is 2, and final recognition result is water entering of steam turbine fault.

What show in accompanying drawing 4 is the diagnosis and distinguish situation of #1 steam turbine equipment fault data on June 1st, 2013 to Dec 31, can find out:

(1) be 85% to the recognition accuracy of Steam Turbine Over-speed Accident fault;

(2) recognition accuracy crossing major break down to Steam Turbine Vibration is 80%;

(3) to damage or the recognition accuracy of fracture defect is 95% running Leaf;

(4) be 75% to the recognition accuracy of water entering of steam turbine fault.

Satisfactory to the recognition accuracy whole structure of four kinds of faults.

In this example, have four data frame, a large I of new data frame loads 1000 real time fail fractional datas; Three large I of old_n data frame load 1000 real time fail fractional datas.If after data frame is loaded with full data, according to the training expansion of method realization to #1 steam turbine fault type and the training again of old fault type of step 3.4, four models built can not be degenerated with steam turbine equipment development working time, possess good adaptive ability.

Four, real time fail acquisition of information and maintenance instruction course

This process and process three are all the fault diagnosis category of real time data, belong to the failure message integrated part of real time fail data.

In this example, after the identification of each svm classifier model of #1 steam turbine equipment real time fail fractional data J under 1vs 1 form of process 3, not only can obtain fault type, and also can obtain according to the method for step 4.1 confidence level that real time fail fractional data J belongs to every class fault type.

According to the recognition credibility algorithm of fault diagnosis, the confidence level that #1 steam turbine equipment real-time testing data i belongs to often kind of fault type is: Steam Turbine Over-speed Accident 5.2%, Steam Turbine Vibration be excessive 1.3%, run Leaf damage or fracture 20.3%, the water entering of steam turbine 73.2%.The fault type that real time fail fractional data J finally determines is that confidence value is maximum, and output fault type is the water entering of steam turbine.

This example can obtain #1 steam turbine fault location information and with corresponding failure characteristic matching degree, concrete grammar is as follows:

Foundation real time fail fractional data vector J mileage value is the rule that negative measuring point is doubtful position of breaking down, and in fault signature data frame Π, the main measuring point of water entering of steam turbine fault signature is main steam temperature, axial displacement, bearing metal temperature, bearing return oil temperature.Real time fail fractional data vector J mileage value is that negative measuring point is axial displacement, bearing metal temperature, bearing return oil temperature are 75%.

When equipment make a definite diagnosis #1 steam turbine equipment there is the water entering of steam turbine fault time, equipment failure intelligent diagnosis system starts to read the maintenance suggestion in expert knowledge library under water entering of steam turbine defect list, and following two the maintenance suggestions giving maintainer according to the weight size of suggestion are:

1, when confirming steam turbine generation water slug, destroying vacuum emergency stop immediately, cutting off relevant vapour, water source, strengthen the hydrophobic of the regarding system such as master, the female pipe of reheating steam pipe road, body extraction line, shaft seal vapour;

2, unit normally runs or opens, shuts down and in varying load process, and main, reheat steam temperature sharply declined 50 DEG C in 10 minutes, should not destroy vacuum and beat gate stop-start.

Although for illustrative purposes; describe illustrative embodiments of the present invention; but it should be appreciated by those skilled in the art that; when not departing from scope of invention disclosed in claims and spirit; the change of various amendment, interpolation and replacement etc. can be carried out in form and details; and all these change the protection domain that all should belong to claims of the present invention; and application claims protection each department of product and method in each step, can combine with the form of combination in any.Therefore, be not intended to limit the scope of the invention to the description of embodiment disclosed in the present invention, but for describing the present invention.Correspondingly, scope of the present invention not by the restriction of above embodiment, but is limited by claim or its equivalent.

Claims (8)

1., based on an equipment failure intelligent diagnosing method for support vector machine, it is characterized in that, comprise the following steps successively:

(1) for device data carries out pretreatment operation;

(2) fault diagnosis case base is built:

(3) fault diagnosis is carried out to support vector machine;

(4) obtain failure message and carry out maintenance guidance;

Wherein said step (1) comprises the following steps:

Step 1.1: regression filtering process is carried out to device data:

Step 1.2: eliminate dimension to residual error data, utilizes residual error data to follow the ratio of training residual error bound to turn fractional form into, eliminates the dimension impact between each equipment measuring point, fault measuring point feature highlighted, obtain fault fractional data;

Step 1.3: PCA dimension-reduction treatment is carried out to fault data, thus prevent the generation of the overfitting in Nonlinear Modeling process and the little measuring point of weight from upsetting normal mapping relations, Feature Dimension Reduction process is carried out to equipment fractional data.

2., as claimed in claim 1 based on the equipment failure intelligent diagnosing method of support vector machine, it is characterized in that: described step 1.1 comprises the following steps:

Step 1.1.1: the history data reading certain hour from device databases, wherein device data comprises fault state data and the nominal situation data of some;

Step 1.1.2: utilize nominal situation data to set up the recurrence filtering model of support vector machine;

Step 1.1.3: all device history service datas obtain residual error data through regression filtering process.

3., as claimed in claim 1 or 2 based on the equipment failure intelligent diagnosing method of support vector machine, it is characterized in that: described step (2) comprises the following steps:

Step 2.1: extract fault signature, with the main feature measuring point of the mode determination fault of the fault contribution degree of measuring point;

Step 2.2: analyze the degree of association between all kinds of fault, using fault signature similarity as distinguishing standard, obtains the degree of association progression of all kinds of fault:

Step 2.3: for all kinds of fault adds fault type label, the fault type that the degree of association is low is designated single failure type, the fault type that the degree of association is higher can merge and is designated a resultant fault type, and wherein resultant fault type proceeds the Fault Identification of particular type in the Fault Identification of next stage.

4., as claimed in claim 1 based on the equipment failure intelligent diagnosing method of support vector machine, it is characterized in that: described step (3) comprises the following steps:

Step 3.1: the disaggregated model of Training Support Vector Machines, utilizes fault data to carry out the training of disaggregated model, comprises the structure of preliminary classification model and the structure of concrete disaggregated model;

Step 3.2: the bandwidth σ parameter and the error penalty factor that utilize cross validation collection optimization gauss kernel function, training data is divided equally 3 equal portions at random, at every turn using 1 wherein part as test set, train as training data for remaining 2 parts, one group of parameter value that discrimination is the highest will as optimum bandwidth σ parameter factors and error penalty factor;

Step 3.3: utilize svm classifier Model Identification fault type, elementary model of cognition is used for determining apparatus data whether single failure type, and concrete model of cognition is for judging the concrete fault type of resultant fault data and exporting failure message;

The incremental learning of step 3.4:SVM disaggregated model, carries out the training of new fault type and the training again of old fault type.

5., as claimed in claim 1 based on the equipment failure intelligent diagnosing method of support vector machine, it is characterized in that: described step (4) comprises the following steps:

Step 4.1, obtains fault type and recognition credibility;

Step 4.2, obtain fault location information and with corresponding failure characteristic matching degree;

Step 4.3, according to expertise library inquiry also for maintainer gives maintenance direction.

6. as claimed in claim 2 based on the equipment failure intelligent diagnosing method of support vector machine, it is characterized in that: described step 1.1.2 concrete steps are: be that the nominal situation data M of [m × n] equidistantly compresses process by form, after determining extracted data spacing distance d, the data in all moment are asked for the vector that Euclid norm obtains a m dimension, from Euclid's vector that m ties up, l state moment is extracted for extracting distance, wherein with distance d floor data is reduced into packed data T;

Then, packed data T is carried out linear normalization process, eliminate each measuring point data dimension, obtain standard exercise data R, wherein normalization formula is as follows:

Wherein x is the numerical value of equipment measuring point, subscripting be " mark " be the form after normalization, subscripting be " former " be raw data form, subscripting be " max " be the maximal value of this measuring point, subscripting be " min " be the minimum value of this measuring point;

Finally, using standard exercise data R as the training data setting up SVM regression model, using all measuring point parameters as input parameter, and measuring point parameter builds a series of multiple-input and multiple-output SVM regression model as the mode exporting target component successively, with reference to historical data, an assessed value vector is simulated to test data vector.

7., as claimed in claim 1 based on the equipment failure intelligent diagnosing method of support vector machine, it is characterized in that: described step 1.2 concrete steps are:

According to healthy residual error data bound T, center normalized is carried out to fault windowed data W, eliminates each measuring point dimension, be specially:

<1> fault windowed data is normalized from small to large successively according to the sequence number of measuring point;

<2> is in measuring point i, and the residual error data being more than or equal to 0 processes in such a way:

Wherein, s _pfor the fault fractional data of measuring point i, T _onfor the residual error data upper limit of measuring point i, ω is the fault windowed data of measuring point i;

<3> is in measuring point i, and the residual error data being less than 0 processes in such a way:

Wherein, s _pfor the fault fractional data of measuring point i, T _onfor the residual error data upper limit of measuring point i, ω is the fault windowed data of measuring point i;

<4> repeats the operation of above-mentioned <2>-<3GreatT.Gre aT.GT two step, until complete the normalized of all measuring points.

8. the equipment failure intelligent diagnosing method based on support vector machine as described in claim 1 or 7, is characterized in that: described step 1.3 concrete steps are:

Utilize PCA method of descent that the fault fractional data S obtained is mapped to low n-dimensional subspace n by orthogonal transformation, be transformed into the feature correlation behind subspace and drop to minimum point, be specially:

Fault fractional data S is solved fault covariance matrix Σ by <1>:

$\mathrm{\&Sigma;}=\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}(S_i-\stackrel{\&OverBar;}{S}){(S_i-\stackrel{\&OverBar;}{S})}^{*}=\left(s_{\mathrm{ij}}\right)$

Wherein, s _ijfor the numerical value of the i-th row, jth row in covariance matrix, n is total number of numerical value in scores vector, S _ifor the i-th fault fractional data in scores vector, for the average of numerical value in scores vector;

<2> solves correlation matrix according to fault covariance matrix Σ

$\hat{R}=\left(r_{\mathrm{ij}}\right),r_{\mathrm{ij}}=\frac{s_{\mathrm{ij}}}{\sqrt{s_{\mathrm{ii}}}\sqrt{s_{\mathrm{jj}}}}$

Wherein, r _ijfor correlation matrix in the i-th row, jth row numerical value, s _ijfor the numerical value of the i-th row, jth row in covariance matrix;

<3> solves ffault matrix major component from fault covariance matrix Σ, obtains eigenvalue λ _iwith proper vector t _i, then its n principal ingredient computing method are:

$\widehat{y_i}=\widehat{t_i^{*}}x,i=\mathrm{1,2},...,n$

Wherein, x is the fault fractional data vector of the i-th measuring point, be the major component numerical value of the i-th measuring point;

<4> asks for the major component feature that front p accumulation contribution rate is greater than 90%, reaches dimensionality reduction effect, wherein major component y _icontribution rate be: the accumulation contribution rate of front p is as follows: wherein p≤n, if the accumulation contribution rate of this P feature chosen is greater than 90%, just achieves fault fractional data S and is tieed up the dimensionality reduction object changing into P and tie up by n;

Train SVM fault grader based on the fault data matrix J after dimensionality reduction, carry out fault diagnosis.