CN104462846B - A kind of equipment fault intelligent diagnosing method based on SVMs - Google Patents

Abstract

A kind of equipment fault intelligent diagnosing method based on SVMs, including carry out pretreatment operation for device data；Build fault diagnosis case base；Fault diagnosis is carried out to SVMs；Obtain fault message and carry out maintenance guidance, equipment fault intelligent diagnosing method based on SVMs farthest highlights the fault signature of equipment, reduce incomplete, the inaccurate situation of device data, possibility is provided to build accurately and reliably fault diagnosis model, diagnostic model is solved the problems, such as with operation hours aging, the misdiagnosis rate of fault diagnosis model is reduced, amplitude peak improves the accuracy and speed of equipment fault diagnosis.

Description

A kind of equipment fault intelligent diagnosing method based on SVMs

Technical field

Present invention relates generally to Diagnosis Technique field, more particularly relates to refer in particular to one kind based on support The diagnosis research field method of vector machine.

Background technology

In important events such as some power plant, iron-smelter, satellite launch sites, factory is the safety for ensuring its key equipment Operation, often ensures the safe operation state of equipment to put into substantial amounts of maintainer.But staff due to itself The problems such as diagnostic techniques or notice, unavoidably ignore the sign of some unit exceptions, made a living once unit exception develops into Failure is produced, huge economic loss will be brought to enterprise.Therefore the researcher of related fields puts into great effort as production Enterprise establishes equipment intelligent diagnosis system, to ensure the safe operation of visual plant in factory.

Under general scenario, the feature of equipment failure state is seldom that single signal shows, often by many features What signal synthesis embodied, for example there is the failure of vacuum in the condenser in power plant steam turbine, not only shows condenser In the precipitate trend of vacuum values, also occur in characteristic signals such as condenser absolute pressure value, exhaust temperature, condensing water temperatures Chain reaction.Traditional fault detection method is to take corresponding mono signal fault detection method for different characteristic signals, Investigated one by one, but this method efficiency is low, single fault-signal has limitation, and nothing on performance fault characteristic Method is probed into out and the relevance between other fault-signals.

For whole useful informations of comprehensive analysis equipment fault, the synthetic fault diagnosis technology based on artificial intelligence technology into To explore the study hotspot of equipment fault diagnosis instantly.Artificial intelligence technology constructs equipment running status by mathematics method for digging With the Nonlinear Mapping model of fault type, so as to realize that Cooperative Analysis equipment whole feature is felt the pulse the mesh of equipment fault jointly Mark.The improved place in need but these Nonlinear Modeling diagnostic techniques remain unchanged：First, its diagnosis effect is with device data sample Present treatment has the efficiency that the characteristics of very big relation, data sample can more embody corresponding failure type can more improve diagnosis.But Existing intellectual technology is excessively coarse in terms of processing equipment data, is confined on single dimensional normalization and Feature Dimension Reduction, Cause the inaccurate incomplete state of data sample.Secondly, the disaggregated model of intelligent fault diagnosis technique construction is substantially all at present It is unalterable, does not account for continuing with operation hours, the fault signature originally extracted, which can degenerate, is caused to knowing The decline of other precision, and the novel fault for occurring having no during equipment is run is the problem of most diagnostic techniques instantly are ignored, such as What realizes that the self-renewing of diagnostic model is also that current device fault diagnosis technology is badly in need of asking for solution using equipment real time data Topic.Finally, the diagnostic result of existing Intelligent Diagnosis Technology is incomplete to fault message, often lacks failure-frequency, trouble location The supplement of the necessary informations such as information, breakdown maintenance guidance.

For example, in paper《The comparative studies of power transformer BP neural network fault diagnosis method》(High-Voltage Electrical Appliances, the 40th Volume, the 3rd phase) in, BP neural network diagnosis is applied among diagnosing fault of power transformer, and it is exhausted by improving transformer Edge oil conventional several dissolved gas analysis standards lift the efficiency of Neural Network Diagnosis.Note that while neutral net has The advantageous property such as parallel processing learning and memory Nonlinear Mapping adaptive ability and robustness, but node in hidden layer is big Small, node weights initial values seriously restrict the convergence rate of neutral net, so that it cannot reach the desired event of neutral net Hinder recognition efficiency.

In order to avoid neutral net convergence problem and problem concerning study is crossed, algorithm of support vector machine is with Structural risk minization Under the theoretical foundation of change, possess the advantage of itself in the classification problem in the case of solving small sample limitation, and the overall situation can be obtained Optimal classification results.National patent document《GIS partial discharge fault type mode identification method based on SVMs》 (number of patent application：CN201310025822.0) propose and carry out GIS partial discharge failure using support vector cassification algorithm The technology of type-scheme identification.Document is told about after the data processing methods such as linear normalization, Feature Dimension Reduction, uses construction Multiple SVM classifiers, to realize the identification to a variety of discharge fault types.But above method still has weak point, passes through Solves the two or more classification of SVM using 1VS 1 method, but have ignored fault data, to belong to different faults classification poll identical Situation.Document is very few by fault message simultaneously, can not eliminate fail operation to lower step and pass through guidance instruction.

For above phenomenon, new intelligent Fault Diagnosis Technique is needed badly to solve above-mentioned produced problem, realization pair is set Standby state rapid identification, the rapid positioning of fault type and the collection of more specific fault message, are so only possible to amplitude peak and carry The high accuracy and speed of equipment fault diagnosis.

The content of the invention

Examined it is an object of the invention to overcome the deficiencies of the prior art and provide a kind of by building accurately and reliably failure Disconnected model, improve current manual's intelligent failure diagnosis method situation not high to equipment fault recognition efficiency.The present invention is by returning Return the mode of filtering farthest to highlight the fault signature of equipment, reduce incomplete, the inaccurate situation of device data, be Accurately and reliably fault diagnosis model provides possibility to structure；The present invention is on the basis of with SVMs, using simple The easily function of mode implementation model incremental learning, to solve the problems, such as diagnostic model with operation hours aging, drop The misdiagnosis rate of low fault diagnosis model；The present invention, by identifying novel fault, realizes event under the basis with SVMs Hinder the constantly improve of knowledge base；The fault diagnosis model that the present invention realizes can provide more detailed failure diagnosis information, not only Simply simple fault type, and belong to the confidence level of every kind of failure, trouble location comprising this failure and know from expert Know the breakdown maintenance guiding opinion extracted in storehouse, amplitude peak improves the accuracy and speed of equipment fault diagnosis.

Equipment fault intelligent diagnosing method based on SVMs, comprises the following steps successively：

(1) pretreatment operation is carried out for device data；

(2) fault diagnosis case base is built：

(3) fault diagnosis is carried out to SVMs；

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

Wherein described step (1) comprises the following steps：

Step 1.1：Regression filtering processing is carried out to device data：

Step 1.2：Dimension is eliminated to residual error data, using residual error data with training the ratio of residual error bound to be turned into fraction Form, the dimension impact between each equipment measuring point is eliminated, failure measuring point feature is highlighted, obtain failure fraction number According to；

Step 1.3：PCA dimension-reduction treatment is carried out to fault data, so as to prevent the overfitting during Nonlinear Modeling Generation and the small measuring point of weight upset normal mapping relations, Feature Dimension Reduction processing is carried out to equipment fractional data.

Preferably, the step 1.1 comprises the following steps：

Step 1.1.1：The history data of certain time is read in slave unit database, wherein device data includes one The fault state data of fixed number amount and nominal situation data；

Step 1.1.2：The recurrence filtering model of SVMs is established using nominal situation data；

Step 1.1.3：All device history service datas handle to obtain residual error data through regression filtering.

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

Step 2.1：Fault signature is extracted, the main feature measuring point of failure is determined in a manner of the failure contribution degree of measuring point；

Step 2.2：The degree of association between all kinds of failures is analyzed, using fault signature similarity as discrimination standard, is obtained each The degree of association series of class failure：

Step 2.3：Fault type label is added for all kinds of failures, the low fault type of the degree of association is identified as single failure class Type, the higher fault type of the degree of association, which can merge, is identified as a resultant fault type, and wherein resultant fault type is in next stage Fault Identification in continue the Fault Identification of particular type.

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

Step 3.1：The disaggregated model of Training Support Vector Machines, the training of disaggregated model is carried out using fault data, including The structure of the structure of preliminary classification model and specific disaggregated model；

Step 3.2：Using the bandwidth σ parameters and error penalty factor of cross validation collection optimization gauss kernel function, will train Data divide equally 3 equal portions at random, and every time using 1 part therein as test set, remaining 2 parts are trained as training data, know One group of parameter value of rate highest will not be used as optimum bandwidth σ parameter factors and error penalty factor；

Step 3.3：Using svm classifier Model Identification fault type, primary identification model is used for whether judging device data Single failure type, specific identification model are used for the specific fault type for judging resultant fault data and output fault message；

Step 3.4：The incremental learning of svm classifier model, carry out training with old fault type again for new fault type Training.

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

Step 4.1, fault type and recognition credibility are obtained；

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

Step 4.3, maintenance direction is given according to expertise library inquiry and for maintainer.

Preferably, the step 1.1.2 is concretely comprised the following steps：Form is carried out equidistantly for the nominal situation data M of [m × n] Compression is handled, and after extraction data spacing distance d is determined, the data at all moment is asked for into Euclid norm and obtain one The vector of m dimensions, it is to extract the l state moment of extraction in Euclid's vector for being tieed up from m of distance using distance d, wherein Floor data is reduced into compressed data T；

Then, compressed data T is subjected to linear normalization processing, eliminates 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 " for the form after normalization, subscripting is " original " For raw data form, subscripting is the maximum for the measuring point of " max ", and subscripting is the minimum for the measuring point of " min " Value.

Finally, using standard exercise data R as the training data for establishing SVM regression models, using all measuring point parameters as Input parameter, and mode of the measuring point parameter successively as output target component builds a series of multiple-input and multiple-output SVM Regression model, one assessed value vector is fitted with reference to historical data to test data vector.

Preferably, the step 1.2 concretely comprises the following steps:

According to healthy residual error data bound T, center normalized is carried out to failure windowed data W, eliminates each survey Point dimension, it is specially：

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

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

Wherein, s_pFor measuring point i failure fractional data, T_OnFor the measuring point i residual error data upper limit, ω is measuring point i failure Windowed data；

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

Wherein, s_pFor measuring point i failure fractional data, T_OnFor the measuring point i residual error data upper limit, ω is measuring point i failure Windowed data；

<4>Repeat above-mentioned<2>-<3>Two steps operate, until completing the normalized of all measuring points.

Preferably, the step 1.3 concretely comprises the following steps:

Obtained failure fractional data S is mapped to lower-dimensional subspace by orthogonal transformation using PCA method of descents, is transformed into Feature correlation behind subspace bottoms out, and is specially：

<1>Failure fractional data S is solved into failure covariance matrix Σ：

Wherein, s_ijFor the i-th row in covariance matrix, the numerical value of jth row, n is the total number of numerical value in scores vector, S_iFor I-th failure fractional data in scores vector,For the average of numerical value in scores vector；

<2>Correlation matrix is solved according to failure covariance matrix Σ

Wherein, r_ijFor correlation matrixIn the i-th row, jth row numerical value, s_ijFor the i-th row, jth in covariance matrix The numerical value of row；

<3>Ffault matrix principal component is solved from failure covariance matrix Σ, obtains eigenvalue λ_iWith characteristic vector t_i, then its n Individual main component computational methods are：

Wherein, x is the failure fractional data vector of the i-th measuring point,For the principal component numerical value of the i-th measuring point.

<4>P accumulation contribution rate is more than 90% principal component feature before asking for, and reaches dimensionality reduction effect, wherein principal component yi Contribution rate be：The accumulation contribution rate of preceding p is as follows：Wherein p≤n, if If the accumulation contribution rate of this P feature chosen is more than 90%, it is achieved that failure fractional data S changes into P dimensions by n dimensions Dimensionality reduction purpose；

SVM fault graders are trained based on the fault data matrix J after dimensionality reduction, carry out fault diagnosis.

The beneficial effects of the invention are as follows：

(1) by the way of regression filtering and coordinate the normalized method in center in terms of processing equipment data, disappearing While except data dimension, the characteristics of greatly having highlighted fault data, it can be more beneficial to by the way of linear normalization than simple The speed and precision that lift scheme is trained in terms of data；

(2) device data period processing mode using plus Hamming time window method, and between two time windows The Duplication for having half can prevent some lofty signals from upsetting to be peeled away by isolated as far as possible, beneficial to disaggregated model pair below In the abundant study and training of fault progression process；

(3) avoided on solving the polytypic problems of SVM using 1VS 1 classification ballot form under one-to-many manner The unbalanced problem of training data ratio, while the defects of error accumulation under directed acyclic mode can be eliminated again, improve model The reliability of classification；

(4) to avoid occurring under 1VS 1 classification form poll identical the problem of can not drawing fault type, using double The pattern of layer diagnosis, regard the high failure collection of fault correlation degree as resultant fault, resultant fault will pass through preliminary classification model Preliminary differentiation after, then the careful identification through specific disaggregated model the problem of successfully avoiding that fault type can not be drawn, effectively carries The high accuracy of fault diagnosis system；

(5) incremental learning of disaggregated model laterally and longitudinally is realized, new fault type data can be added to failure in time and know Know in storehouse, extend the diagnostic area of disaggregated model, while old fault type is trained again again, strengthen disaggregated model Memory to old fault type, it is therefore prevented that the degeneration of disaggregated model recognition performance；

(6) more useful fault messages are obtained, for example, belong to the reliability of all kinds of failures, position occurs for failure, specially Maintenance guidance etc. is given by family, more more rich than the simple fault type mode information content that provides, and contributes to maintenance personal rationally next The arrangement and expansion of the troubleshooting effort of step.

Brief description of the drawings

Fig. 1 equipment fault diagnosis method flow charts

Fig. 2 SVMs diagnostic model structure charts

The one-to-one more classification ballot decision machine schematic diagrames of Fig. 3 SVM.

Recognition effect figure of Fig. 4 fault diagnosis systems to test data

The main measuring point tendency chart of feature of all kinds of failures of Fig. 5 #1 steam turbine equipments

Embodiment

The following detailed description of the specific implementation of the present invention, it is necessary to it is pointed out here that, implement to be only intended to this hair below Bright further explanation, it is impossible to be interpreted as limiting the scope of the invention, art skilled person is according to above-mentioned Some nonessential modifications and adaptations that the content of the invention is made to the present invention, still fall within protection scope of the present invention.

The present invention includes four main process, is preprocessing process, the fault diagnosis case base of device data respectively Building process, SVMs failure diagnostic process, fault message obtain and maintenance instruction course.One or two part therein Belong to device data processing category；Part III belongs to Fault Identification category；Part IV belongs to fault message and obtains category.

Equipment fault intelligent diagnosing method based on SVMs, as shown in Figure 1, the whole design of the present invention include with Under several processes：

Step 1 is the pretreatment operation of device history data, and main purpose is by process of data preprocessing that equipment is former The fault signature of barrier data shows as much as possible, and this process mainly includes three key steps.

The regression filtering processing 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, and substantially operating process is： After selected measuring point is obtained from PI databases, the relevant device enough time is read from power plant's database using ACS softwares History run status data；Then fault data G and floor data M is carried out to all data using certain screening rule to draw Point, if occur in data trend numerical value transfinite, fluctuate acutely, without fluctuate phenomena such as can be divided into fault data；Finally All historical failure datas are matched using expertise knowledge and equipment fault historical record, by the careful division of fault data Into different types of failure.

When device data of one device databases importing with n measuring point after nan data screenings has included m Carve, then can regard the vector of n dimension as in the data strip of each point of observation at j moment, be represented by：

ω(t_j)=[ω_j1,ω_j2,ω_j3,...ω_jn]^T

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

Step 1.1.2 establishes SVM and returns filtering model

First, form is subjected to equidistant compression processing for the nominal situation data M of [m × n].Specific method is：It is determined that After extracting data spacing distance d well, the data at all moment are asked for into Euclid norm can obtain the vector of a m dimension, so L state moment is extracted in the Euclid's vector tieed up afterwards using distance d as extraction distance from m, whereinThus can be Floor data is reduced into compressed data T；

Then, compressed data T is subjected to linear normalization processing, to eliminate each skimble-scamble shadow of measuring point data dimension Ring, obtain standard exercise data R, wherein normalization formula is as follows：

Wherein x is the numerical value of equipment measuring point, subscripting be " mark " for the form after normalization, subscripting is " original " For raw data form, subscripting is the maximum for the measuring point of " max ", and subscripting is the minimum for the measuring point of " min " Value.

Finally, standard exercise data R can be as the training data for establishing SVM regression models.SVM Regressions are will be multiple Measuring point parameter is mapped in the feature space of higher-dimension, is then constructed strongly-convex problem using target component as output, is passed through seizure Feature space Margin Vector (i.e. SVMs) constantly Tuning function system is gone to approach one and can solved the above problems simultaneously Full function expression formula comprising input parameter with output parameter complex nonlinear mapping relations again.To realize multi-measuring point parameter Output, the present invention will be used as output target ginseng successively using all measuring point parameters as these measuring point parameters while input parameter Several modes builds a series of multiple-input and multiple-output SVM regression models.

Nonlinear problem f (X)=W ψ (X)+b is exactly changed into Lagrange introduced below by SVM regression modelings principle The quadratic programming problem of the factor：

The Lagrange factor wherein optimizedL is the number of SVMs, The in store Lagrange factor α of SVM models^*, SVMs l, slack variable ξ, the important parameter such as error penalty factor.

N SVM regression model can be built using standard exercise data R, can be with real time to test data vector reference Historical data fits an assessed value vector.

Step 1.1.3 fault datas are melted into residual error data

The definition of residual error is instantaneous value and the difference of 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 includes two parts, first, healthy residual error bound and fault residual data are obtained, second, being residual error Data segment carries out windowing process.

First, after n SVM regression models structure is completed, the normal training data M shaped like [m × n] is sequentially input Recurrence assessment is all carried out in SVM regression models can obtain the residual matrix of one [m × n], and each measuring point of residual matrix is most Value can be used as healthy residual error bound T.

Then, according to same way as above, all kinds of fault data G are sequentially input into SVM regression models respectively, obtain event After the assessed value for hindering data, all kinds of fault residual data U are turned to.

Finally, load time window processing is carried out to all kinds of residual error data U respectively, reason is what each failure was necessary to Lead-time, the data at single time point can not comprehensively embody the characteristic attribute of failure.

The size of time window is defined as the most long cycle time of lead-time in all fault types of the equipment, thus may be used Ensure that the residual error data after the processing of elapsed time window can cover the faulty development time of institute, time window size determines that method is：

t_w=max [t₁, t₂..., t_i... t_n]

I ∈ [1, n]

(5)

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

The formal character of time window is Hamming window (Hamming).Hamming window (Hamming) retention fault can be sent out as far as possible The feature of mid-term is opened up, the unstable feature in failure late period morning is slackened, can possess periodic feature using fault data, beneficial to complete Whole expression fault signature attribute.With the 1/2 of the velocity magnitude time window of positioning size of time shift window, so front and rear two to connect Window just has 50% Duplication, it is ensured that the stationarity of similar fault data feature, avoids lofty data mode It is extracted.Here is expression formula of the Hamming window in time domain：

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

The specific method that fault data is handled through Hamming window is, after each adding window, the data after windowing process are averaged The global feature level of the fault signature of whole time window is represented, such fault residual data U changes into failure windowed data W, often The data volume m of individual measuring point is reduced toFailure windowed data form is as follows：

Wherein, w_ijThe numerical value arranged for the i-th row, jth, D (i) are i-th of fault data in window, and w (i) is i-th of number of faults According to window coefficient.

Step 1.2, residual error data eliminates dimension processing

According to healthy residual error data bound T, the failure windowed data W obtained by step 1.1.3 is subjected to center normalization Processing, eliminates each skimble-scamble problem of measuring point dimension.Shown in comprising the following steps that：

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

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

Wherein, s_pFor measuring point i failure fractional data, T_OnFor the measuring point i residual error data upper limit, ω is measuring point i failure Windowed data.

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

Wherein, s_pFor measuring point i failure fractional data, T_UnderFor the measuring point i residual error data upper limit, ω is measuring point i failure Windowed data.

(4) above-mentioned two steps operation is repeated, until completing the normalized of all measuring points.

So failure windowed data W will be normalized to failure fractional data S.If the advantages of so doing is exactly this moment hair If abnormal alarm occurs in generating apparatus state, the fraction of at least one measuring point of meeting is in the failure fractional data vector at this moment It is negative, conversely, whole measuring points will occur without negative fractional value.

Step 1.3, fault data PCA dimension-reduction treatment

The failure fractional data S that step 1.2 is obtained can be mapped to by principal component analysis (PCA) method of descent by orthogonal transformation The process of lower-dimensional subspace.The optimal correlation in view of data in mapping process, so it is transformed into the feature behind subspace Correlation bottoms out.

Comprising the following steps that for principal component analysis (PCA) is shown：

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

Wherein, s_ijFor the i-th row in covariance matrix, the numerical value of jth row, n is the total number of numerical value in scores vector, S_iFor I-th failure fractional data in scores vector,For the average of numerical value in scores vector.

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

Wherein, r_ijFor correlation matrixIn the i-th row, jth row numerical value, sij be covariance matrix in the i-th row, jth The numerical value of row.

(3) ffault matrix principal component is solved from failure covariance matrix Σ,

Mainly obtain eigenvalue λ_iWith characteristic vector t_i, then its n main component computational methods be：

Wherein, x is the failure fractional data vector of the i-th measuring point,For the principal component numerical value of the i-th measuring point.

(4) p accumulation contribution rate is more than 90% principal component feature before asking for, and reaches dimensionality reduction effect

Method is：Principal component y_iContribution rate be：

And the accumulation contribution rate of preceding p (p≤n) is as follows：

If if the accumulation contribution rate of this P feature chosen is more than 90%, it is achieved that failure fractional data S is turned by n dimensions It is melted into the dimensionality reduction purpose of P dimensions.

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

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

Step 2.1, fault signature is extracted

The purpose of this step is to determine the performance characteristic of failure, and a kind of failure is certain in equipment different from another kind of failure Showed on the intensity of anomaly of measuring point.The present invention determines the main feature measuring point of failure in a manner of the failure contribution degree of measuring point,

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

(2) the failure score matrix J of the i-th class is assumed_iSize isThe then failure scores vector form at each moment It is as follows：

J_ij=[j_ij1,j_ij2,...,j_ijt..., j_ijp] (15)

Wherein, j_ijtFor the i-th class failure score matrix jth moment, the failure fraction numerical value of t measuring points.

Each measuring point contribution degree and accumulation contribution degree are asked successively according to formula below for the scores vector at per moment.Choose Accumulate fault signature main measuring point of preceding K measuring point of the contribution degree more than 70% as this moment.

When all measuring points obtain the main measuring point of fault signature at corresponding moment, count each measuring point and be selected as fault signature master The probability of measuring point, measuring point of the probability more than 70% will be chosen to pick out as the i-th final main measuring point of class fault signature.

The formula of measuring point t failure contribution rates is as follows：

Wherein, j_ijtFor the i-th class failure score matrix jth moment, the failure fraction numerical value of t measuring points, G_ijtFor the event of the i-th class Hinder score matrix jth moment, the failure contribution numerical value of t measuring points.

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

Wherein, j_ijtFor the i-th class failure score matrix jth moment, the failure fraction numerical value of t measuring points, G_ijtFor first K survey Point cumulative failure contribution numerical value.

(3) above-mentioned two steps operation is repeated, until the fault signature for completing all fault types is decided.

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

Step 2.2, fault correlation degree is analyzed

This step determines the degree of association rank of all kinds of failures between any two using fault signature similarity as discrimination standard； If if there are k kind fault types, then need to match k (k+1)/2 group fault type pair.

Degree of association rank is divided into 2 grades：It is rudimentary and advanced.Specific matching process is as follows：

First, when fault type i and fault type j carry out asking for fault correlation spending, carried from fault signature data frame Π Take the main measuring point Π of respective fault signature_iAnd Π_j

Then, common factor H is asked for above, when common factor H measuring point number accounts for Π_iAnd Π_jNumber 50% when, be considered as therefore The degree of association rank for hindering type i and fault type j is height, otherwise to be low.

Finally, the degree of association rank of k (k+1)/2 group fault type pair is all sought out, is stored in fault type pass It is standby in connection degree matrix K.

Step 2.3, fault type label is added

The low fault type of the degree of association is identified as single failure type by this step, and the higher fault type of the degree of association can close And a resultant fault type is identified as, it will continue the Fault Identification of particular type in the Fault Identification of next stage.

For existing k kinds fault type, start single failure type conducts all successively first by integer 1 Its fault type label, then remaining resultant fault type is identified successively again, identification means are resultant fault type Sub- fault type be all identified as same integer type label, be stored 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, is not in label 0 in training data.

Training and failure diagnostic process of the step 3 for svm classifier model, belong to the fault type recognition portion of fault diagnosis Point, mainly include four steps：

Step 3.1, svm classifier model is trained

This part carries out the training of disaggregated model, including the training of preliminary classification model and specific point using fault data The training of class model；

SVM principle of classification summary：SVM is initially that to seek optimum segmentation face the point in training set can be made remote as much as possible From two classification problems of the plane.

The double optimization problem can be expressed as：

Dual problem is translated into again, and form is as follows：

Most classification function is positioned as at last：

Wherein, x be independent variable measuring point data vector, y be target measuring point data vector, segmentations of the w between two classes The distance value in face, b are constant vector, a_iFor SVMs.

The present invention establishes the classification ballot decision-making mechanisms of 1vs 1 (as shown in Figure 3) on the basis of svm classifier principle.With Solve the polytypic problems of SVM.

1vs 1 classification ballot decision-making mechanism construction method be：K kind fault types match two-by-two, by structure k (k+1)/2 Individual supporting vector disaggregated model.For example there are tetra- kinds of fault types of A, B, C, D, then the present invention will establish [AB], [AC], [AD], [BC], [BD], [CD] amount to six svm classifier models.In four kinds of fault types for assuming the above simultaneously, A, B, C are single Fault type, and if D includes sub- fault type E, F, G for resultant fault type, in mother stock class model [AD], [BD], [CD] D roots under can derive [EF], [FG], [EG] three sub- disaggregated models.Fault data matrix J in step 1.3 is as training Data trained six mother stock class models with four major class failure modes first, then take again three small according to svm classifier principle Fault data three sub- disaggregated models of training of class, the situation of final svm classifier model are 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 the svm classifier model of table 1

The bandwidth σ parameters of step 3.2 optimization gauss kernel function

SVM is to realize that the approach that an optimum segmentation face can be searched out under Arbitrary Dimensions is to seek inner product by kernel function Method feature space be mapped to more high-dimensional reciprocal of duty cycle look for optimum segmentation face (i.e. SVMs), and Gaussian kernel letter Several bandwidth σ plays vital effect with error penalty factor for SVM classifying quality.Therefore the present invention is using friendship Fork checking collection carrys out the bandwidth σ parameters and penalty factor of optimization gauss kernel function.

The present invention runs cross validation collection method as follows：

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

Then, for disaggregated model i, 3 equal portions will be divided equally at random per a kind of failure training data, every time will be therein 1 part is used as test set, and remaining 2 parts are trained as training data, right using grid data service within the limits prescribed SVM bandwidth parameter σ and penalty factor carries out optimizing operation, and a pair of parameter values of discrimination highest will be used as optimum bandwidth σ Parameter and penalty factor；

Finally, all mother and sons' disaggregated models have been optimized in sequence, have been saved in relevant parameter vector P.

Step 3.3, svm classifier model carries out fault diagnosis

It present invention can be suitably applied to carry out fault diagnosis for online real time data, the idiographic flow of fault diagnosis is as follows：

(1) real time data rt enters in the SVM regression models that process 1 is established, and carries out regression filtering processing, normalization successively Processing, PCA dimensionality reduction data processings are converted into real-time fractional data J；

(2) first, if real-time fractional data J turns to failure fractional data J if negative value being present, if in the absence of if negative value It is judged to trouble-free nominal situation；Then, after failure fractional data J sequentially enters the parameter optimization built in step 3.1 Mother stock class model collection is classified；Next carries out ballot statistics, and whether the type for judging tentatively to identify is resultant fault type, if It is that the subclassification Models Sets that failure fractional data J also needs to sequentially enter in step 3.2 after parameter optimization corresponding to structure are thrown Ticket counts, and identifies its specific fault type, if it is not, the type tentatively identified is exactly specific fault type, idiographic flow is for example attached Shown in Fig. 2.

The method of the classification ballot decision-making mechanisms of 1vs 1 is sequentially entered whole female (sons) point in real time fail fractional data J After class model, the poll of every kind of fault type is counted, the fault type that the most fault type of number of votes obtained as finally determines, if When the fault type for most polls occur is multiple, then it is determined as new fault type, label is set to 0.

Step 3.4, the incremental learning of svm classifier model

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

During real-time fault diagnosis, if occurring if being judged as fault type label 0, that is, there is new fault type, The real time data can be preserved in the data frame that a label is new, data frame size is [1000 × p]；If it is judged as If fault type label is non-zero, that is, there is old fault type, the real time data can be preserved corresponding to label be that (n is old_n Fault type label) data frame in, the size of data frame is [1000 × p]；

(2) if after new data frame is filled with data, new fault type data are added into original failure training data In, start to carry out the incremental learning of svm classifier according to the method for step 3.1, while new data frame can be cleared；If After old_n data frame is filled with data, old acquaintance is hindered into categorical data and added in original failure training data, is started according to step 3.1 method carries out the incremental learning of svm classifier, while old_n data frame can be cleared；

(3) finally, the parameter for the svm classifier model completed according to the method for step 3.2 to renewal is optimized, updated Parameter vector P.

Process 4 is that fault message obtains and overhauled instruction course

For this process with the fault diagnosis category that process 3 is all real time data, the fault message for belonging to real time fail data is whole Part is closed, this process includes 3 key steps：

Step 4.1, fault type and recognition credibility are obtained

In the present invention, after the identification of each svm classifier model under the forms of 1vs 1 of the real time data rt by process 3, The confidence level for belonging to each fault type can be obtained, the fault type finally determined is to belong to confidence level highest fault type.

The recognition credibility algorithm for how obtaining real time data rt fault diagnosises is detailed below：

(1) real time data rt can not only obtain fault type when carrying out type identification using svm classifier model, and The decision value of a signed can be obtained, decision value can regard vertical range of the data apart from optimum segmentation face as, very The numerical symbol of the obvious decision value determines identification types, and the size of the decision value absolute value represents it and possesses category category The power of property.Confidence level ep is identified to real time data according to below equation and asks for by the present invention：

MaxD therein represents the decision value of maximum absolute value in the training data of the fault type, and minD represents the failure The minimum decision value of absolute value in the training data of type, rtD represent the categorised decision value of real time data.

Understand that the training data recognition credibility ep of decision value maximum absolute value is 100% according to formula (22), and decision-making It is 60% to be worth the minimum training data recognition credibility ep of absolute value.Therefore, a real time data rt can obtain the credible of two classes Degree.

Real time data rt passes through all svm classifier models successively, can obtain the confidence level of all fault types, specific to calculate Shown in method equation below：

Wherein, n is fault type, and N is the number for the failure training data structure svm classifier model that type is n, and epi is Confidence score of the real time data to fault type i.

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

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

After the fault type that step 4.1 determines real time data rt, by real time data rt be converted into failure fractional data J to Amount form, wherein vectorial J mileages value is the doubtful position M to break down for negative measuring point.

If real time data rt fault type is n type failure, corresponding class is consulted from fault signature data frame Π Fault signature Π _ n corresponding to type n, and fault signature matching degree will be solved according to following formula：

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

Step 4.3, maintenance direction is given according to expertise library inquiry and for maintainer

The present invention will be broken down on this equipment among the guiding opinion importing software repaired, in case adjusting at any time With.

When equipment, which is made a definite diagnosis, certain known fault occurs, equipment fault intelligent diagnosis system starts to read in expert knowledge library Maintenance suggestion under corresponding failure catalogue, maintainer is showed successively according to the weight size of suggestion, while according to step 4.2 The size for the matchDG values obtained calculates maintenance time in case maintainer's reasonable arrangement maintenance time.

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

The present invention is to #1 steam turbine local equipments based on SVMs equipment fault intelligent diagnosing method key step It is as follows：

First, data prediction operation is carried out to the historical data of #1 steam turbine equipments

Selected #1 steam turbine equipments are one to have a multiple faults type multi-measuring point exemplary apparatus, and this example is from power plant's PI data Storehouse obtains 15 selected related measuring points, and they are main oil pump outlet oil pressure (kpa), bearing metal temperature (DEG C), steamer respectively Machine rotating speed (r/s), returning-oil temperature (DEG C), bearing amplitude (mm/s), atmospheric pressure (kpa), axial displacement (mm/s), thrust bearing shoe valve temperature Spend (DEG C), main steam temperature (DEG C), reheat steam temperature (DEG C), thrust bearing shoe valve metal temperature (DEG C), steam flow (t/h), steam discharge Temperature (DEG C).

#1 steam turbine equipments are read from December 31 1 day to 2013 January in 2011 from power plant's database using ACS softwares Number 3 years whole historical datas, data volume is about 1,300,000；We are by two-and-a-half years on May 31,1 day to 2013 January in 2011 Data are used to build svm classifier model as training data, and the data using on June 1st, 2013 to December 31 are used as test The effect of data detection svm classifier Symbolic fault diagnosis.

We need to isolate nominal situation data from #1 steam turbine equipment training datas using existing screening rule With the class data of abnormal data two.Nominal situation data represent the normal operating condition of #1 steam turbine equipments, and this partial data is just Can be as the training data T of structure SVM regression models_Return.Nominal situation data represent #1 steam turbine equipments and transported in this period of history All fault types occurred during row, it is all using expertise knowledge and equipment fault historical record matching Historical failure data, different types of failure is divided into by whole abnormal datas are careful, by fault data types of nuclear to after, two Four kinds of typical faults of steam turbine equipment are contained in the abnormal data in year half：Steam Turbine Over-speed Accident, Steam Turbine Vibration is excessive, runs Middle blade damage or fracture and the water entering of steam turbine.

Training data T_ReturnData be mainly used for build regression model：It is using linear normalization method that training data is each Individual measuring point dimension is zoomed between [0,1], then training data T is compressed in a manner of step 1.2_ReturnTo 3000 or so；Finally, press The SVM regression models of 10 multiple input single outputs are built using 3000 training datas after compression according to SVM Regressions.

Training data T_ReturnCan be training data output size phase after being sequentially inputted to the above-mentioned SVM regression models of 10 Same healthy residual matrix；Simultaneously by all kinds of fault data T_FailureAfter being sequentially inputted to the SVM regression models of 10 respectively, also may be used Fault data is changed into fault residual data.Simultaneously to excavate the characteristic attribute of fault residual data comprehensively, according to formula (5), the mode of formula (6) carries out the failure windowed data of hamming window processing to it, due to four kinds of Steam Turbine Over-speed Accidents, steamers Machine vibration is excessive, in operation blade damage or fracture and water entering of steam turbine inaction interval be respectively 48.1 seconds, 13.3 seconds, 35.6 seconds, 20.5 seconds, therefore the time window size of hamming windows is uniformly defined as 50 seconds.

#1 steam turbine equipment failure windowed datas carry out center normalized according still further to the way of formula (7), (8), disappear The dimension impact of measuring point unless each, and be that the state at each moment makes a call to the fraction of an identical standard situation, after PCA dimensionality reductions The dimensions of #1 steam turbine equipment failure fractional datas be reduced to 9 from 13, screened out and other feature correlations be high and principal component tribute 4 low features of degree of offering：Reheat steam temperature (DEG C), thrust bearing shoe valve metal temperature (DEG C), steam flow (t/h), exhaust temperature (℃)。

2nd, #1 steam turbine equipment fault diagnosis case bases are built

In order to preferably embody the feature of all kinds of fault datas of #1 steam turbine equipments, this part is still taken off from data plane Show the essential attribute and the degree of association each other of each failure, therefore still belong to the pretreatment stage of device data.

First, fault signature is extracted from #1 steam turbine equipment failure fractional datas.According to failure contribution degree and failure The calculation of contribution rate is accumulated, can preferably extract the substantive characteristics of every kind of fault type, following table is with Steam Turbine Over-speed Accident Failure fractional data exemplified by, the failure contribution degree of exposition moment each feature measuring point, it can be seen that in Steam Turbine Over-speed Accident In this failure, feature measuring point of the cumulative failure contribution rate more than 70% is primary outlet oil pressure, bearing metal temperature, steam turbine turn Fast three main feature measuring points.

The Steam Turbine Over-speed Accident failure fractional data measuring point failure contribution degree of table 2

In the way of more than, the spy of four kinds of major failures of #1 steam turbine equipments can be extracted from all kinds of failures successively Sign：

(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) blade damage or fracture in running

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 to analyze the height of all kinds of fault correlation degree, according to the judgement of fault correlation degree in step 2.2 Rule can relatively easily judge similarity between blade damage or fracture defect and water entering of steam turbine failure in operation Up to more than 50%, this two kinds of failures should first be merged into a kind of resultant fault to improve the accuracy of fault diagnosis.

Analyzed more than, in four kinds of fault types of #1 steam turbines, Steam Turbine Over-speed Accident, Steam Turbine Vibration are excessive to belong to single One fault type, blade damage or fracture defect merge into a kind of resultant fault type with water entering of steam turbine failure in operation.

Therefore, in this example, we are Steam Turbine Over-speed Accident, Steam Turbine Vibration is excessive adds fault type label 1,2 respectively, Resultant fault type addition label 3, and blade damage or fracture and steam turbine in the sub- failure operation under resultant fault type Label 4,5 is added in water slug.

3rd, the disaggregated model and fault diagnosis of #1 steam turbines are trained

To solve four kinds of failures of svm classifier #1 steam turbines, according to the 1vs1 thinkings of step 3.1, we are according to svm classifier Principle, four svm classifier models are constructed to carry out the identification of four kinds of fault types.

(1) case mold A is used to identify that Steam Turbine Over-speed Accident failure crosses major break down with Steam Turbine Vibration, adds label 1,2 respectively；

(2) case mold B is used to identify Steam Turbine Over-speed Accident failure and resultant fault, adds label 1,2 respectively；

(3) case mold C is used to identify that major break down and resultant fault are crossed in Steam Turbine Vibration, adds label 1,2 respectively；

(4) submodel D is used to identify blade damage or fracture defect and water entering of steam turbine failure in operation, adds respectively Label 1,2.

All disaggregated models above are optimized in SVM models according to step 3.2 using the method for cross validation collection Gaussian kernel function bandwidth σ parameters and error penalty factor, and parameter optimization method carries out optimizing using grid data service, wherein Gaussian kernel function bandwidth σ scopes are [0.1,10], and step-size in search 0.1, error penalty factor scope is [1,100], search step A length of 1.Training recognition effect after optimization is as follows：

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

Recognition effect of 3 four disaggregated models of table under optimal parameter

The data on June 1st, 2013 to December 31 are carried out fault diagnosis by this example.

In order to intactly capture all fault types of #1 steam turbine equipments, set device intelligent fault diagnosis in this example System is when i is 25 seconds integral multiples at runtime, reads the i-th -49 to i of #1 steam turbine equipments and amounts to 50 test datas and carries out Fault diagnosis.This 50 test datas are subjected to data prediction operation first, i.e., SVM regression filterings, center in process 1 are returned Four kinds of one change, load time window, Feature Dimension Reduction preprocess methods, are eventually converted into a real time fail fractional data J；Then, Examine in real time fail fractional data J and whether include negative value, further fault diagnosis is needed if having for abnormal data, if nothing Then fault diagnosis need not be carried out for nominal situation data.After abnormal data is judged to, real time fail fractional data J sequentially enters mother Model A, B, C, if if being identified as single failure type, fault type recognition is completed, if it is identified as resultant fault type If, submodel D need to be gone successively to and continue to identify, obtain final fault type.If there are most poll identical situations, Then fault type is sentenced and makees new fault type, type 0.

Citing, enter in this section of real time data in 25 minutes October 9 day 11 point of 35 minutes to 2013 October 9 day 10 point in 2013 After equipment fault intelligent diagnosis system, change 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 negative value phenomenon be present, should sentence and make abnormal data, sequentially enter three case mold A, B, C result is followed successively by 1,2,2.Resultant fault type obtains two tickets, and it is 2 to go successively to submodel D judged results, final identification knot Fruit is water entering of steam turbine failure.

It is the diagnosis knowledge to the #1 steam turbine equipment fault datas on June 1st, 2013 to December 31 shown in accompanying drawing 4 Sorrow of separation condition, it can be seen that：

(1) recognition accuracy to Steam Turbine Over-speed Accident failure is 85%；

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

(3) it is 95% to the recognition accuracy of blade damage in operation or fracture defect；

(4) recognition accuracy to water entering of steam turbine failure is 75%.

It is satisfactory to the recognition accuracy whole structure of four kinds of failures.

In this example, four data frames are had, the big I of a new data frame loads 1000 real time fail fraction numbers According to；The big I of three old_n data frames loads 1000 real time fail fractional datas.After if data frame is fully loaded with data, according to The method of step 3.4 realizes that the training again with old fault type is expanded in the training to #1 steam turbine fault types so that structure Four models will not with steam turbine equipment run time develop and degenerate, possess good adaptive ability.

4th, real time fail acquisition of information and maintenance instruction course

For this process with the fault diagnosis category that process three is all real time data, the fault message for belonging to real time fail data is whole Close part.

In this example, it is each under the forms of 1vs 1 of the #1 steam turbine equipments real time fail fractional data J Jing Guo process 3 After the identification of svm classifier model, fault type is not only can obtain, and real time fail can be also obtained according to the method for step 4.1 Fractional data J belongs to the confidence level of every class fault type.

According to the recognition credibility algorithm of fault diagnosis, #1 steam turbine equipment real-time testing data i belongs to every kind of failure classes The confidence level of type is：Steam Turbine Over-speed Accident 5.2%, Steam Turbine Vibration are excessive 1.3%, blade damage or fracture 20.3% in operation, The water entering of steam turbine 73.2%.The fault type that real time fail fractional data J is finally determined is that confidence value is maximum, exports failure Type is the water entering of steam turbine.

This example can obtain #1 steam turbine failures location information and with corresponding failure characteristic matching degree, specific method is such as Under：

It is the rule that negative measuring point is the doubtful position broken down according to real time fail fractional data vector J mileages value Then, the main measuring point of water entering of steam turbine fault signature is main steam temperature, axial displacement, bearing metal in fault signature data frame Π Temperature, bearing return oil temperature.Real time fail fractional data vector J mileage values are that negative measuring point is axial displacement, bearing metal temperature Degree, bearing return oil temperature 75%.

When equipment, which makes a definite diagnosis #1 steam turbine equipments, water entering of steam turbine failure occurs, equipment fault intelligent diagnosis system starts The maintenance suggestion under water entering of steam turbine defect list in expert knowledge library is read, maintenance people is given according to the weight size of suggestion Following two of member repair suggestion：

1st, when confirming that water slug occurs for steam turbine, vacuum emergency shutdown is destroyed immediately, cuts off relevant vapour, water source, strengthen Master, reheating steam pipe road, body extraction line, axle envelope vapour main pipe etc. are about the hydrophobic of system；

2nd, unit normal operation or open, shut down and varying load process in, main, reheat steam temperature drastically declined 50 in 10 minutes DEG C, vacuum should not be destroyed and beat gate stop-start.

Although for illustrative purposes, it has been described that illustrative embodiments of the invention, those skilled in the art Member it will be understood that, can be in form and details in the case of the scope and spirit for not departing from invention disclosed in appended claims The upper change for carrying out various modifications, addition and replacement etc., and all these changes should all belong to appended claims of the present invention Protection domain, and each step in each department of claimed product and method, can be in any combination Form is combined.Therefore, to disclosed in this invention embodiment description be not intended to limit the scope of the present invention, But for describing the present invention.Correspondingly, the scope of the present invention is not limited by embodiment of above, but by claim or Its equivalent is defined.

Claims (2)

1. a kind of equipment fault intelligent diagnosing method based on SVMs, it is characterised in that comprise the following steps successively：

(1) pretreatment operation is carried out for device data；

(2) fault diagnosis case base is built；

(3) fault diagnosis is carried out to SVMs；

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

Wherein described step (1) comprises the following steps：

Step 1.1：Regression filtering processing is carried out to device data, including：

Step 1.1.1：The history data of certain time is read in slave unit database, wherein device data includes certain number The fault state data of amount and nominal situation data；

Step 1.1.2：The recurrence filtering model of SVMs is established using nominal situation data, the step 1.1.2 is specific Step is：Form is subjected to equidistant compression processing for the nominal situation data M of [m × n], data spacing distance is extracted determining After d, the data at all moment are asked for into Euclid norm and obtain the vector of a m dimension, are tieed up using distance d as distance is extracted from m Euclid's vector in extract l state moment, whereinFloor data is reduced into compressed data T；

Then, compressed data T is subjected to linear normalization processing, eliminates 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 " for the form after normalization, subscripting is " original " for original Beginning data mode, subscripting are the maximum for the measuring point of " max ", and subscripting is the minimum value for the measuring point of " min "；

Finally, using standard exercise data R as the training data for establishing SVM regression models, input is used as using all measuring point parameters Parameter, and mode of the measuring point parameter successively as output target component returns to build a series of multiple-input and multiple-output SVM Model, one assessed value vector is fitted with reference to historical data to test data vector；

Step 1.1.3：All device history service datas handle to obtain residual error data through regression filtering；

Step 1.2：Dimension is eliminated to residual error data, using residual error data with training the ratio of residual error bound to be turned into fraction shape Formula, the dimension impact between each equipment measuring point is eliminated, failure measuring point feature is highlighted, obtains failure fractional data, The step 1.2 concretely comprises the following steps:

According to healthy residual error data bound T, center normalized is carried out to failure windowed data W, eliminates each measuring point amount Guiding principle, it is specially：

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

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

Wherein, s_pFor measuring point i failure fractional data, T_OnFor the measuring point i residual error data upper limit, ω is measuring point i failure adding window number According to；

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

Wherein, s_pFor measuring point i failure fractional data, T_OnFor the measuring point i residual error data upper limit, ω is measuring point i failure adding window number According to；

<4>Repeat above-mentioned<2>-<3>Two steps operate, until completing the normalized of all measuring points；

Step 1.3：PCA dimension-reduction treatment is carried out to fault data, so as to prevent the hair of the overfitting during Nonlinear Modeling Raw and small weight measuring point upsets normal mapping relations, and Feature Dimension Reduction processing, the step are carried out to equipment fractional data 1.3 concretely comprise the following steps:

Obtained failure fractional data S is mapped to lower-dimensional subspace by orthogonal transformation using PCA method of descents, is transformed into sub- sky Between after feature correlation bottom out, be specially：

<1>Failure fractional data S is solved into failure covariance matrix Σ：

<mrow> <mi>&amp;Sigma;</mi> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mi>n</mi> <mo>-</mo> <mn>1</mn> </mrow> </mfrac> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <mrow> <mo>(</mo> <msub> <mi>S</mi> <mi>i</mi> </msub> <mo>-</mo> <mover> <mi>S</mi> <mo>&amp;OverBar;</mo> </mover> <mo>)</mo> </mrow> <msup> <mrow> <mo>(</mo> <msub> <mi>S</mi> <mi>i</mi> </msub> <mo>-</mo> <mover> <mi>S</mi> <mo>&amp;OverBar;</mo> </mover> <mo>)</mo> </mrow> <mo>*</mo> </msup> <mo>=</mo> <mrow> <mo>(</mo> <msub> <mi>s</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <mo>)</mo> </mrow> </mrow>

Wherein, s_ijFor the i-th row in covariance matrix, the numerical value of jth row, n is the total number of numerical value in scores vector, S_iFor fraction I-th failure fractional data in vector,For the average of numerical value in scores vector；

<2>Correlation matrix R is solved according to failure covariance matrix ∑：

<mrow> <mover> <mi>R</mi> <mo>^</mo> </mover> <mo>=</mo> <mrow> <mo>(</mo> <msub> <mi>r</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <mo>)</mo> </mrow> <mo>,</mo> <msub> <mi>r</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <mo>=</mo> <mfrac> <msub> <mi>s</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <mrow> <msqrt> <msub> <mi>s</mi> <mrow> <mi>i</mi> <mi>i</mi> </mrow> </msub> </msqrt> <msqrt> <msub> <mi>s</mi> <mrow> <mi>j</mi> <mi>j</mi> </mrow> </msub> </msqrt> </mrow> </mfrac> </mrow>

Wherein, r_ijFor correlation matrixIn the i-th row, jth row numerical value, s_ijArranged for the i-th row, jth in covariance matrix Numerical value；

<3>Ffault matrix principal component is solved from failure covariance matrix ∑, obtains eigenvalue λ_iWith characteristic vector t_i, then its n master The composition computational methods are wanted to be：

<mrow> <mover> <msub> <mi>y</mi> <mi>j</mi> </msub> <mo>^</mo> </mover> <mo>=</mo> <mover> <msubsup> <mi>t</mi> <mi>i</mi> <mo>*</mo> </msubsup> <mo>^</mo> </mover> <mi>x</mi> <mo>,</mo> <mi>i</mi> <mo>=</mo> <mn>1</mn> <mo>,</mo> <mn>2</mn> <mo>,</mo> <mo>...</mo> <mo>,</mo> <mi>n</mi> </mrow>

Wherein, x is the failure fractional data vector of the i-th measuring point,For the principal component numerical value of the i-th measuring point；

<4>P accumulation contribution rate is more than 90% principal component feature before asking for, and reaches dimensionality reduction effect, wherein principal component y_iContribution Rate is：The accumulation contribution rate of preceding p is as follows：Wherein p≤n, if this chosen If the accumulation contribution rate of P feature is more than 90%, it is achieved that failure fractional data S is changed into the dimensionality reduction mesh of P dimensions by n dimensions 's；

Wherein, the step (2) comprises the following steps：

Step 2.1：Fault signature is extracted, the main feature measuring point of failure is determined in a manner of the failure contribution degree of measuring point；

Step 2.2：The degree of association between all kinds of failures is analyzed, using fault signature similarity as discrimination standard, obtains all kinds of events The degree of association series of barrier：

Step 2.3：Fault type label is added for all kinds of failures, the low fault type of the degree of association is identified as single failure type, The higher fault type of the degree of association, which can merge, is identified as a resultant fault type, wherein event of the resultant fault type in next stage Continue the Fault Identification of particular type in barrier identification；

The step (3) comprises the following steps：

Step 3.1：The disaggregated model of Training Support Vector Machines, the training of disaggregated model, including primary are carried out using fault data The structure of the structure of disaggregated model and specific disaggregated model；

Step 3.2：Using the bandwidth σ parameters and error penalty factor of cross validation collection optimization gauss kernel function, by training data Divide equally 3 equal portions at random, every time using 1 part therein as test set, remaining 2 parts are trained as training data, discrimination One group of parameter value of highest will be used as optimum bandwidth σ parameter factors and error penalty factor；

Step 3.3：Using svm classifier Model Identification fault type, primary identification model is used to judge whether device data is single Fault type, specific identification model are used for the specific fault type for judging resultant fault data and output fault message；

Step 3.4：The incremental learning of svm classifier model, carry out the training of new fault type and the training again of old fault type.

2. the equipment fault intelligent diagnosing method based on SVMs as claimed in claim 1, it is characterised in that：The step Suddenly (4) comprise the following steps：

Step 4.1, fault type and recognition credibility are obtained；

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

Step 4.3, maintenance direction is given according to expertise library inquiry and for maintainer.