CN105320126A - Secondary equipment hidden danger excavation method based on big data technology - Google Patents

Abstract

The purpose of the invention is to provide a secondary equipment hidden danger excavation method based on a big data technology. The method comprises two processes of model training and abnormal sample identification which are based on secondary equipment history data, wherein the model training process comprises the steps of obtaining a characteristic matrix of secondary equipment monitoring data, analyzing main components of secondary equipment sample data and calculating reliable statistics probability distribution; and the abnormal sample identification process is based on a sample state evaluation model, monitored samples of one device in a time interval are calculated and evaluated, and related quantified indexes are given. By adopting the method, the equipment having hidden dangers, which is characterized in that state monitoring data of the equipment is seemed to be normal, but the equipment is already in a slow degrading stage, can be early discovered. The performance stability and the abnormity degree of the equipment are evaluated in a quantified manner, the defect that the health state of the equipment cannot be accurately understood in the current secondary equipment maintenance process is overcome, maintenance personnel are enabled to early discover the hidden dangers before the failure of the secondary equipment, the range of equipment maintenance is shortened, and the maintenance cost is reduced.

Description

A kind of secondary device hidden danger method for digging based on large data technique

Technical field

The present invention relates to electric power system data disposal route, specifically propose for a kind of secondary device hidden danger method for digging based on large data technique of electric system.

Background technology

Electrical equipment is different according to function, can be divided into primary equipment and secondary device.Electrical secondary equipment mainly comprises relay protection, aut.eq., failure wave-recording, on the spot monitoring and telemechanical.The operation of their normal reliable is the basic demand ensureing the stabilization of power grids and power equipment safety.Along with microcomputer is in the widespread use of relay protection and aut.eq., the reliability of relay protection, the dirigibility of fixed value adjusting improve greatly.

Traditional repair method of the secondary device of electric system, generally based on the time, namely the actual state of tube apparatus how, does not expire required.Its lacks comprehensive analysis to equipment, and the complicate statistics method fallen behind due to many dependence means carrys out coaching device maintenance, is not often that maintenance is excessive, overhauls deficiency exactly.And regular inspection process of the test is complicated, inconsiderately during test other switch malfunction may be caused.Within the time between overhauls(TBO), protection exception can not Timeliness coverage., having there is the intelligent alarm relying on supervisory system and Microcomputer Protection in the current development along with technology, to be given the alarm signal, then site operation personnel's alerting signal, and come according to on-the-spot experience by the control system of routine.The subject matter that this pattern exists at that time was, when reporting to the police appears in equipment or be some not too important alarms, or was the alarm of equipment grievous injury.And being in the stage of " normally " on the Condition Monitoring Data surface of equipment, equipment has been in deteriorated process slowly, is in hidden danger state that is normal and fault.

Therefore, be badly in need of a kind of method that the hidden danger state of equipment is realized, the hidden danger state of Timeliness coverage secondary device, or identify the trend that develops to malfunction of secondary device, thus provide valuable information for the repair based on condition of component of secondary device.

Summary of the invention

The object of the present invention is to provide a kind of hidden danger method for digging of the secondary device Condition Monitoring Data based on large data technique, utilize data digging method, from the magnanimity Monitoring Data that the secondary device of model of the same race accumulates in history, find the relation between equipment state and Monitoring Data, the hidden danger state of Timeliness coverage secondary device, or identify the trend that develops to malfunction of secondary device, thus provide valuable information for the repair based on condition of component of secondary device.

Object of the present invention realizes by following technical measures:

Based on a secondary device hidden danger method for digging for large data technique, comprise two processes: based on model training and the exceptional sample identification of secondary device historical data.

Wherein, first process comprise obtain secondary device Monitoring Data eigenmatrix, the principal component analysis (PCA) of secondary device sample data, reliable statistics probability distribution, the model of output is dimensionality reduction transition matrix, mapping matrix and sample state evaluation model.The identification process of these model support exceptional samples.

The dimensionality reduction conversion of second process, sample map, sample identification depends on dimensionality reduction transition matrix in first process, mapping matrix and sample state evaluation model respectively.The exceptional sample identifying index (comprising intensity of anomaly index and degree of stability index) of final output, as the analysis and evaluation result to Monitoring Data in single device a period of time.

1) described as follows based on the particular content of the model training step of secondary device historical data:

1.1) the eigenmatrix detailed process obtaining secondary device Monitoring Data is as follows:

The relevant information of the secondary devices such as merge cells in transformer station, intelligent terminal, protection supervisory equipment, straight-flow system, the network equipment is related generally to according to professional secondary device; and the Testing index of every class specialty is different, as information off-line, online information, source book quantity of state, ruuning situation quantity of state, checking experiment situation quantity of state etc.Also need to carry out pre-service and standardization to the different raw information of all kinds of secondary device, forming device feature samples matrix at construction phase.When structural attitude sample matrix, by the distinct device of same model, mixing is in the matrix of sample data, and the matrix of described sample data is shown below:

Rawdata _j＝[IDTf _j,1f _j,2...f _j,m]

(j＝1,2,…,N)(1)

In formula (2), Rawdata _jrepresent the state-detection sample data of certain secondary device of jth bar (being numbered ID) at T sometime, such secondary device has m characteristic index, forms monitoring feature data successively.The eigenmatrix of monitoring feature data is wherein:

Feature _j＝[f _j,1f _j,2...f _j,m]

(j＝1,2,…,N)(3)

Equally, j is sample index, and m represents m characteristic index, and N represents sample number;

1.2) secondary device sample data principal component analysis (PCA) detailed process is as follows:

Factor data excavates and often relates to vector, matrix operation, and such calculates the increase along with dimension, calculated amount exponentially doubly increases, i.e. so-called " dimension calamity " phenomenon, and a lot of information of matrix of the original sample data of structure is repetition and redundancy, is also unfavorable for the visual of Result.Secondary device sample data principal component analysis (PCA) target is reducing data set dimension while, retains the information of former data set as much as possible.

Suppose a data set X={x _m, wherein m=1,2 ..., M, and each x _mit is the row vector of a N dimension.Principal component analysis (PCA) is mapped to by X in a K dimension space, the wherein dimension <N of K, makes the data remained can maximize the information representing raw data set simultaneously.

By described step 1.1) in the eigenmatrix [f of Monitoring Data of sample data _{j, 1}f _{j, 2}... f _j,m], after principal component analysis (PCA), eigenmatrix FEATURES_PCA after generation dimensionality reduction _j=[f _{j, 1}f _{j, 2}... f _j,k], namely contain former data set overwhelming majority quantity of information by k major component, preserve dimensionality reduction transition matrix Q simultaneously _t.

1.3) reliable statistics probability distribution concrete steps are calculated as follows:

Generate eigenmatrix FEATURES_PCA after dimensionality reduction _j=[f _{j, 1}f _{j, 2}... f _j,k] after, need statistical probability distribution.Center and the shape of distribution are represented by average and covariance respectively, are foundations during subsequent calculations distance.But because average and covariance are all very easily by the impact of abnormity point, therefore need a kind of algorithm can getting rid of abnormity point, reliable estimated probability distribution, algorithm steps is as follows:

Initialization: randomly draw H sample from dimensionality reduction eigenmatrix FEATURES_PCA, wherein N/2≤H≤3N/4, calculates its sample average T1 and covariance matrix S1;

Sample departs from centre distance and calculates, and computing formula is as follows here:

$\begin{array}{c}{d}_j=\sqrt{(FEATURES\_{\mathrm{PCA}}_j-T_1)S_1^{-1}{(FEATURES\_{\mathrm{PCA}}_j-T_1)}^T}\\ (j=1,2,...,N)\end{array}---\left(4\right)$

Wherein T1 is sample average and S1 is covariance matrix.

Iteration optimization: select H the sample that corresponding mahalanobis distance is minimum from dimensionality reduction eigenmatrix, its sample average T2 of iterative computation and covariance matrix S2, when meeting det (S2)=det (S1) or det (S2)=0, expect the reliable estimation of T and variance S using T1 and S1 as dimensionality reduction eigenmatrix population distribution.

By a large amount of historical sample through before 1)-3) calculating after step, obtain sample state evaluation model, that is: dimensionality reduction eigenmatrix population distribution expects the reliable estimator of T and variance S, namely based on expectation T and the variance S of mahalanobis distance.Wherein, this mahalanobis distance obeys card side's distribution that degree of freedom is K, when meeting d > d _thresholdtime be considered as exceptional sample.

2) detailed process of exceptional sample identification

The target of exceptional sample identification is based on sample state evaluation model, the monitor sample of certain equipment a period of time is carried out calculating and evaluating, and provides relevant quantizating index: intensity of anomaly and degree of stability.Intensity of anomaly index refers to that reflection equipment departs from the degree of normal condition in this time period operating mode, degree of stability index is whether this equipment of reflection is stablized in the performance of this time period, it is supplementary (the more typical situation to intensity of anomaly, some equipment, its intensity of anomaly index and local mean value may be normal, but its local variance is larger, illustrate that this equipment may be in the boundary stage that is normal and deterioration).

2.1) ID of target object and secondary device is set, and the observation time section [t1t2] of this secondary device, from sample data, take out the data subset satisfied condition, as sample to be tested;

2.2) step 1.2 is utilized) the middle dimensionality reduction transition matrix Q generated _tdimensionality reduction is carried out to sample to be tested, obtains the sample to be tested after dimensionality reduction;

2.3) utilize step 1.3) in generate covariance matrix S as mapping matrix, the sample to be tested after dimensionality reduction is projected to population sample space, obtains normalized sample to be tested.

2.4) local mean value and the variance of standardization sample to be tested is calculated.Using sample to be tested average and step 1.3) population sample average after iteration optimization carries out distance and calculates as intensity of anomaly desired value, using sample to be tested variance as degree of stability desired value.

The present invention contrasts prior art, has the following advantages:

Hidden danger method for digging based on secondary device monitoring historical data of the present invention, Discovery Status Monitoring Data surface is early in the stage of ' normally ', and has been in the hidden danger equipment in degradation trend slowly.The method is based on the large data of Historical Monitoring, through flow processs such as sample data structure, principal component analysis (PCA), distance calculating, reliable statistics probability distribution and exceptional sample identifications, the evaluation quantized has been had to the stability of the performance of equipment and intensity of anomaly, overcome the deficiency correctly cannot grasping equipment health status in current secondary overhaul of the equipments, be convenient to operation maintenance personnel scented a hidden danger early before secondary device fault, the scope of reduction equipment maintenance, reduces the cost of maintenance.Meanwhile, in the same way the Monitoring Data after maintenance can also be carried out distance to calculate, to the assessment that the effect after maintenance quantizes.

Accompanying drawing explanation

Fig. 1 is secondary device hidden danger method for digging schematic diagram of the present invention;

Fig. 2 is pivot contribution degree curve synoptic diagram;

Fig. 3 (a) and 3 (b) are the intensity of anomaly of equipment state and the degree of stability schematic diagram of state in two setting-up time sections.

Embodiment

For certain secondary device, illustrate how applicating history Monitoring Data carries out hidden danger excavation, and method step is as shown in Figure 1.

First, collect the Condition Monitoring Data of the secondary device of certain model, as shown in table 1 below, sample number is 32915:

ID	Time	F1	F2	F3	F4
						115686216099168301	2015-02-1318:39:45.0	1.142	0.482	0.249	0.668
115686216099168302	2015-02-1318:39:45.0	2.721	1.056	6.095	18.486
						115686216099168304	2015-02-1318:39:45.0	1.22	0.404	1.467	6.212
115686216099168312	2015-02-1318:39:45.0	1.182	0.404	2.121	7.022
						…	…	…	…	…	…

Table 1

Wherein ID is classified as certain equipment of this model, and Time is Monitoring Data acquisition time, and F1-F4 is the status data gathered.

And then structural attitude matrix is as follows:

$Feature=\left[\begin{array}{cccc}1.142& 0.482& 0.249& 0.668\\ 2.721& 1.056& 6.095& 18.486\\ 1.22& 0.404& 1.467& 6.212\\ ...& ...& ...& ...\end{array}\right]$

The line number of eigenmatrix is the record number gathered.The data gathered in this example are sample data.

And then dimension-reduction treatment is carried out to eigenmatrix.When K=1 time, be mapped on straight line by former data set, suppose that rectilinear direction is by unit column vector w ₁represent, the variance after mapping is such as formula shown:

$Variance=\frac{1}{M}{\mathrm{\&Sigma;}}_{m=1}^M{\{x_m{w}_1-\stackrel{\&OverBar;}{x}{w}_1\}}^2=\frac{1}{M}{w}_1^T{\mathrm{Sw}}_1---\left(5\right)$

Wherein, S represents the covariance matrix of former data set

$S={\mathrm{\&Sigma;}}_{m=1}^M{(x_m-\stackrel{\&OverBar;}{x})}^T(x_m-\stackrel{\&OverBar;}{x})={(X-\stackrel{\&OverBar;}{X})}^T(X-\stackrel{\&OverBar;}{X})---\left(6\right)$

it is the vector of each column mean composition.Target finds vector of unit length w exactly ₁formula (6) is maximized, namely

$w_1=\arg \underset{\left|\right|w_1=1\left|\right|}{max}{w}_1^T{\mathrm{Sw}}_1---\left(7\right)$

By introducing Lagrange multiplier λ ₁, formula (7) is converted into a unconstrained optimization problem,

${\mathrm{maxw}}_1^T{\mathrm{Sw}}_1+{\mathrm{\&lambda;}}_1(1-w_1^T{w}_1)---\left(8\right)$

By to w ₁differentiate, obtains the expression formula of extreme point,

Sw ₁＝λ ₁w ₁(9)

From above formula, the λ at extreme value place ₁the eigenwert of covariance matrix, w ₁it is the proper vector of covariance matrix.And formula (9) is substituted into formula (8), then have

$Variance=\frac{1}{M}{w}_1^T{\mathrm{Sw}}_1=\frac{1}{M}{\mathrm{\&lambda;}}_1{w}_1^T{w}_1=\frac{1}{M}{\mathrm{\&lambda;}}_1---\left(10\right)$

For making maximum variance, λ ₁must be the maximum eigenwert of covariance matrix, and w ₁be and eigenvalue of maximum characteristic of correspondence vector, i.e. first principal component.As K>1, former data set be down to K dimension and retain information as much as possible, need former data set, be mapped to by the K dimension space being base with K eigenvalue of maximum characteristic of correspondence vector before the covariance matrix of former data set, this K proper vector is called as front K major component of former data set.Therefore through linear orthogonal transformation, under data set being mapped to one group of new coordinate system (namely generate one group new feature).First dimension of new coordinate system is that former data set changes maximum direction, is called first major component.Second dimension of new coordinate system and first dimension perpendicular, be the secondary direction greatly of former data set change, be called Second principal component, by that analogy.By eigenwert according to order arrangement from big to small, if λ ₁>=λ ₂>=...>=λ _p, normal orthogonal proper vector corresponding is with it designated as γ respectively ₁, γ ₂..., γ _p, primitive character matrix has p to tie up.And then obtain the p after changing and tie up the contribute information degree of main metadata to primitive character matrix and be:

As shown in Fig. 2, front bidimensional (k=2) pivot more than 95% is reached to the contribution degree of primitive character matrix, therefore, get front bidimensional pivot as [λ ₁, λ ₂], form pivot characteristic matrix F EATURES_PCA, preserve characteristic of correspondence vector [γ ₁, γ ₂].

$FEATURES\_PCA=\left[\begin{array}{cc}4.28& -1.33\\ 4.68& -0.32\\ 3.01& 3.18\\ ...& ...\end{array}\right]$

Then cluster centre estimation is carried out and distance threshold is determined.H the sample selecting corresponding mahalanobis distance minimum from dimensionality reduction eigenmatrix FEATURES_PCA, calculate its sample average T2 and covariance matrix S2, when meeting det (S2)=det (S1) or det (S2)=0, expect the reliable estimation of T and variance S using T1 and S1 as dimensionality reduction eigenmatrix population distribution.Otherwise, the mahalanobis distance dj of all samples is recalculated based on T2 and S2, and H the sample selecting corresponding mahalanobis distance minimum, calculate its sample average T3 and covariance matrix S3 ... so repeatedly, until stop iteration when det (Sn+1)=det (Sn) or det (Sn+1)=0.Finally obtain:

$S_n=\left[\begin{array}{cc}0.8409& -1.2827\\ -1.2827& 2.9991\end{array}\right]$

T _n＝[-4.2196-0.0155]

Expect that using Tn and Sn as dimensionality reduction eigenmatrix population distribution the reliable estimation of T and variance S stores.

And then by range formula, calculate the distance of each sample distance cluster centre Tn:

$\begin{array}{c}{d}_j=\sqrt{({\stackrel{\&RightArrow;}{x}}_j-T_n)S_n^{-1}{({\stackrel{\&RightArrow;}{x}}_j-T_n)}^T}\\ (j=1,2,...,N)\end{array}$

Wherein for the vector that an original jth feature samples transforms through dimensionality reduction eigenmatrix.

Sample mahalanobis distance obeys card side's distribution that degree of freedom is K, when meeting time be considered as exceptional sample, α is level of significance.And in this example, K=2, setting degree of confidence is 95%, therefore distance threshold thus can by d _jbe greater than the exceptional sample point identification of threshold value.

Finally, the state estimation of individual equipment different time sections is carried out.Using ID be the individual equipment of 115686216099168312 as object of observation, selected will to be divided into the time period apart from comparatively new data in last image data 10 days, and the data older before 10 days apart from last image data.Same according to method before, the calculating of expectation and variance that this Liang Pi local data is carried out distributing, thus obtain Local Clustering center.Finally two Local Clustering results are placed in overall cluster, contrast respectively, assess the intensity of anomaly of equipment state and the degree of stability of state in two setting-up time sections.

As shown in Figure 3, figure (b) is the partial enlarged drawing of figure (a), circle points is through carries out cluster analysis and Distance Judgment to all sample points, with the sample point of 95% confidence declaration normal condition, large stain is the sample point of ID=115686216099168312 individual equipment older time period (apart from last image data before 10 days), and triangle form point is equipment sample point of (apart from last image data before 10 days) within the nearlyer time period for this reason.As seen from the figure, the centre distance normal condition center of the sample point of time period far away is comparatively far away, and local center (center, circle points place) the distance normal condition center of the sample point in the nearlyer time period is comparatively near, and comparatively stable.A conclusion can be drawn thus, from Monitoring Data, ID=115686216099168312 individual equipment, although be still in abnormality at the status and appearance of nearly ten days, the trend that is improved of the status and appearance of (before 10 days) than ever.

Embodiments of the present invention are not limited thereto; under stating basic fundamental thought prerequisite on the invention; according to the ordinary technical knowledge of this area and customary means to content of the present invention make the amendment of other various ways, replacement or change, all drop within rights protection scope of the present invention.

Embodiments of the present invention are not limited thereto; under stating basic fundamental thought prerequisite on the invention; according to the ordinary technical knowledge of this area and customary means to content of the present invention make the amendment of other various ways, replacement or change, all drop within rights protection scope of the present invention.

Claims (2)

1. based on a secondary device hidden danger method for digging for large data technique, it is characterized in that comprising two processes: based on model training and the exceptional sample identification of secondary device historical data;

1) described as follows based on the particular content of the model training step of secondary device historical data:

1.1) the eigenmatrix detailed process obtaining secondary device Monitoring Data is as follows:

Pre-service and standardization are carried out to the different raw information of all kinds of secondary device, forming device feature samples matrix, wherein, when structural attitude sample matrix, by the distinct device of same model, mixing is in the matrix of sample data, and the matrix of described sample data is shown below:

Rawdata _j＝[IDTf _j,1f _j,2...f _j,m]

(j＝1,2,…,N)(1)

In formula (2), Rawdata _jrepresent the state-detection sample data of certain secondary device being numbered ID of jth bar at T sometime, such secondary device has m characteristic index, forms monitoring feature data successively;

The eigenmatrix of monitoring feature data is wherein:

Feature _j＝[f _j,1f _j,2...f _j,m]

(j＝1,2,…,N)(3)

Equally, j is sample index, and m represents m characteristic index, and N represents sample number;

1.2) described secondary device sample data principal component analysis (PCA) detailed process is as follows:

By described step 1.1) in the eigenmatrix [f of Monitoring Data of sample data _{j, 1}f _{j, 2}... f _j,m], through principal component analysis (PCA), be mapped in a K dimension space by eigenmatrix, the wherein dimension <N of K, make the data remained can maximize the information representing raw data set, and preserve dimensionality reduction transition matrix Q _t, eigenmatrix FEATURES_PCA after generation dimensionality reduction _j=[f _{j, 1}f _{j, 2}... f _j,k];

1.3) reliable statistics probability distribution concrete steps are as follows:

Initialization: randomly draw H sample from dimensionality reduction eigenmatrix FEATURES_PCA, wherein N/2≤H≤3N/4, calculates its sample average T1 and covariance matrix S1;

Calculate sample and depart from centre distance:

$\begin{array}{c}{d}_j=\sqrt{(FEATURES\_{\mathrm{PCA}}_j-T_1)S_1^{-1}{(FEATURES\_{\mathrm{PCA}}_j-T_1)}^T}\\ (j=1,2,\mathrm{...},N)\end{array}---\left(4\right)$

Iteration optimization: select H the sample that corresponding mahalanobis distance is minimum from dimensionality reduction eigenmatrix, its sample average T2 of iterative computation and covariance matrix S2, when meeting det (S2)=det (S1) or det (S2)=0, expect the reliable estimation of T and variance S using T1 and S1 as dimensionality reduction eigenmatrix population distribution;

By a large amount of historical sample through before 1)-3) calculating after step, obtain sample state evaluation model, that is: dimensionality reduction eigenmatrix population distribution expects the reliable estimator of T and variance S;

2) detailed process of described exceptional sample identification is as follows:

2.1) ID of target object and secondary device is set, and the observation time section [t1t2] of this secondary device, from sample data, take out the data subset satisfied condition, as sample to be tested;

2.2) step 1.2 is utilized) the middle dimensionality reduction transition matrix Q generated _tdimensionality reduction is carried out to sample to be tested, obtains the sample to be tested after dimensionality reduction;

2.3) utilize step 1.3) in generate covariance matrix S as mapping matrix, the sample to be tested after dimensionality reduction is projected to population sample space, obtains normalized sample to be tested;

2.4) the standardization local mean value of sample to be tested and variance is calculated: using sample to be tested average and step 1.3) population sample average after iteration optimization carries out distance and calculates as intensity of anomaly desired value, using sample to be tested variance as degree of stability desired value.

2. method according to claim 1, is characterized in that: described mahalanobis distance obeys card side's distribution that degree of freedom is K, when meeting d > d _thresholdtime be considered as exceptional sample.