CN103713628A - Fault diagnosis method based on signed directed graph and data constitution - Google Patents

Abstract

The invention discloses a fault diagnosis method based on a signed directed graph and data constitution. The fault diagnosis method includes the steps of firstly, collecting normal data to conduct offline training, decomposing pre-processed data through a PCA, and solving a control limit of an SPE; secondly, setting up an SDG model according to a system flow graph, and determining parameters of all variables V-mask after setting the missing alarm rate, the false alarm rate and the detection offset; thirdly, collecting process data of system unknown state in real time, monitoring a CUSUM of all the variables and the SPE of a sample, if the SPE exceeds the control limit, diagnosing that a fault happens to a system, determining an effective node through the CUSUM, and conducting reconstitution on data in all compatible path directions by searching for all possible compatible paths, wherein the direction with the largest fault isolation index is taken as a real fault propagation direction, the starting node in the direction is taken as a causal variable of the fault, and an event which causes the node to be abnormal is taken as the root cause of the fault.

Description

Method for diagnosing faults based on signed digraph and data reconstruction

Technical field

The present invention relates to the fault diagnosis system field of multivariate complication system, relate in particular to a kind of method for diagnosing faults based on signed digraph and data reconstruction.

Background technology

Along with industrial process control system is constantly towards large-scale, intellectuality and complicated future development, safety problem becomes one of subject matter of everybody care day by day.As one of core component of Process Control System, the fault detection and diagnosis of dynamic system (FDD) technology is in order to adapt to industrial system to improving reliability and reducing the needs of accident risk and formation and development.In the past few decades, troubleshooting issue has obtained the extensive concern of Chinese scholars, emerges the whole bag of tricks about fault detect and isolation.These methods can be divided into quilitative method and the large class of quantitative test two on the whole.Wherein, in quantitative analysis method, the method based on data-driven is to pay close attention in recent years maximum methods.

In the method for diagnosing faults based on data-driven, the method for diagnosing faults based on multivariate statistics is one of method receiving much concern in recent years.Traditional multivariate statistical method based on contribution plot is to differentiate which variable makes corresponding statistic exceed the most popular method of normal value, and those variablees that statistic is had to a maximum contribution value are considered to cause the causal variable of fault.But the deficiency of the method maximum is, contribution margin is easily transferred to its dependent variable from a variable, and the variable of contribution margin maximum not necessarily causes the basic reason of fault.In addition, traditional multivariate statistical method based on contribution plot is not considered the propagation problem of fault in system, is therefore difficult to detect the basic reason that causes fault.

Therefore, those skilled in the art is devoted to develop a kind of method for diagnosing faults based on signed digraph and data reconstruction, considers the propagation problem of fault in system, effectively diagnoses out the basic reason that causes fault.

Summary of the invention

Because the above-mentioned defect of prior art, technical matters to be solved by this invention is to provide a kind of method for diagnosing faults based on signed digraph and data reconstruction, the method use square prediction error (SPE) and accumulation and (CUSUM) statistic are carried out fault detect, by the sample data of all compatible path directions of SDG being reconstructed when fault occurs, after reconstruct, residual error changes the direction of propagation that maximum direction is considered to fault, the start node that the party makes progress is the causal variable that causes fault, effectively diagnoses out the basic reason that causes fault.

For achieving the above object, the invention provides a kind of method for diagnosing faults based on signed digraph and data reconstruction, comprise the steps:

Step 1: the multivariate normal data of some in acquisition system operational process, and described multivariate normal data is carried out to pre-service, as known state measurement data;

Step 2: the measurement data of known state described in step 1 is carried out to PCA decomposition, the measurement data of described known state is decomposed into pivot part and residual error part, and the control of square prediction error of obtaining the measurement data of described known state is limit;

Step 3: according to the architectural characteristic of described system and response characteristic, set up the oriented graphical diagram (SDG) of described system, the node of described oriented graphical diagram is system single argument parameter, and set rate of false alarm, rate of failing to report parameter, according to the SDG of system, determine the parameter value of described each variable V-mask of known state measurement data;

Step 4: the multivariable process data of some in real-time acquisition system operational process, and described multivariable process data are carried out to pre-service, as unknown state measurement data, add up the square prediction error SPE of described unknown state measurement data accumulation and CUSUM statistical value and described unknown state measurement data, if SPE exceeds the control limit of step 2, expression system breaks down;

Step 5: if system does not break down, repeating step four, if system breaks down, the both arms by V-mask judge that whether the node of described SDG is effective, for the described process data variable that surpasses the upper underarm of V, node symbol is respectively "+" and "-" number;

Step 6: determine after all node states, search for compatible paths all in described SDG, adjacent node symbol multiplies each other for positive path;

Step 7: on all effective node direction and described compatible path direction, described unknown state measurement data is reconstructed, the direction of reconstruct index maximum is just made as to the travel path of fault, judges that the start node on the travel path of described fault is the basic reason that causes fault to occur.

In better embodiment of the present invention, described multivariate normal data pre-treatment step in described step 1 is: first the multivariate normal data of the described some gathering is deducted to the average of described multivariate normal data, then divided by the variance of described multivariate normal data.

In another better embodiment of the present invention, in described step 2, PCA method is chosen pivot according to eigenwert contribution rate, requires contribution rate more than 85%.

In better embodiment of the present invention, in described step 2, square prediction error is controlled and is limit computing formula to be

wherein (1-α) * 100% is the degree of confidence of described control limit,

here λ _ithe large eigenwert of i for described pretreated described multivariate normal data sample covariance matrix.

In another better embodiment of the present invention, described in described step 3, the parameter of V-mask is: and h=d*k, the slope that wherein k is V, d be nearest sampled point from the distance of V fixed point, h be nearest sampled point from the distance of the upper underarm of V, α is rate of false alarm, β is rate of failing to report, the side-play amount (multiple of sample standard deviation) of δ for detecting, σ _xstandard deviation for sample.

In better embodiment of the present invention, if reconstruct direction is exactly the travel path of fault in described step 7, the square prediction error of the unknown state measurement data after reconstruct should obtain maximum reducing so, square prediction error reduce degree as shown in Equation (1):

${\mathrm{\η}}^2=\frac{y^{\mathrm{SPE}}-y_j^{\mathrm{SPE}}}{y^{\mathrm{SPE}}}=\frac{f_j^2{\left|\right|\widetilde{\mathrm{\ζ}}\left|\right|}^2}{y^{\mathrm{SPE}}}---\left(1\right)$

Its span is between 0～1, wherein y ^sPEfor the square prediction error of sample before reconstruct,

for the square prediction error of the unknown state measurement data after reconstruct,

for the projection of ζ in residual error space,

the principal component space projection matrix of C for measuring, in formula (1)

the η making ²maximum direction is just the travel path of fault, the basic reason that the start node on travel path occurs for fault.

Method for diagnosing faults based on signed digraph and data reconstruction provided by the invention has been considered the propagation problem of fault in system, SDG and data reconstruction are combined, by the sample on all compatible paths of SDG being reconstructed when fault occurs, the direction of wherein reconstruct index maximum is considered to the direction of propagation of physical fault, and start node on the direction of propagation is considered to cause the basic reason variable of fault.By the architectural characteristic of system is combined with data-driven method, make the contribution plot method that the present invention is more traditional have better diagnosis effect to causing the basic reason of fault.

Below with reference to accompanying drawing, the technique effect of design of the present invention, concrete structure and generation is described further, to understand fully object of the present invention, feature and effect.

Accompanying drawing explanation

Fig. 1 is the signed digraph of the TE simulating experimental system of a preferred embodiment of the present invention;

Fig. 2 is the Troubleshooting Flowchart of a preferred embodiment of the present invention;

Fig. 3 is the test data square prediction error figure of the TE procedure fault 1 of a preferred embodiment of the present invention;

Fig. 4 is the fault isolation index η of a preferred embodiment of the present invention ²histogram.

Embodiment

A method for diagnosing faults based on signed digraph and data reconstruction, carries out off-line training by gathering normal data, as known state measurement data, by PCA, pretreated data is decomposed, and then obtains the control limit of SPE.Then according to system flow, set up SDG model, after setting rate of failing to report, rate of false alarm, detection side-play amount, determine the parameter of each variable V-mask.Then the process data of real-time acquisition system unknown state, as unknown state measurement data, the CUSUM of each variable and sample SPE are monitored, if having surpassed, SPE controls limit, there is fault in expression system, by CUSUM statistic, determine effective node afterwards, by searching for all possible compatible path, data on all compatible path directions are reconstructed, the direction of its fault isolation index maximum, it is real fault propagation direction, the start node that the party makes progress is considered to the causal variable of fault, and cause the abnormal event of this node to be considered to produce the basic reason of fault.

Fig. 1 is the signed digraph of the typical emulation platform TE of fault diagnosis field process, node in figure is a single argument, the parameter of certain parts in expression system, solid line represents just to affect, be that start node increases, the node that solid arrow points to also increases, and dotted line is just contrary, in figure, in circle, represent node

As shown in Figure 2, concrete steps are as follows for method for diagnosing faults flow process based on signed digraph and data reconstruction:

Off-line modeling process:

1, data acquisition.The multivariate normal data of acquisition system process some under system normal operating condition.

2, data pre-service, by data center's nondimensionalization.As known state measurement data.The multivariate normal data collecting is lined up to matrix X _{n * m}, the number that wherein n is collecting sample, m is variable number, for X _{n * m}in each column vector x _i(i=1 ... m), pass through formula

obtaining average is that 0 variance is 1 new vectorial y _i, wherein

and S _ibe respectively x _iaverage and variance, matrix X like this _{n * m}after pre-service, become matrix Y _{n * m}.

3, pretreated data are carried out to PCA decomposition, and calculate the SPE control limit of fault detect.First obtain covariance matrix Y ^tthe eigenwert of Y and individual features vector λ _iand p _i(i=1 ... m), the pivot number k getting sets according to 96% of eigenwert summation, and the eigenwert characteristic of correspondence vector that front k is large has formed principal component space, and other m-k eigenwert characteristic of correspondence vectors have formed residual error space.Thus can be by sample vector y(m dimensional vector) be decomposed into

pivot part wherein

represent that sample vector y is at principal component space S _pprojection,

residual error part

represent that sample vector y is in residual error space S _rprojection,

wherein I is unit matrix.The square prediction error formula of sample y is $\mathrm{SPE}\left(y\right)={\left|\right|\tilde{y}\left|\right|}^2=y^T{(I-P)}^T(I-P)y.$ Control is limited to ${\mathrm{\δ}}^2=g^{\mathrm{SPE}}{\mathrm{\χ}}_{\mathrm{\α}}^2\left(h^{\mathrm{SPE}}\right),$ Wherein (1-α) * 100% is degree of confidence, and has:

$g^{\mathrm{SPE}}=\frac{{\mathrm{\θ}}_2}{{\mathrm{\θ}}_1},h^{\mathrm{SPE}}=\frac{{\mathrm{\θ}}_1^2}{{\mathrm{\θ}}_2}$

Here

λ _ii the eigenwert for covariance matrix.

4, set up system symbol Directed Graph Model, set the whether abnormal parameters of detection node variable.According to systematic procedure flow graph, draw the signed digraph of system, as shown in Figure 1.For whether decision node in diagnostic procedure is effective, whether the CUSUM statistic of carrying out detection node variable by V-mask here overrun.V-mask isolates the effective tool of undesired data in CUSUM table, and its shape is opening V font left, and in the same horizontal line, both distances are d for its summit and recently sampled point, and underarm is that slope is the slope line of k.Data in V word all represent normal data, otherwise are abnormal data.The major parameter of V-mask can be determined according to formula below:

$k=\frac{\mathrm{\δ}{\mathrm{\σ}}_x}{2}$

$d=\frac{2}{{\mathrm{\δ}}^2}\ln \left(\frac{1-\mathrm{\β}}{\mathrm{\α}}\right)$

h＝d*k

The slope that wherein k is V, d be nearest sampled point from the distance of V fixed point, h is that nearest sampled point is from the distance of the upper underarm of V.α is rate of false alarm; β is rate of failing to report; The side-play amount (multiple of sample standard deviation) of δ for detecting; σ _xstandard deviation for sample.

Inline diagnosis process:

5, real-time acquisition system process data, as unknown state measurement data.

6, calculate the square prediction error SPE of real time data, and itself and SPE control limit are compared, if surpass and control limit, return to 5 collections next one process datas.

If 7 surpass SPE, control limit, add up the accumulation of each variable and CUSUM statistic and the both arms by V-mask judges whether the node variable of SDG is effective node variable, for surpassing the above variable of underarm of V, node symbol is respectively "+" and "-".The computing formula of accumulation and CUSUM statistic is:

wherein m is number of samples,

by being surveyed sample average,

the average of i the variable that step 1 is calculated.

8, in SDG, search for all possible compatible path, adjacent node symbol multiplies each other on positive Gai path, path, if the symbol of SDG is "+", the analog value of correspondence direction vector is 1, if the symbol of SDG is "-", the analog value of correspondence direction vector is-1, otherwise is 0.For example, the respective nodes that detects the 1st variable and the 3rd variable is "+" and "-", other nodes are all that 0 and the 1st variable and the 3rd variable are on compatible path, value on direction vector correspondence position on this path is exactly corresponding node state 1 ,-1 or 0 so, and the direction vector that obtains this path after unit is

9, on all effective node direction and compatible path direction, unknown state measurement data is surveyed and is reconstructed.The object of reconstruct will be measured exactly sample and move along reconstruct direction, until from the distance close to principal component space, if reconstruct direction is exactly the direction of propagation of fault, the square prediction error of the unknown state measurement data after so mobile should obtain maximum reducing.Square prediction error reduce degree as shown in Equation (1):

${\mathrm{\η}}^2=\frac{y^{\mathrm{SPE}}-y_j^{\mathrm{SPE}}}{y^{\mathrm{SPE}}}=\frac{f_j^2{\left|\right|\widetilde{\mathrm{\ζ}}\left|\right|}^2}{y^{\mathrm{SPE}}}---\left(1\right),$

Its span is between 0～1, wherein y ^sPEfor the square prediction error of sample before reconstruct,

for the square prediction error of unknown state measurement data after reconstruct,

for the projection of ζ in residual error space, in formula (1)

the η making ²maximum direction is just the travel path of fault, the basic reason that the start node on travel path occurs for fault.

TE is an emulation platform, its emulation 20 faults, Fig. 3 is TE process while occurring (a certain fault) test data square prediction error figure, in figure, dotted line is to control limit.

Fig. 4 is the fault isolation index η after reconstruct on all effective nodes and compatible path ²histogram, horizontal ordinate represents the numbering in effective node and compatible path, wherein numbers 1～12 for effective node, numbering 13～15 is compatible path, is respectively:

As can be known from Fig. 4, fault isolation index on numbering 15 these compatible paths is maximum, therefore the travel path that this path is fault, be that XC change causes greatly reactor pressure P7 to increase, thereby cause trap pressure P13 and desorb pressure tower P16 to increase, owing to only having fault 1 can make XC component become large in 20 faults of TE process, XB component is constant, can conclude that thus fault 1 has occurred.

More than describe preferred embodiment of the present invention in detail.Should be appreciated that those of ordinary skill in the art just can design according to the present invention make many modifications and variations without creative work.Therefore, all technician in the art, all should be in the determined protection domain by claims under this invention's idea on the basis of existing technology by the available technical scheme of logical analysis, reasoning, or a limited experiment.

Claims (6)

1. the method for diagnosing faults based on signed digraph and data reconstruction, is characterized in that, comprises the steps:

Step 1: the multivariate normal data of some in acquisition system operational process, and described multivariate normal data is carried out to pre-service, as known state measurement data;

Step 2: the measurement data of known state described in step 1 is carried out to PCA decomposition, described known state measurement data is decomposed into pivot part and residual error part, and obtain the control limit of the square prediction error of described known state measurement data;

Step 3: according to the architectural characteristic of described system and response characteristic, set up the oriented graphical diagram (SDG) of described system, the node of described oriented graphical diagram is system single argument parameter, and set rate of false alarm, rate of failing to report parameter, according to the SDG of system, determine the parameter value of described each variable V-mask of known state measurement data;

Step 4: the multivariable process data of some in real-time acquisition system operational process, and described multivariable process data are carried out to pre-service, as unknown state measurement data, add up the square prediction error SPE of described unknown state measurement data accumulation and CUSUM statistical value and described unknown state measurement data, if SPE exceeds the control limit of step 2, expression system breaks down;

Step 5: if system does not break down, repeating step four, if system breaks down, the both arms by V-mask judge that whether the node of described SDG is effective, for the described process data variable that surpasses the upper underarm of V-mask, node symbol is respectively "+" and "-" number;

Step 6: determine after all node states, search for compatible paths all in described SDG, adjacent node symbol multiplies each other for positive path;

Step 7: on all effective node direction and described compatible path direction, described unknown state measurement data is reconstructed, the direction of reconstruct index maximum is just made as to the travel path of fault, judges that the start node on the travel path of described fault is the basic reason that causes fault to occur.

2. the method for diagnosing faults based on signed digraph and data reconstruction as claimed in claim 1, it is characterized in that, described multivariate normal data pre-treatment step in described step 1 is: first the multivariate normal data of the described some gathering is deducted to the average of described multivariate normal data, then divided by the variance of described multivariate normal data.

3. the method for diagnosing faults based on signed digraph and data reconstruction as claimed in claim 1, is characterized in that, in described step 2, PCA method is chosen pivot according to eigenwert contribution rate, requires contribution rate more than 85%.

4. the method for diagnosing faults based on signed digraph and data reconstruction as claimed in claim 1, is characterized in that, in described step 2, square prediction error is controlled and limit computing formula to be

, wherein (1-α) * 100% is the degree of confidence of described control limit,

here

λ _ithe large eigenwert of i for described pretreated described multivariate normal data sample covariance matrix.

5. the method for diagnosing faults based on signed digraph and data reconstruction as claimed in claim 1, is characterized in that, described in described step 3, the parameter value of V-mask is:

and h=d*k, the slope that wherein k is V, d be nearest sampled point from the distance of V fixed point, h be nearest sampled point from the distance of the upper underarm of V, α is rate of false alarm, β is rate of failing to report, the side-play amount (multiple of sample standard deviation) of δ for detecting, σ _xstandard deviation for sample.

6. the method for diagnosing faults based on signed digraph and data reconstruction as claimed in claim 1, it is characterized in that, if reconstruct direction is exactly the travel path of fault in described step 7, the square prediction error of the unknown state measurement data after reconstruct should obtain maximum reducing so, square prediction error reduce degree as shown in Equation (1):

${\mathrm{\η}}^2=\frac{y^{\mathrm{SPE}}-y_j^{\mathrm{SPE}}}{y^{\mathrm{SPE}}}=\frac{f_j^2{\left|\right|\widetilde{\mathrm{\ζ}}\left|\right|}^2}{y^{\mathrm{SPE}}}---\left(1\right)$

Its span is between 0～1, wherein y ^sPEfor the square prediction error of sample before reconstruct,

for the square prediction error of the unknown state measurement data after reconstruct,

for the projection of ζ in residual error space,

the principal component space projection matrix of C for measuring, in formula (1)

make η ²maximum direction is just the travel path of fault, the basic reason that the start node on the travel path of fault occurs for fault.

Images