CN103760889B - Fault based on Bayesian network separates fast method - Google Patents

Abstract

Fault based on Bayesian network separates a fast method, comprises the following steps: by historical data, obtain Bayesian network, the average of historical data and the variance-covariance matrix of historical data of system; The rate of false alarm of regulation system, and ask for and control limit by rate of false alarm, this control limits the use of in judging whether system breaks down; Read the sampled value of current time, and formed a vector, calculate the T of current system²Statistic; Judge current T²Whether statistic is less than is controlled limit, if so, represents that system, in normal operating condition, does not have fault, otherwise carries out next step operation; To T²Statistic is decomposed, and can derive unique MYT decompose based on Bayesian network, compared with decomposing, original MYT is decomposed to calculative decomposition and reduce to p item with traditional MYT, greatly reduces amount of calculation, fast to T²Statistic is decomposed, thereby orients the source of trouble fast; The method is applicable to inline diagnosis process and the system high to requirement of real-time.

Description

Fault based on Bayesian network separates fast method

Technical field

The present invention relates to fault diagnosis field, specifically one is applicable at industrial control process, especially in complicated workWhen process breaks down in industry control procedure, orient fast the method for the source of trouble.

Background technology

Once having an accident, complex industrial process may cause harmful effect to production safety, efficiency or product quality, forPrevent the generation of this type of thing, need to carry out in real time fault diagnosis to industrial process. For traditional fault based on modelDiagnostic method, is difficult to that a complex process is set up to accurate physical model and manages monitoring. And in complex industrial mistakeCheng Zhong, the moment produces a large amount of data that reflect process operation mechanism and running status to various sensors again. Therefore drive based on dataMoving method for diagnosing faults, when to process operation data analysis, without knowing system accurate Analysis model, can completeTo the fault detection and diagnosis of system.

Method for diagnosing faults based on complex industrial process data is a kind of effective detection and separation process faultMethod. The method is passed through industrial process key position sensor installation, by sensor is produced to data analysis, and canWhether judge exactly this industrial process and break down, the source of trouble where.

Whether break down, locate the source of trouble in order to ensure that industrial process normally moves, to judge, also not only need accuratelyNeed fast, therefore those skilled in the art are devoted to find rapid fault diagnosis method.

Summary of the invention

The object of this invention is to provide a kind of rapid fault diagnosis method, realize that quick diagnosis is also in complex industrial processIsolated fault, the guarantee course of work is normally moved.

For achieving the above object, the present invention passes through in conjunction with Bayesian network, fast to T²Statistic is decomposed, thereby fastOrient the source of trouble, ensure that industrial process normally moves.

The invention provides a kind of method for diagnosing faults based on Bayesian network, it is characterized in that, comprise the following steps:

(1), by the historical data of industrial process, construct the Bayesian network of industrial process, and calculate industrial process faultControl limit;

(2) the current T of calculating industrial process²Statistic;

(3) if current T²Statistic is less than controls limit, and industrial process, in normal operating condition, performs step (2);If current T²Statistic is greater than controls limit, and industrial process breaks down;

(4) in conjunction with Bayesian network, to current T²Statistic is carried out MYT decomposition;

(5) MYT step (4) being obtained decomposes item and detects successively the control limit that whether exceedes its correspondence;

(6) exceeding the MYT that controls limit, to decompose a corresponding Factors be exactly the source of trouble of fault, if Factors numberBe 0, MYT decomposes a source of trouble that corresponding variable is industrial process;

Factors number refers to that MYT decomposes a number that relies on variable.

Further, in step (1), calculate the control limit of industrial process fault, comprise the following steps:

(11) according to the historical data of industrial process, the mean vector of the non-fault data of calculating industrial process and variance-Covariance matrix;

(12) data of synchronization step (11) being calculated form a vector, and the dimension of vector represents synchronizationThe number of the number of probes of sampling or the variable of industrial process;

(13), according to the rate of false alarm of default, calculate the control limit of industrial process fault.

Further, in step (13), the method for the described control limit of the described industrial process fault of calculating is:

$\mathrm{CL}=T_{\mathrm{\α}}^2=\frac{p(n-1)(n+1)}{n(n-p)}{F}_{\mathrm{\α}}(p,n-p),$

Wherein: F_α(p, n-p) is that the free degree is that p and n-p, confidence level are the F distribution of 1-α, and p is system variable number, and n is sampleGiven figure.

Further, step (2) is calculated the current T of industrial process²Statistic, comprises the following steps:

(21) read the sampled value of the sensor of current time industrial process;

(22) sampled value of sensor, according to the order of the sampled value of sensor in historical data, is combined into sample vector;

(23), according to mean vector, sample vector is normalized;

(24), according to normalized sample vector and variance-covariance matrix, calculate the T that sample vector is corresponding²SystemMetering.

Further, in step (23), the method for the mean vector of calculating sample vector is:

$\mathrm{\μ}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{X}_i$

The mean vector that wherein μ is described sample vector, n is sample number, X_iFor i sample in described sample vectorThis.

Further, 6, the method for diagnosing faults based on Bayesian network as claimed in claim 1, it is characterized in that stepSuddenly control limit computational methods corresponding to (5) are: calculate variable X_i-1＝[x₁,…,x_i-1]^TOn variable x_iCorresponding control limitCL_i-1：

${\mathrm{CL}}_{i-1}=\frac{p(n-1)(n+1)}{n(n-p)}{F}_{\mathrm{\α}}(k,n-k),$

Wherein F_α(p, n-p) is that the free degree is that k and n-k, confidence level are the F distribution of 1-α, and p is system variable number, and n is sampleGiven figure, the sum that k is Factors, i-1 is variable x_iRely on variable number.

7, the method for diagnosing faults based on Bayesian network as claimed in claim 1, is characterized in that, knot in step (4)Close described Bayesian network, to described current T²The method that statistic is carried out MYT decomposition is: calculate variable X_i-1＝[x₁,…,x_i-1]^TOn variable x_iCorresponding MYT decomposes item

$T_{i\·1,...,i-1}^2={[{(X_{i-1}-\stackrel{\‾}{X_{i-1}})}^T{\mathrm{\Σ}}_{i-1}^{-1}{S}_{i-1}-(x_i-{\mathrm{\μ}}_i)]}^2/{\mathrm{\Δ}}_i,$

Wherein:For variable X_i-1Mean vector, Σ_i-1For variable X_i-1Variance-covariance matrix, S_i-1For variableX_i-1With variable x_iCovariance, μ_iFor variable x_iAverage, Δ_iFor variable x_iVariance.

Further, the MYT in step (5), step (4) being obtained decomposes item and detects successively the control that whether exceedes its correspondenceSystem limit, comprises the following steps:

(51) according to Factors number number MYT decomposed to item carry out classification;

(52) from the few rank of Factors number, detect MYT and decompose, whether exceed the corresponding limit of controlling separately;

(53) MYT in same rank decompose detect complete, if exist the MYT that transfinites to decompose, stop underMYT in one rank decomposes item and detects, and the MYT if there is no transfiniting decomposes item, continues the MYT in next rankDecompose item and detect, detect complete until all MYT decompose item.

Further, MYT decomposes a corresponding variable and comprises measurand and control variables.

Compared with prior art, the method for diagnosing faults based on Bayesian network provided by the invention has the following beneficial effect that hasReally:

Can derive unique MYT based on Bayesian network and decompose, compared with decomposing with traditional MYT, original MYT be decomposed and neededThe decomposition item calculating is from p2^p-1Reduce to p item, greatly reduce amount of calculation, can thereby can obtain rapidly fault separating resultingGreatly reduce amount of calculation, fast to T²Statistic is decomposed, thereby orients the source of trouble fast; The method is applicable to examine onlineDisconnected process and the system high to requirement of real-time.

Brief description of the drawings

Fig. 1 is the flow chart of the fault separating method based on Bayesian network of one embodiment of the present of invention;

Fig. 2 is the flow chart of an Application Example TE model of the present invention;

Fig. 3 is the Bayesian network of the TE model shown in Fig. 2;

Fig. 4 is the T of fault 6 correspondences of the TE model shown in Fig. 2²Statistic;

Fig. 5 A is fault 6 sources of trouble of the TE model shown in Fig. 2 amplitude change curves to dependent variable 20;

Fig. 5 B is fault 6 sources of trouble of the TE model shown in Fig. 2 amplitude change curves to dependent variable 44.

Detailed description of the invention

Below in conjunction with accompanying drawing, the fault separation fast method that the present invention is based on Bayesian network is elaborated.

Referring to Fig. 1, a kind of fault based on Bayesian network separates fast method, comprises the following steps:

The first step: structure Bayesian network, computation of mean values vector sum variance-covariance matrix, the rate of false alarm of regulation system, asksObtain Fault Control limit. Before complex industrial engineering is carried out to real-time sampling, need to construct according to the historical data of this processIts Bayesian network, in the time of structure Bayesian network, choose the variable number of this process, and variable number has determined Bayesian networkNodes and T²The amount of calculation that statistic is decomposed. Then calculate mean vector according to the non-fault data in historical dataAnd variance-covariance matrix. Mean vector, for by sample vector normalization, is then used from variance-covariance matrix oneCalculate T²Statistic. Specify again the fault misdescription rate of this process, and limit fault control by the Fault Control that rate of false alarm is tried to achieve this processThe T of system limit and the sampled value in each moment²Statistic compares.

Second step: read the sampled value of each sensor of current time, calculate T²Statistic. The sampled value of each sensorBe combined in a certain order sample vector, the sensor sample value that this order will be corresponding with each historical data simultaneouslySequence consensus. Then by the sample vector normalization obtaining, then to obtain this sample vector corresponding with variance-covariance matrixT²Statistic. In order to ensure the uniformity at whole flow processing sample vector, need to be normalized sample.

The 3rd step: judge T²Whether statistic is less than is controlled limit CL. The T that second step is obtained²Statistic and the first step are initialThe control limit CL comparison that stage calculates, if T²Statistic is less than controls limit, thinks that process does not break down, and enters nextThe sampling period in moment; If T²Statistic is greater than controls limit, thinks that process breaks down, and proceeds next step faultSeparate; If T²Statistic equals to control limit, and at this moment process is in critical condition, for the healthy operation of guarantee process, the present inventionAlso will carry out in the case fault mask work.

The 4th step: in conjunction with Bayesian network, to T²Statistic is carried out MYT decomposition, MYT (Mason, YoungandTracy)Decomposition is a kind of by T²The method of decomposing, by the Bayesian network of this process, as shown in Figure 3, to T²Statistic is carried out MYT decompositionObtain totally 52 decomposition. Wherein: Factors number is totally 14 of the decomposition items of 0, Factors number be 1 decomposition item altogether22, Factors number is 14 of the decomposition items of 2, and Factors number exceedes 2 of the decomposition items of 2. Here, condition because ofSubnumber represents that this MYT decomposes a number that relies on variable in Bayesian network.

The 5th step: the MYT obtaining is decomposed to item and detect successively the control limit that whether exceedes its correspondence. The present invention is according to conditionBecause of the size of subnumber from small to large these 52 of hierarchical detection decompose item and whether exceed the corresponding limit of controlling separately. Once detect with one-levelAfter find to have the decomposition item that transfinites, stop the detection to next stage.

The 6th step: exceed a MYT decomposition source of trouble that corresponding Factors is this fault of controlling limit. In previous stepIn, the corresponding Factors of decomposition item that exceedes control limit is just the source of trouble. If corresponding Factors number is 0, shouldDecompose a corresponding variable source of trouble for process for this reason. After to the further inspection and maintenance of the source of trouble, just entering fault diagnosisIn normal flow process, wait for that entering the sampling period in next moment proceeds fault diagnosis.

Provide an Application Example of the present invention below in conjunction with Fig. 2, Fig. 3, Fig. 4, Fig. 5 A and Fig. 5 B. To the present invention do intoThe description of one step. Wherein, Fig. 2 is the TE(TennesseeEastman of this application embodiment) model flow chart, TE model be byDowns and Vogel do according to the actual process flow process of Eastman chemical company that a few modifications proposes, and this process is no matter be backLine structure or course of reaction are all very complicated, the complex working condition in can well simulating reality. This model comprises fiveFormant, respectively: reactor, condenser, compressor, separator and desorber.

Fig. 3 is the TE model Bayes net shown in Fig. 2; Fig. 4 is the T of the fault 6 in the present embodiment²Statistic schematic diagram;In Fig. 5 A and Fig. 5 B, show fault 6 times the amplitude situation of change of the source of trouble to dependent variable. The present embodiment adopts TE modelTest, this model has 41 measurands and 12 control variables. TE model comprises 21 class faults, comprising 16 classesKnown type fault and 5 class UNKNOWN TYPE faults. Wherein the step of fault 1-7 and system variable change relevant, fault 8-12 withThe changeability of system variable increases relevant, and fault 13 is relevant with drift, and fault 14,15 and 21 is with valve viscous relevant, and all the other areUNKNOWN TYPE fault. The present invention tests all 21 faults, except fault 3,9,11,15 and 21, to all the other faultsFault detect accuracy rate is all more than 90%, and wherein the fault detect rate to fault 6,7,1,4,5,14 and 12 (is arranged from high to lowRow) reach especially 99%. The present invention will stress fault 6.

In the present embodiment, select all 41 measurands and front 11 control variables totally 52 variable composition dataCollection. 500 samples that history data set comprises normal condition down-sampling, the time in sampling interval is 3 minutes. Pass through historical dataCollect, train the Bayesian network of TE model, as shown in Figure 3. In Fig. 3, each node represents a variable, wherein: father's nodeNumber is totally 14 of the variablees of 0, and father's node number is totally 22 of the variablees of 1, and father's node number is the variable 14 of 2Individual, and father's node number exceedes 2 of the variablees of 2.

All have a test data set for 21 class faults in TE model, wherein, each data set comprises 960 samplesThis, front 160 samples are the data under normal condition, introduce fault since the 161st sample.

By the sample under 500 normal conditions, i.e. history data set, calculates mean vectorμ isThe column vector of one 52 dimension, variance-covariance matrix Σ=XX^T, wherein X=[x₁,…,x_500], Σ is one 52 × 52Matrix. In the present embodiment, set fault misdescription rate, α=0.01, the rate of false alarm is here the confidence level in statistics namely.Because variance-covariance matrix is to estimate to get from sample covariance matrix, control and be limited to so:

$\mathrm{CL}=T_{\mathrm{\α}}^2=\frac{p(n-1)(n+1)}{n(n-p)}{F}_{\mathrm{\α}}(p,n-p),$

Wherein: F_α(p, n-p) is that the free degree is that p and n-p, confidence level are the F distribution of 1-α, and p is system variable number, and n is sampleGiven figure. Substitution numerical value obtains Fault Control and is limited to:

F_0.01(52,868) are by the gained F that tables look-up_0.01(52,868) ≈ 1.54. Finally, corresponding Fault Control limit $\mathrm{CL}=T_{0.01}^2\≈85.$

According to sample vector X, mean vector μ and variance-covariance matrix Σ, calculate the T that sample vector is corresponding²SystemMetering:

T²＝(X-μ)^TΣ^-1(X-μ)

As shown in Figure 4, totally 960 the test data sample calculation in fault 6 are obtained to 960 T²Statistic. Wherein160 T that sample is corresponding²Statistic is all being controlled below limit, since the 161st sample, T²Statistic all control limit withOn. After fault 6 introducing processes, T²Statistic can well detect out of order existence.

Tradition MYT decompose exist p! Plant and decompose, it decomposes item is p2^p-1, taking 3 variablees as example, according to traditionalMYT decomposes, and it exists 6 kinds of decomposition, and it decomposes item is 12. Its decomposition is as follows:

$T^2=\begin{array}{c}{T}_1^2+T_{2\·1}^2+T_{3\·\mathrm{1,2}}^2=T_1^2+T_{3\·1}^2+T_{2\·\mathrm{1,3}}^2\\ =T_2^2+T_{1\·2}^2+T_{3\·\mathrm{1,2}}^2=T_2^2+T_{3\·2}^2+T_{1\·\mathrm{2,3}}^2\\ T_3^2+T_{1\·3}^2+T_{2\·\mathrm{1,3}}^2=T_3^2+T_{2\·3}^2+T_{1\·\mathrm{2,3}}^2\end{array}$

Decompose and Bayesian network is introduced to MYT, decomposing item is p item, with regard to 3 variablees above, only exists a kind of MYT to decomposeWith 3 decomposition.

Taking TE model used in the present invention as example, according to the Bayesian network of TE model, its MYT decomposes as follows:

Compared with decomposing with traditional MYT, can derive unique MYT based on Bayesian network and decompose, original MYT is decomposed and neededThe decomposition item calculating is from p2^p-1Reduce to p item, greatly reduce amount of calculation, thereby can obtain rapidly fault separating resulting.

Mention above, MYT decomposition obtains totally 52 decomposition. Wherein: Factors number is totally 14 of the decomposition items of 0,Factors number is totally 22 of the decomposition items of 1, and Factors number is 14 of the decomposition items of 2, and Factors number exceedes 22 of individual decomposition items.

According to the rate of false alarm α of system variable number p, sample number n used and regulation, each decomposes a corresponding eventBarrier is controlled and is limited to:

$T_{i\·1,...,i-1}^2=\frac{p(n-1)(n+1)}{n(n-p)}{F}_{\mathrm{\α}}(k,n-k),$

The sum that wherein k is Factors.

The decomposition item that is first 0 from Factors number starts to detect, and exceedes until detect the decomposition item of controlling limit.

Experiment of the present invention is carried out on training data. Therefore regulation fault misdescription rate is less than 0.01, namely beforeThe T2 statistic of 160 the normal samples number sum that number and rear 800 fault data T2 statistics do not transfinite that transfinites is less than10(960*0.01=9.6 ≈ 10) individual.

Decompose item at 14 that are 0 to Factors number and detect, the item that wherein transfinites isWithThey are corresponding respectivelyVariable 20 and variable 44, therefore variable 20 and variable 40 are fault variable.

The detection number corresponding according to fault misdescription rate above-mentioned, finally obtaining fault variable is variable 20 and variable44, do not need to carry out the detection of next stage. Therefore, the source of trouble of fault 6 is variable 20 and variable 44. Fault 6 is that " A charging is damagedLose (stream 1) ", the corresponding source of trouble is variable 20(compressor horsepower) and variable 44(A inlet amount (stream 1)). When fault 6 occursTime, the compressor horsepower rapid drawdown being connected with A charging aperture, control variables (being variable 44) has increased the inlet amount of A. Fig. 5 A and Fig. 5 BIn shown fault 6 times, the amplitude situation of change of variable 20 and variable 44.

More than describe preferred embodiment of the present invention in detail. Should be appreciated that those of ordinary skill in the art withoutNeed creative work just can design according to the present invention make many modifications and variations. Therefore, all technology in the artPersonnel are available by logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's ideaTechnical scheme, all should be in by the determined protection domain of claims.

Claims (8)

1. the method for diagnosing faults based on Bayesian network, is characterized in that, comprises the following steps:

(1), by the historical data of industrial process, construct the Bayesian network of described industrial process, and calculate described industrial processThe control limit of fault;

(2) calculate the current T of described industrial process²Statistic;

(3) if described current T²Statistic is less than described control limit, and described industrial process, in normal operating condition, is carried outStep (2); If described current T²Statistic is greater than described control limit, and described industrial process breaks down;

(4) in conjunction with described Bayesian network, to described current T²Statistic is carried out MYT decomposition;

(5) MYT step (4) being obtained decomposes item and detects successively the control limit that whether exceedes its correspondence;

(6) it is exactly the source of trouble of described fault that the described MYT that exceedes described control limit decomposes a corresponding Factors, if instituteStating Factors number is 0, and described MYT decomposes a source of trouble that corresponding variable is described industrial process;

Described Factors number refers to that described MYT decomposes a number that relies on variable;

Control limit computational methods corresponding to step (5) are: calculate variable X_i-1＝[x₁,…,x_i-1]^TOn variable x_iCorresponding controlSystem limit CL_i-1：

${\mathrm{CL}}_{i-1}=\frac{p(n-1)(n+1)}{n(n-p)}{F}_{\mathrm{\α}}(k,n-k),$

Wherein F_α(p, n-p) is that the free degree is that k and n-k, confidence level are the F distribution of 1-α, and p is system variable number, and n is sample number, kFor the sum of Factors, i-1 is variable x_iRely on variable number.

2. the method for diagnosing faults based on Bayesian network as claimed in claim 1, is characterized in that, calculates institute in step (1)The control limit of stating industrial process fault, comprises the following steps:

(11), according to the described historical data of described industrial process, calculate the mean vector of the non-fault data of described industrial processAnd variance-covariance matrix;

(12) data of synchronization step (11) being calculated form a vector, and the dimension of described vector represents synchronizationThe number of the number of probes of sampling or the variable of described industrial process;

(13), according to the rate of false alarm of default, calculate the described control limit of described industrial process fault.

3. the method for diagnosing faults based on Bayesian network as claimed in claim 2, is characterized in that, step described in (13) isThe rate of false alarm that system is set is α, and the method for calculating the described control limit of described industrial process fault is:

$CL=T_{\mathrm{\α}}^2=\frac{p(n-1)(n+1)}{n(n-p)}{F}_{\mathrm{\α}}(p,n-p),$

Wherein: F_α(p, n-p) is that the free degree is that p and n-p, confidence level are the F distribution of 1-α, and p is system variable number, and n is sample number.

4. the method for diagnosing faults based on Bayesian network as claimed in claim 2, is characterized in that, described in step (2) is calculatedThe current T of industrial process²Statistic, comprises the following steps:

(21) read the sampled value of the sensor of industrial process described in current time;

(22) sampled value of described sensor, according to the order of the sampled value of sensor described in historical data, form sample toAmount;

(23) calculate the mean vector of described sample vector, according to described mean vector, described sample vector is normalized;

(24) according to normalized described sample vector and described variance-covariance matrix, calculate described sample vector pairThe T answering²Statistic.

5. the method for diagnosing faults based on Bayesian network as claimed in claim 4, is characterized in that, calculates institute in step (23)The method of stating the mean vector of sample vector is:

$\mathrm{\μ}=\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}{x}_i$

The mean vector that wherein μ is described sample vector, n is sample number, X_iFor i sample in described sample vector.

6. the method for diagnosing faults based on Bayesian network as claimed in claim 1, is characterized in that, in step (4) in conjunction with instituteState Bayesian network, to described current T²The method that statistic is carried out MYT decomposition is: calculate variable X_i-1＝[x₁,…,x_i-1]^TOnVariable x_iCorresponding MYT decomposes item

$T_{i\·1,...,i-1}^2={[{(X_{i-1}-{\stackrel{\‾}{X}}_{i-1})}^T{\mathrm{\Σ}}_{i-1}^{-1}{S}_{i-1}-(x_i-{\mathrm{\μ}}_i)]}^2/{\mathrm{\Δ}}_i,$

Wherein:For variable X_i-1Mean vector, Σ_i-1For variable X_i-1Variance-covariance matrix, S_i-1For variable X_i-1WithVariable x_iCovariance, μ_iFor variable x_iAverage, Δ_iFor variable x_iVariance.

7. the method for diagnosing faults based on Bayesian network as claimed in claim 1, is characterized in that, in step (5) to step(4) MYT obtaining decomposes item and detects successively the control limit that whether exceedes its correspondence, comprises the following steps:

(51) according to described Factors number number described MYT decomposed to item carry out classification;

(52) from the few rank of described Factors number, detect described MYT and decompose, whether exceed corresponding control separatelyLimit;

(53) it is complete that the described MYT in same described rank decomposes a detection, if there is the described MYT decomposition of transfiniting,Stop that the described MYT in rank described in next is decomposed to item and detect, the described MYT if there is no transfiniting decomposes item,Continue that the described MYT in rank described in next is decomposed to item and detect, detect complete until all described MYT decompose item.

8. the method for diagnosing faults based on Bayesian network as claimed in claim 1, is characterized in that, it is right that described MYT decomposes itemThe variable of answering comprises measurand and control variables.