CN105259895A  Method and monitoring system for detecting and separating micro fault in industrial process  Google Patents
Method and monitoring system for detecting and separating micro fault in industrial process Download PDFInfo
 Publication number
 CN105259895A CN105259895A CN201510661060.2A CN201510661060A CN105259895A CN 105259895 A CN105259895 A CN 105259895A CN 201510661060 A CN201510661060 A CN 201510661060A CN 105259895 A CN105259895 A CN 105259895A
 Authority
 CN
 China
 Prior art keywords
 variable
 fault
 improvement
 data
 industrial
 Prior art date
 Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
 Granted
Links
 238000004519 manufacturing process Methods 0.000 title claims abstract description 48
 230000000875 corresponding Effects 0.000 claims abstract description 25
 238000004458 analytical method Methods 0.000 claims abstract description 24
 238000000513 principal component analysis Methods 0.000 claims abstract description 20
 238000000926 separation method Methods 0.000 claims abstract description 16
 239000011159 matrix material Substances 0.000 claims description 36
 238000000034 method Methods 0.000 claims description 34
 238000001514 detection method Methods 0.000 claims description 26
 239000012141 concentrate Substances 0.000 claims description 16
 238000005259 measurement Methods 0.000 claims description 16
 238000004364 calculation method Methods 0.000 claims description 11
 238000010606 normalization Methods 0.000 claims description 5
 230000002159 abnormal effect Effects 0.000 claims description 4
 238000003745 diagnosis Methods 0.000 description 7
 210000000299 Nuclear Matrix Anatomy 0.000 description 5
 238000010586 diagram Methods 0.000 description 5
 238000005516 engineering process Methods 0.000 description 4
 210000003324 RBC Anatomy 0.000 description 3
 238000004422 calculation algorithm Methods 0.000 description 3
 230000000694 effects Effects 0.000 description 3
 230000001755 vocal Effects 0.000 description 3
 241000347972 Caucasus prunus virus Species 0.000 description 2
 230000015572 biosynthetic process Effects 0.000 description 2
 238000000354 decomposition reaction Methods 0.000 description 2
 238000009826 distribution Methods 0.000 description 2
 238000005755 formation reaction Methods 0.000 description 2
 230000006399 behavior Effects 0.000 description 1
 230000001276 controlling effect Effects 0.000 description 1
 230000001186 cumulative Effects 0.000 description 1
 239000000203 mixture Substances 0.000 description 1
 230000000717 retained Effects 0.000 description 1
 239000004065 semiconductor Substances 0.000 description 1
 230000035945 sensitivity Effects 0.000 description 1
 230000001568 sexual Effects 0.000 description 1
 238000004088 simulation Methods 0.000 description 1
 238000003860 storage Methods 0.000 description 1
Classifications

 G—PHYSICS
 G05—CONTROLLING; REGULATING
 G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
 G05B23/00—Testing or monitoring of control systems or parts thereof
 G05B23/02—Electric testing or monitoring
 G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
Abstract
The invention discloses a method and monitoring system for detecting and separating a micro fault in an industrial process. The method comprises steps of: acquiring and using sensor data in a normal condition of the industrial process as training data and establishing a principal component analysis model of the training data; giving an appropriate sliding time window width and computing an improvement reconstruction contribution value of each variable of each sample in the training data; determining the control limit of the improvement reconstruction contribution of each variable; acquiring and using sensor data in a realtime condition as test data; computing the improvement reconstruction contribution of each variable in the test data and comparing the improvement reconstruction contribution with the corresponding control limit in order to analyze the fault of the test data; and if a fault analysis result indicates that a fault happens, identifying the variable with a maximum improvement reconstruction contribution value as a fault variable in order to achieve fault separation. Compared with a method in the prior art, the method does not require a mathematic model of the industrial process and may detect and separate the micro fault in the industrial process.
Description
Technical field
The invention belongs to industrial process monitoring and fault diagnosis field, be specifically related to a kind of detection of industrial process small fault and separation method and monitoring system thereof.
Background technology
Modern industry coursescale is huge, complex structure, once process occurs extremely just may causing huge economic loss, even jeopardizes personal safety.Process monitoring and fault diagnosis technology effectively can improve reliability, the equipment maintainability of system and reduce accident risk, have become one of study hotspot of active procedure control field.In addition, comparatively serious fault is developed by small fault usually, many major disaster sexual behaviors occurred in history therefore be also failed to be caused by Timeliness coverage and solution by minor anomaly in system.Therefore, the harm of small fault can not be ignored, and detects timely and be separated small fault, and takes effective safeguard procedures significant to guarantee industrial process safety, Effective Function.
Because the complicacy of industrial process constantly increases, its accurate analytical mathematical models is difficult to obtain usually, thus traditional method for diagnosing faults based on model is restricted.And based on the process monitoring of data and method for diagnosing faults without the need to the accurate analytic model of process, just utilize process large amount measurement data in normal conditions to set up corresponding data model, and by this models applying in inline diagnosis.As the important branch based on data fault diagnostic field, multivariate statistical process monitoring technology achieves tremendous development in two times more than ten years in the past, and is successfully applied to the such as industrial process such as petrochemical complex, semiconductor manufacturing.Fault detect and fault reconstruction are two elementary objects of process monitoring, and whether the former there occurs fault for deterministic process, and the latter is for determining the kind of fault or finding out the variable broken down.In multivariate statistical process monitoring, pca method is widely used in fault detect; Reconstruct contribution plot method (AlcalaCF; QinSJ.Reconstructionbasedcontributionforprocessmonitori ng.Automatica; 2009; 45 (7): 15931600.) usually based on Principal Component Analysis Model, for the fault reconstruction after fault detect.But the detection perform of traditional pca method to small fault is poor, thus make fault rate of failing to report higher (verification and measurement ratio is lower).In addition, traditional reconstruct contribution plot method when processing small fault separation problem, also easily causing trouble variable location of mistake.Existing small fault diagnostic techniques is mainly improved conventional failure detection algorithm, thus make it to small fault sensitivity and obtain good detection perform, but seldom relate to fault reconstruction, and the computation complexity of some algorithm is higher, is unfavorable for practical application.Therefore, need badly propose a kind of new method for diagnosing faults and corresponding monitoring system thereof realize to the detection of industrial process small fault be separated.
Summary of the invention
For the abovementioned technical matters existed in prior art, the present invention proposes a kind of detection and separation method of industrial process small fault, it adopts following technical scheme:
The detection of industrial process small fault and a separation method, comprise the steps:
One section of sensor measurement data under a collection industrial process nominal situation as training dataset, and sets up the Principal Component Analysis Model of this training dataset;
The given suitable sliding time window width of b, the improvement that each variable of each sample of calculation training data centralization has reconstructs contribution margin;
C determines the control limit of the improvement reconstruct contribution of each variable;
Sensor measurement data under d collection industrial process realtime working condition is as test data, and the measurand in test data is corresponding with the measurand of training dataset in step a;
E calculates the improvement reconstruct contribution of each variable in test data, and the control corresponding with step c is limit and compared, and carries out fault analysis to described test data;
If the result of fault analysis in f step e occurs for there being fault, then the variable with greatest improvement reconstruct contribution margin is defined as fault variable, to realize fault reconstruction.
Preferably, step b is specially:
Improvement reconstruct contribution margin according to each variable of each sample of following formula calculation training data centralization has:
In formula, IRBC
_{i,k}subscript i, improvement reconstruct contribution IRBC, the wherein k=1 of what k represented calculating is training data concentrates ith variable of a kth sample ..., N represents that training data is concentrated and comprises N number of sample, i=1 ..., m represents that training data is concentrated and comprises m variable; Matrix Φ represents that the fault detect index in the Principal Component Analysis Model adopted when calculating IRBC is combined index, and account form is
wherein,
d=P Λ
^{1}p
^{t}, P is the load matrix of pivot,
for the load matrix of residual error, Λ is pivot characteristic value matrix, δ
^{2}and τ
^{2}be respectively fault detect index S PE and T in Principal Component Analysis Model
^{2}control limit; ξ
_{i}represent that m ties up the ith row of unit matrix;
when representing that sliding time window arrives training dataset kth sample, the sample average in window width, solve mode as shown in the formula:
In formula, W represents given sliding time window width, x
_{w}represent that training data concentrates w sample after having carried out data normalization (variable that training data is concentrated has zeromean, unit variance).
Preferably, described step c is specially:
The improvement reconstruct contribution margin IRBC of ith variable of all samples is concentrated according to the training data asked in step b
_{i,k}, k=1 ..., N, determines the control limit of the improvement reconstruct contribution of ith variable:
The control of the improvement reconstruct contribution of note ith variable is limited to
subscript CL represents control limit (controllimit), namely threshold value;
according to following logic value: training data concentrates all N number of improvement reconstruct contribution IRBC of ith variable
_{i,k}, k=1 ..., allow the ratio of 1% to exceed at the most in N set and control limit
wherein, 1% represents level of significance, and its mathematical expression is expressed as:
In formula, Prob represents probability (ratio), and α=0.01 is level of significance, within namely training data concentrates the improvement of ith of all samples variable reconstruct contribution margin to have the ratio of 99% to be all positioned at control limit.
Preferably, described step e is specially:
First calculate the improvement reconstruct contribution of each variable in test data, be shown below:
In formula,
subscript test represent ith variable of a kth sample in test data improvement reconstruct contribution;
when representing that sliding time window comprises test data kth sample, the sample average in window width, solve mode as shown in the formula:
In formula,
represent in test data w sample after having carried out data normalization;
Then, will to calculate
the control corresponding to step c is limit
compare:
If the improvement reconstruct contribution of ith variable of a kth sample in test data
exceed and control limit accordingly
then think that this moment k has abnormal generation;
Next, in order to reduce fault misdescription rate, fault analysis is carried out according to following logic:
In test data, if there are continuous three samples, the improvement reconstruct of its certain variable i is contributed all to exceed and is controlled limit accordingly, then think and have fault to occur; Otherwise, think that process is in nominal situation.
Preferably, described step f is specially:
If the result of fault analysis in step e occurs for there being fault, then determine the variable broken down according to following formula:
In formula, i
_{f}be the variable being confirmed as breaking down, at this variable of current time k, there is maximum improvement reconstruct contribution, thus achieve fault reconstruction.
In addition, present invention also offers a kind of detection and separating monitoring system of industrial process small fault, this monitoring system adopts following technical scheme:
The detection of industrial process small fault and a separating monitoring system, comprising:
For gathering one section of sensor measurement data under industrial process nominal situation as training dataset, and set up the module of the Principal Component Analysis Model of this training dataset;
For the module of the improvement reconstruct contribution margin that each variable of each sample of calculation training data centralization has;
For determining the module of the control limit of the improvement reconstruct contribution of each variable;
For gathering the module of the sensor measurement data under industrial process realtime working condition as test data, wherein, the measurand in test data is corresponding with the measurand of training dataset;
For calculating the improvement reconstruct contribution of each variable in test data, and limit with corresponding control and compare, described test data is carried out to the module of fault analysis;
For the variable with greatest improvement reconstruct contribution margin is defined as fault variable, to realize the module of fault reconstruction.
Tool of the present invention has the following advantages:
The detection of the industrial process small fault in the present invention and separation method and monitoring system thereof, under utilizing process nominal situation, data carry out modeling, without the need to the accurate mathematical model of industrial process, also without the need to the data under industrial process fault condition, are convenient to practical application; Achieve fault detect and fault reconstruction simultaneously; The fault analysis logic adopted can reduce rate of false alarm effectively.
Accompanying drawing explanation
Fig. 1 is a kind of detection of industrial process small fault and the schematic flow sheet of separation method in the embodiment of the present invention;
Fig. 2 is the schematic flow sheet of offline modeling process in the embodiment of the present invention;
Fig. 3 is the schematic flow sheet of inline diagnosis process in the embodiment of the present invention;
Fig. 4 is failure detection result (all variablees) schematic diagram based on the inventive method in the present invention one example;
Fig. 5 is failure detection result (variable) schematic diagram based on the inventive method in the present invention one example;
Fig. 6 is the failure detection result schematic diagram based on traditional pca method in the present invention one example;
Fig. 7 is the fault reconstruction result schematic diagram based on the inventive method in the present invention one example;
Fig. 8 is the fault reconstruction result schematic diagram based on conventional reconstruction contribution plot method in the present invention one example.
Embodiment
Basic thought of the present invention is: based on traditional pca method (principalcomponentanalysis, and reconstruct contribution plot method (reconstructionbasedcontribution PCA), RBC), by time slipwindow vocal imitation skill, propose new statistical indicator in order to the detection that realizes industrial process small fault be separated.
Below in conjunction with accompanying drawing and embodiment, the present invention is described in further detail:
Shown in composition graphs 1, a kind of detection of industrial process small fault and separation method, comprise the following steps:
One section of sensor measurement data under step S110 collection industrial process nominal situation as training dataset, and sets up the Principal Component Analysis Model of this training dataset;
The given suitable sliding time window width of step S120, the improvement that each variable of each sample of calculation training data centralization has reconstructs contribution margin;
Step S130 determines the control limit of the improvement reconstruct contribution of each variable;
Sensor measurement data under step S140 collection industrial process realtime working condition is as test data, and the measurand in test data is corresponding with the measurand of training dataset in step S110;
Step S150 calculates the improvement reconstruct contribution of each variable in test data, and the control corresponding with step S130 is limit and compared, and carries out fault analysis to described test data;
If the result of fault analysis in step S160 step S150 occurs for there being fault, then the variable with greatest improvement reconstruct contribution margin is defined as fault variable, to realize fault reconstruction.
Wherein, step S110, S120 and S130 are offline modeling process, and step S140, S150 and S160 are inline diagnosis process, i.e. online checkingi and detachment process.
Fig. 2 is the schematic flow sheet of the offline modeling process of the embodiment of the present invention, specific implementation step as mentioned below:
Step S210 gathers sensor measurement data under industrial process nominal situation as training dataset, is stored as the data matrix of two dimension, to and data matrix carries out standardization.
Particularly, the training data collected is stored as the data matrix X of two dimension
_{0}∈ R
^{n × m}, wherein, often row represents a sample, and often row represent a sensor variable, and data matrix comprises N number of sample, a m variable.
To X
_{0}carry out standardization, by X
_{0}each row cancellation average, unit variance data.Note X
_{0}average and the standard deviation of a middle m variable are respectively μ
_{i}and σ
_{i}, i=1 ..., m, the data matrix after standardization is designated as X.
Step S220 runs pivot analysis algorithm to the data matrix after standardization, obtains the parameter of Principal Component Analysis Model.
Particularly, X is utilized to calculate covariance matrix S, S=X
^{t}x/ (N1), carries out to S the matrix that Eigenvalues Decomposition obtains all proper vectors formation
with corresponding characteristic root matrix
wherein, characteristic root is according to descending sort.In an embodiment of the present invention, pivot number l is chosen based on accumulative variance contribution ratio criterion.According to the pivot number l Matrix dividing chosen
with
wherein, P is the load matrix of pivot, by
front l row form;
for the load matrix of residual error.Λ is pivot characteristic value matrix,
for residual error eigenvalue matrix, and Λ=diag{ λ
_{1}..., λ
_{l},
$\stackrel{~}{\mathrm{\Λ}}=diag\{{\mathrm{\λ}}_{l+1},...,{\mathrm{\λ}}_{m}\}.$
Step S230 calculates the nuclear matrix of fault detect index in Principal Component Analysis Model.
Particularly, in Principal Component Analysis Model, the three kinds of fault detect indexs often adopted are respectively square prediction error SPE, T
^{2}statistic and combined index φ.SPE and T
^{2}the nuclear matrix of statistic is respectively
with D=P Λ
^{1}p
^{t}.The nuclear matrix of combined index φ is
wherein, δ
^{2}, τ
^{2}be respectively SPE and T
^{2}control limit, can utilize card side distribution calculating try to achieve.
The given sliding time window width of step S240, the improvement that each variable of each sample of calculation training data centralization has reconstructs contribution margin.
Particularly, according to the improvement reconstruct contribution margin that each variable of each sample of following formula calculation training data centralization has:
In formula, IRBC
_{i,k}subscript i, the improvement reconstruct contribution IRBC of what k represented calculating is training data concentrates ith variable of a kth sample, wherein, k=1 ..., N represents that training data is concentrated and comprises N number of sample, i=1 ..., m represents that training data is concentrated and comprises m variable; Matrix Φ represents that the fault detect index in the Principal Component Analysis Model adopted when calculating IRBC is combined index φ, account form provides in step S230, adopts combined index φ to consider that the reconstruct contribution plot method based on combined index has higher fault reconstruction accuracy usually here; ξ
_{i}represent that m ties up the ith row of unit matrix;
when representing that sliding time window arrives training dataset kth sample, the sample average in window width, solve mode as shown in the formula:
In formula, W represents given sliding time window width, x
_{w}w the sample that after representing standardization, training data is concentrated.
Step S250 determines the control limit of the improvement reconstruct contribution of each variable.
Particularly, for ith variable, concentrate the improvement reconstruct contribution margin IRBC of ith variable of all samples according to the training data asked in step S240
_{i,k}, k=1 ..., N, determines the control limit of the improvement reconstruct contribution of ith variable:
The control of the improvement reconstruct contribution of note ith variable is limited to
subscript CL represents control limit (controllimit, CL), namely threshold value.
according to following logic value: training data concentrates all N number of improvement reconstruct contribution IRBC of ith variable
_{i,k}, k=1 ..., allow the ratio of 1% to exceed at the most in N set and control limit
wherein, 1% represents level of significance.Being expressed as mathematical expression is:
In formula, Prob represents probability (ratio), and α=0.01 is level of significance, within namely training data concentrates the improvement of ith of all samples variable reconstruct contribution margin to have the ratio of 99% to be all positioned at control limit.
Consider that sliding time window width is W, before training data concentrates ith variable, the improvement reconstruct contribution margin of W1 sample cannot be asked for, these can be improved reconstruct contribution margin in actual use procedure and carry out zero setting process, at N enough large or much larger than window width W when, this process can not affect the accuracy of offline modeling.
The inventive method is determining the control of fault detect in limited time, concentrates the improvement reconstruct contribution margin of certain variable under all sample situations based on training data, by given suitable level of significance, determines the control limit of the improvement reconstruct contribution of this variable.This mode determining to control to limit based on historical data, can avoid the difference due to Data distribution8 form to calculate to controlling limit theoretical value the trouble brought effectively.
Fig. 3 is the schematic flow sheet of the inline diagnosis process of the embodiment of the present invention, specific implementation step as mentioned below:
Sensor measurement data under step S310 collection industrial process realtime working condition is as test data, and measurand is wherein corresponding with the measurand that training data in offline modeling process is concentrated, and note test data is
Step S320 utilizes the mean variable value of training data in offline modeling process and standard deviation to carry out standardization to test data.Particularly, the average μ of m variable is concentrated according to offline modeling process training data
_{i}and standard deviation sigma
_{i}(i=1 ..., m) to test data
carry out standardization, be about to
in each element value deduct corresponding average μ
_{i}afterwards divided by corresponding standard deviation sigma
_{i}, the test data after note standardization is x
^{test}.
After step S330 normalized, in test data, the improvement reconstruct of each variable is contributed, and compares to the corresponding control limit obtained in offline modeling process, carries out fault analysis to test data.
Particularly, first calculate the improvement reconstruct contribution of each variable in test data, be shown below:
In formula,
subscript test represent ith variable of a kth sample in test data improvement reconstruct contribution, with distinguish training data in offline modeling process improvement reconstruct contribution;
represent that sliding time window comprises current uptodate test data, namely during kth test data sample, the sample average in window width, solve mode as shown in the formula:
In formula,
represent in test data w sample after having carried out data normalization.
By what calculate
the control corresponding to offline modeling process steps S250 is limit
compare:
If the improvement reconstruct contribution of ith variable of a kth sample in test data
exceed and control limit accordingly
then think that this moment k has abnormal generation.
Next, in order to effectively reduce fault misdescription rate, fault analysis is carried out according to following logic:
In test data, if there are continuous three samples, the improvement reconstruct of its certain variable i is contributed all to exceed and is controlled limit accordingly, then think and have fault to occur; Otherwise, think that process is in nominal situation.
Step S340, if the result of fault analysis occurs for there being fault, is then defined as fault variable by the variable with greatest improvement reconstruct contribution margin, thus realizes fault reconstruction.
Particularly, if the result of fault analysis in step S330 occurs for there being fault, then the variable broken down is determined according to following formula:
In formula, i
_{f}be the variable being confirmed as breaking down, at this variable of current time k, there is maximum improvement reconstruct contribution, thus achieve fault reconstruction.
The inventive method utilizes industrial process to run data in normal conditions to carry out modeling, and this model is used for inline diagnosis, thus without the need to the mathematics mechanism model of industrial process, simultaneously without the need to the data under industrial process fault condition, easy to implement.
The inventive method is based on pivot analysis PCA and reconstruct contribution plot RBC method, by time slipwindow vocal imitation skill, by proposing the new detection of statistical indicator realization to small fault, compared to the fault detect index in conventional P CA, more responsive to small fault; Meanwhile, the inventive method, by by time slipwindow vocal imitation skill, utilizes the reconstruct contribution improved to carry out fault reconstruction, compared to traditional RBC method, effectively can improve the accuracy of fault reconstruction; Compared to the existing technology detected for small fault, the inventive method not only solves small fault test problems, also solves small fault separation problem simultaneously.
In order to help to understand the inventive method, showing that it detects and the effect be separated for small fault simultaneously intuitively, below an example being described: this example is based on Matlab instrument, and the present invention will be described to utilize numerical simulation case.
(1) produce training data, and set up the Principal Component Analysis Model of this training data:
This example uses following equation to produce N=1000 normal sample:
Wherein, x
_{i}, i=1 ..., 6 represent 6 measurands, i.e. m=6; t
_{1}, t
_{2}, t
_{3}be zeromean gaussian variable, its standard deviation is respectively 1.0, and 0.8,0.6; In formula (7), noise item represents zero mean Gaussian white noise, and standard deviation is 0.1.Represent a sample according to every a line, abovementioned 1000 sample storage are twodimentional data matrix X by the mode that each row represents a variable
_{0}∈ R
^{1000 × 6}.To X
_{0}carry out standardization, note X
_{0}average and the standard deviation of a middle m variable are respectively μ
_{i}and σ
_{i}, i=1 ..., m, the data matrix after standardization is designated as X.X is utilized to calculate covariance matrix S, S=X
^{t}x/ (N1), carries out to S the matrix that Eigenvalues Decomposition obtains all proper vectors formation
with corresponding characteristic root matrix
wherein, characteristic root is according to descending sort.Choose pivot number l based on accumulative variance contribution ratio CPV criterion, the threshold might be set of CPV is 90%, and front l the characteristic root sum namely retained accounts for more than 90% of all characteristic root sums.In this example, pivot number is 3.According to pivot number Matrix dividing
with
wherein P is the load matrix of pivot, by
front l=3 row form;
for the load matrix of residual error.Λ is pivot characteristic value matrix,
for residual error eigenvalue matrix, and Λ=diag{ λ
_{1}..., λ
_{3},
Utilize the parameter of abovementioned Principal Component Analysis Model to calculate the nuclear matrix Φ of combined index φ in fault detect index, computing formula is
wherein,
with D=P Λ
^{1}p
^{t}be respectively fault detect index S PE and T in Principal Component Analysis Model
^{2}the nuclear matrix of statistic, δ
^{2}, τ
^{2}be respectively SPE and T
^{2}control limit, can utilize card side distribution calculating try to achieve.
(2) preset time window width, the improvement reconstruct contribution that each variable of each sample of calculation training data centralization has.
In this example, window width is chosen for W=100, according to the improvement reconstruct contribution margin IRBC that each variable of each sample in training dataset X after expression formula (1) and (2) normalized has
_{i,k}.
(3) the control limit of the improvement reconstruct contribution of each variable is determined.
For ith variable, concentrate the improvement reconstruct contribution margin IRBC of ith variable of all samples according to the training data calculated in this example (2) step
_{i,k}, k=1 ..., N, determines the control limit of the improvement reconstruct contribution of ith variable
particularly, control limit can be obtained according to formula (3)
in formula (3), α=0.01.Variable i traversal 1 ~ 6 just can obtain the control limit of the improvement reconstruct contribution of all 6 variablees.
(4) produce test data, and standardization is carried out to test data.
1000 and training dataset independently test sample book data are produced again based on expression formula (7), and from the 401st sample, at variable x
_{2}upper interpolation amplitude is the constant deviation fault of 0.6.It should be noted that this fault amplitude is less than variable x under nominal situation
_{2}standard deviation, thus think that this fault is small fault in this application.This example (1) step training data is utilized to concentrate the average μ of m=6 variable
_{i}and standard deviation sigma
_{i}(i=1 ..., m) standardization is carried out to the test data comprising fault.
(5) after normalized, in test data, the improvement reconstruct of each variable is contributed, and carries out fault analysis to test data.
First, according to the improvement reconstruct contribution of each variable in test data after formula (4) and (5) normalized, wherein, the window width used in formula (5) is consistent with the window width in this example (2) step, is also W=100.Then, will to calculate
control accordingly to limit in (in test data, the improvement reconstruct of ith variable of a kth sample is contributed) and this example (3) step
compare.If
exceed and control limit accordingly
then think this time be carved with abnormal generation.Next, in order to reduce fault misdescription rate, carry out fault analysis: according to following logic in test data, if there are continuous three samples, the improvement reconstruct of its certain variable i is contributed all to exceed and is controlled limit accordingly, then think and have fault to occur; Otherwise, think that process is in nominal situation.
After Fig. 4 illustrates standardization, in test data, the improvement reconstruct of each variable is contributed and controls limit accordingly, and 6 subgraphs represent the monitoring 6 variablees being improved to reconstruct contribution successively.As seen in Figure 4, after breaking down (from the 401st sample), the improvement reconstruct contribution of each variable all there occurs larger change, and especially the improvement reconstruct contribution of the 2nd, 5,6 variable is all far beyond corresponding control limit, indicates the generation of fault.In order to show the effect that put forward the methods of the present invention detects small fault more intuitively.Fig. 5 carries out refinement to the subgraph of the 2nd in Fig. 4, to illustrate after standardization variable x in test data
_{2}improvement reconstruct contribution and control accordingly limit.As seen in Figure 5, before the failure occurs (before the 400th sample), variable x
_{2}improvement reconstruct contribution be all positioned at control limit within (because window width is 100, front 99 improve reconstruct contributions be zero), fault misdescription rate is zero; After breaking down, through only having the delay of 9 samples, variable x
_{2}improvement reconstruct contribution just exceed control limit and remain on always control limit more than, there is very high fault detect rate.
Fig. 6 illustrates three kinds of fault detect indexs, i.e. SPE, T of traditional pca method
^{2}statistic and combined index φ carry out the effect of fault detect to test data.As seen in Figure 6, for the small fault comprised in test data, the fault detect rate of three kinds of fault detect indexs of traditional pca method is all lower (being respectively 20.3%, 1.8%, 15.8%), responsive not to small fault.
(6) fault reconstruction after fault detect.
If the result of fault analysis occurs for there being fault, then the variable with greatest improvement reconstruct contribution margin is defined as fault variable.Particularly, for the out of order test sample book of tool, determine the variable broken down according to formula (6), thus realize fault reconstruction.
Fig. 7 illustrates the result that the inventive method carries out fault reconstruction: horizontal ordinate representative sample, show only the sample comprising fault here, namely from the 401st sample; Ordinate expresses possibility variable i=1 of breaking down ..., 6; The point that in figure, any one is represented by asterisk, its implication is, is confirmed as fault variable in this sample moment (being determined by the horizontal ordinate of this point) variable i (value of i is determined by the ordinate of this point).In test data, fault has been added on variable 2.As seen in Figure 7, just in the starting stage that fault occurs, variable 1 and variable 5 are erroneously determined to fault variable, along with the movement of sliding time window, the cumulative effect of fault in window strengthens, and variable 2 is all correctly separated by all sample moment afterwards.Through calculating, the accuracy of fault reconstruction, namely in all samples broken down, variable 2 is confirmed as fault scaling of variables, is 95.7%.
Fig. 8 illustrates the fault reconstruction result of conventional reconstruction contribution plot method based on combined index, although correctly variable 2 is defined as fault variable more than the sample of half in fault sample, also has a lot of sample that variable 5 or variable 6 are defined as fault variable.The fault reconstruction accuracy of conventional reconstruction contribution plot method is adopted to only have 58.2%.
In addition, the invention allows for a kind of detection and separating monitoring system of industrial process small fault, this monitoring system uses the detection of a kind of industrial process small fault as above and separation method to monitor process.This supervisory system comprises:
For gathering one section of sensor measurement data under industrial process nominal situation as training dataset, and set up the module of the Principal Component Analysis Model of this training dataset;
For the module of the improvement reconstruct contribution margin that each variable of each sample of calculation training data centralization has;
For determining the module of the control limit of the improvement reconstruct contribution of each variable;
For gathering the module of the sensor measurement data under industrial process realtime working condition as test data, wherein, the measurand in test data is corresponding with the measurand of training dataset;
For calculating the improvement reconstruct contribution of each variable in test data, and limit with corresponding control and compare, described test data is carried out to the module of fault analysis;
For the variable with greatest improvement reconstruct contribution margin is defined as fault variable, to realize the module of fault reconstruction.
Certainly; more than illustrate and be only preferred embodiment of the present invention; the present invention is not limited to enumerate abovedescribed embodiment; should be noted that; any those of ordinary skill in the art are under the instruction of this instructions; made all equivalently to substitute, obvious form of distortion, within the essential scope all dropping on this instructions, protection of the present invention ought to be subject to.
Claims (6)
1. the detection of industrial process small fault and a separation method, is characterized in that, comprise the steps:
One section of sensor measurement data under a collection industrial process nominal situation as training dataset, and sets up the Principal Component Analysis Model of this training dataset;
The given suitable sliding time window width of b, the improvement that each variable of each sample of calculation training data centralization has reconstructs contribution margin;
C determines the control limit of the improvement reconstruct contribution of each variable;
Sensor measurement data under d collection industrial process realtime working condition is as test data, and the measurand in test data is corresponding with the measurand of training dataset in step a;
E calculates the improvement reconstruct contribution of each variable in test data, and the control corresponding with step c is limit and compared, and carries out fault analysis to described test data;
If the result of fault analysis in f step e occurs for there being fault, then the variable with greatest improvement reconstruct contribution margin is defined as fault variable, to realize fault reconstruction.
2. the detection of a kind of industrial process small fault according to claim 1 and separation method, it is characterized in that, described step b is specially:
Improvement reconstruct contribution margin according to each variable of each sample of following formula calculation training data centralization has:
In formula, IRBC
_{i,k}subscript i, the improvement reconstruct contribution IRBC of what k represented calculating is training data concentrates ith variable of a kth sample, wherein, k=1 ..., N represents that training data is concentrated and comprises N number of sample, i=1 ..., m represents that training data is concentrated and comprises m variable; Matrix Φ represents that the fault detect index in the Principal Component Analysis Model adopted when calculating IRBC is combined index, and account form is
wherein,
d=P Λ
^{1}p
^{t}, P is the load matrix of pivot,
for the load matrix of residual error, Λ is pivot characteristic value matrix, δ
^{2}and τ
^{2}be respectively fault detect index S PE and T in Principal Component Analysis Model
^{2}control limit; ξ
_{i}represent that m ties up the ith row of unit matrix;
when representing that sliding time window arrives training dataset kth sample, the sample average in window width, solve mode as shown in the formula:
In formula, W represents given sliding time window width, x
_{w}represent that training data concentrates w sample after carrying out data normalization.
3. the detection of a kind of industrial process small fault according to claim 2 and separation method, it is characterized in that, described step c is specially:
The improvement reconstruct contribution margin IRBC of ith variable of all samples is concentrated according to the training data asked in step b
_{i,k}, k=1 ..., N, determines the control limit of the improvement reconstruct contribution of ith variable:
The control of the improvement reconstruct contribution of note ith variable is limited to
subscript CL represents control limit;
according to following logic value: training data concentrates all N number of improvement reconstruct contribution IRBC of ith variable
_{i,k}, k=1 ..., allow the ratio of 1% to exceed at the most in N set and control limit
wherein, 1% represents level of significance, and its mathematical expression is expressed as:
In formula, Prob represents probability, and α=0.01 is level of significance, within namely training data concentrates the improvement of ith of all samples variable reconstruct contribution margin to have the ratio of 99% to be all positioned at control limit.
4. the detection of a kind of industrial process small fault according to claim 3 and separation method, it is characterized in that, described step e is specially:
First calculate the improvement reconstruct contribution of each variable in test data, be shown below:
In formula,
subscript test represent ith variable of a kth sample in test data improvement reconstruct contribution;
when representing that sliding time window comprises test data kth sample, the sample average in window width, solve mode as shown in the formula:
In formula,
represent in test data w sample after having carried out data normalization;
Then, will to calculate
the control corresponding to step c is limit
compare:
If the improvement reconstruct contribution of ith variable of a kth sample in test data
exceed and control limit accordingly
then think that this moment k has abnormal generation;
Next, fault analysis is carried out according to following logic:
In test data, if there are continuous three samples, the improvement reconstruct of its certain variable i is contributed all to exceed and is controlled limit accordingly, then think and have fault to occur; Otherwise, think that process is in nominal situation.
5. the detection of a kind of industrial process small fault according to claim 4 and separation method, it is characterized in that, described step f is specially:
If the result of fault analysis in step e occurs for there being fault, then determine the variable broken down according to following formula:
In formula, i
_{f}be the variable being confirmed as breaking down, at this variable of current time k, there is maximum improvement reconstruct contribution margin, thus achieve fault reconstruction.
6. the detection of industrial process small fault and a separating monitoring system, is characterized in that, comprising:
For gathering one section of sensor measurement data under industrial process nominal situation as training dataset, and set up the module of the Principal Component Analysis Model of this training dataset;
For the module of the improvement reconstruct contribution margin that each variable of each sample of calculation training data centralization has;
For determining the module of the control limit of the improvement reconstruct contribution of each variable;
For gathering the module of the sensor measurement data under industrial process realtime working condition as test data, wherein, the measurand in test data is corresponding with the measurand of training dataset;
For calculating the improvement reconstruct contribution of each variable in test data, and limit with corresponding control and compare, described test data is carried out to the module of fault analysis;
For the variable with greatest improvement reconstruct contribution margin is defined as fault variable, to realize the module of fault reconstruction.
Priority Applications (1)
Application Number  Priority Date  Filing Date  Title 

CN201510661060.2A CN105259895B (en)  20151014  20151014  A kind of detection of industrial process small fault and separation method and its monitoring system 
Applications Claiming Priority (1)
Application Number  Priority Date  Filing Date  Title 

CN201510661060.2A CN105259895B (en)  20151014  20151014  A kind of detection of industrial process small fault and separation method and its monitoring system 
Publications (2)
Publication Number  Publication Date 

CN105259895A true CN105259895A (en)  20160120 
CN105259895B CN105259895B (en)  20170811 
Family
ID=55099631
Family Applications (1)
Application Number  Title  Priority Date  Filing Date 

CN201510661060.2A Active CN105259895B (en)  20151014  20151014  A kind of detection of industrial process small fault and separation method and its monitoring system 
Country Status (1)
Country  Link 

CN (1)  CN105259895B (en) 
Cited By (7)
Publication number  Priority date  Publication date  Assignee  Title 

CN105629959A (en) *  20160223  20160601  清华大学  Industrial process fault detection method 
CN105700518A (en) *  20160310  20160622  华中科技大学  Fault diagnosis method during industrial process 
CN105894027A (en) *  20160331  20160824  华北电力科学研究院有限责任公司  Principal element degree of association sensor fault detection method and apparatus based on density clustering 
CN108595380A (en) *  20180314  20180928  山东科技大学  A kind of Blast Furnace Abnormal Conditions Judgement detection method 
CN109947076A (en) *  20190314  20190628  华中科技大学  A kind of industrial process method for diagnosing faults based on bayesian information criterion 
CN110293949A (en) *  20190606  20191001  山东科技大学  A kind of bullet train Pneumatic brake systems small fault detection method 
EP3674946A1 (en) *  20181228  20200701  AO Kaspersky Lab  System and method for detecting anomalies in cyberphysical system with determined characteristics 
Citations (4)
Publication number  Priority date  Publication date  Assignee  Title 

WO2008042759A2 (en) *  20060929  20080410  FisherRosemount Systems, Inc.  Online multivariate analysis in a distributed process control system 
CN101308385A (en) *  20080711  20081119  东北大学  Nonlinear procedure failure testing method based on twodimensional dynamic kernel principal component analysis 
CN101617279A (en) *  20060929  20091230  费舍柔斯芒特系统股份有限公司  Online multivariate analysis in the distributed process control system 
CN103279123A (en) *  20130521  20130904  沈阳化工大学  Method of monitoring faults in sections for intermittent control system 

2015
 20151014 CN CN201510661060.2A patent/CN105259895B/en active Active
Patent Citations (5)
Publication number  Priority date  Publication date  Assignee  Title 

WO2008042759A2 (en) *  20060929  20080410  FisherRosemount Systems, Inc.  Online multivariate analysis in a distributed process control system 
WO2008042759A3 (en) *  20060929  20090827  FisherRosemount Systems, Inc.  Online multivariate analysis in a distributed process control system 
CN101617279A (en) *  20060929  20091230  费舍柔斯芒特系统股份有限公司  Online multivariate analysis in the distributed process control system 
CN101308385A (en) *  20080711  20081119  东北大学  Nonlinear procedure failure testing method based on twodimensional dynamic kernel principal component analysis 
CN103279123A (en) *  20130521  20130904  沈阳化工大学  Method of monitoring faults in sections for intermittent control system 
NonPatent Citations (1)
Title 

郭小萍等: "基于改进重构贡献图的故障定位方法", 《仪器仪表学报》 * 
Cited By (11)
Publication number  Priority date  Publication date  Assignee  Title 

CN105629959A (en) *  20160223  20160601  清华大学  Industrial process fault detection method 
CN105700518A (en) *  20160310  20160622  华中科技大学  Fault diagnosis method during industrial process 
CN105700518B (en) *  20160310  20180612  华中科技大学  A kind of industrial process method for diagnosing faults 
CN105894027A (en) *  20160331  20160824  华北电力科学研究院有限责任公司  Principal element degree of association sensor fault detection method and apparatus based on density clustering 
CN105894027B (en) *  20160331  20190614  华北电力科学研究院有限责任公司  Pivot degree of association Transducerfault Detecting Method and device based on Density Clustering 
CN108595380A (en) *  20180314  20180928  山东科技大学  A kind of Blast Furnace Abnormal Conditions Judgement detection method 
EP3674946A1 (en) *  20181228  20200701  AO Kaspersky Lab  System and method for detecting anomalies in cyberphysical system with determined characteristics 
CN109947076A (en) *  20190314  20190628  华中科技大学  A kind of industrial process method for diagnosing faults based on bayesian information criterion 
CN109947076B (en) *  20190314  20200602  华中科技大学  Industrial process fault diagnosis method based on Bayesian information criterion 
CN110293949A (en) *  20190606  20191001  山东科技大学  A kind of bullet train Pneumatic brake systems small fault detection method 
CN110293949B (en) *  20190606  20210924  山东科技大学  Method for detecting tiny faults of air brake system of highspeed train 
Also Published As
Publication number  Publication date 

CN105259895B (en)  20170811 
Similar Documents
Publication  Publication Date  Title 

CN105259895A (en)  Method and monitoring system for detecting and separating micro fault in industrial process  
CN105700518B (en)  A kind of industrial process method for diagnosing faults  
CN103713628B (en)  Fault diagnosis method based on signed directed graph and data constitution  
CN103853152A (en)  Batch process failure monitoring method based on ARPCA (Autoregressive Principal Component Analysis)  
CN103631145B (en)  Multiple operating modes process monitoring method and system based on monitor control index switching  
CN102539192B (en)  Failure prediction method based on ICA reconstruction  
CN103389701B (en)  Based on the level of factory procedure fault Detection and diagnosis method of distributed data model  
CN108268893B (en)  Chemical industry park early warning method and device based on machine learning  
CN109489977B (en)  KNNAdaBoostbased bearing fault diagnosis method  
Shang et al.  Recursive dynamic transformed component statistical analysis for fault detection in dynamic processes  
CN109522948A (en)  A kind of fault detection method based on orthogonal locality preserving projections  
CN105676833A (en)  Power generation process control system fault detection method  
CN110348150A (en)  A kind of fault detection method based on dependent probability model  
CN103646013B (en)  Multiple fault reconstruction method based on covariance matrix norm approximation  
CN109407652A (en)  Multivariable industrial process fault detection method based on majorminor pca model  
CN104880217B (en)  A kind of fault sensor signal reconstruct method based on the measured value degree of association  
CN110990788A (en)  Bearing residual life prediction method based on ternary wiener process  
CN104035431B (en)  The acquisition methods of kernel functional parameter and system for nonlinear process monitoring  
CN107862324B (en)  MWSPCAbased CBR prediction model intelligent early warning method  
CN104615123B (en)  Knearest neighbor based sensor fault isolation method  
US20160369777A1 (en)  System and method for detecting anomaly conditions of sensor attached devices  
CN103853144A (en)  Onsite sensor fault detection method based on oilextraction production data  
CN111780800B (en)  Method and system for monitoring, isolating and reconstructing sensor fault  
CN103995985A (en)  Fault detection method based on Daubechies wavelet transform and elastic network  
CN105389475B (en)  A kind of electric power factory equipment fault detection method based on WAVELET PACKET DECOMPOSITION 
Legal Events
Date  Code  Title  Description 

C06  Publication  
PB01  Publication  
C10  Entry into substantive examination  
SE01  Entry into force of request for substantive examination  
GR01  Patent grant  
GR01  Patent grant 