CN105893700B - Based on the online fault detection and diagnosis technology of physics-big data mixed model Chemical Manufacture - Google Patents
Based on the online fault detection and diagnosis technology of physics-big data mixed model Chemical Manufacture Download PDFInfo
- Publication number
- CN105893700B CN105893700B CN201610263402.XA CN201610263402A CN105893700B CN 105893700 B CN105893700 B CN 105893700B CN 201610263402 A CN201610263402 A CN 201610263402A CN 105893700 B CN105893700 B CN 105893700B
- Authority
- CN
- China
- Prior art keywords
- model
- data
- variable
- performance
- performance indicator
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/70—Machine learning, data mining or chemometrics
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Medical Informatics (AREA)
- General Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- Chemical & Material Sciences (AREA)
- Crystallography & Structural Chemistry (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Evolutionary Biology (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
The invention discloses a kind of based on the online fault detection and diagnosis technology of physics-big data mixed model Chemical Manufacture, it is characterized in that, the operating method of the fault detection and diagnosis technology is as follows: after choosing object run unit, all historical datas of the unit will be scanned, after checking parameter, accident knowledge base and parameter model are established;During subsequent on-line checking, online data is importing directly into parameter model, the fault data in online data is obtained after scanning, is sounded an alarm, fault data and data in accident knowledge base are compared, obtain failure cause;Method of the invention is applied in determining single chemical engineering unit operation, establish reliable physical model, and the method for making Chemical Measurement is adopted in entire production process, introduce the real time data of big data processing technique processing several years, it is conceived to the failure of operator's performance indicator in controllable operating range of variables, eliminates the influence to fault detection of uncontrollable production process variable.
Description
Technical field
The present invention relates to safety detection in chemical production process and control fields, more particularly to a kind of physics-that is based on to count greatly
According to the online fault detection and diagnosis technology of the Chemical Manufacture of mixed model.
Background technique
Chemical production process is extremely complex, along with the physical and chemical reaction much not yet verified.To continuous process work
For industry, maintain a stable operating condition, can not only reduce accident and secondary disaster, also to stabilized product quality, optimizing
Operating condition under carry out optimal cost operate to obtain maximum economic benefit.Thus failure referred herein not only includes influencing dress
Safe abnormal conditions (being referred to as safety failure) is set, also includes deviation (the being referred to as property of product quality and Optimum Operation operating condition
Energy failure).
For complicated chemical production process, currently used fault detection and diagnosis method is all built upon data mostly
On the basis of driving model.In past 20 years, the professional of chemical industry and message area has carried out a large amount of beneficial explorations, especially
It is the construction in product practice library and universal before and after 2000, is deep into chemical production field to informationization technology and brings
Opportunity expedites the emergence of out many application on site, but also the application on site of fault detection and diagnosis technology is rapidly developed.Often at present
Fault detection and diagnosis method is all built upon on the basis of data-driven model mostly.
Statistical method based on production history data --- chemometrics modeling, i.e., by dividing a large amount of historical datas
Analysis, obtains the internal association of data, and note abnormalities (destruction of internal association usually between manufacturing parameter, or entire behaviour
Making operating condition has deviation relative to normal value) after, it accurately navigates in unit, equipment or the parameter of generation problem, passes through association
Influence factor displaying synchronous with analysis result, provides timely, accurate fault detection and examines for operator and administrative staff
Disconnected information.
With the rise of big data technology, the statistical processing methods of big data can be integrated, big data technology is compared to tradition
Statistical method the characteristics of more highlighting and only ask association, pay no attention to cause and effect, focusing on the relevance between card analysis data, big data skill
Art is complicated, implicit and influence relationship being difficult to the variable measured between factory is solved, and has in its distinctive feature, such as production unit A
Production status influences whether the quality (such as purity and sensitive impurity content) of unit A output intermediate products, if intermediate products
Quality can not on-line testing or testing cost it is high, these implicit relationships also will affect the production feelings of subsequent production unit B
Condition not only can be monitored and examine to the failure of unit A, but also in time series by big data analysis technology, can be right
The failure of production unit B carries out early warning, and reminds operator to make and timely adjust.But for Chemical Manufacture, very much
Unit operation especially physical process (as exchanged heat, compression, rectifying etc.), there is determining quality and energy balance relations, and
The forward/backward operation unit of process is commonly present physical isolation (two separative units such as are isolated by relay reservoir), between physical parameter not
In the presence of correlation physically, big data analysis technology is directlyed adopt, because existing noise effect between measurement data, or
It is to be influenced by public work (such as power steam temperature, fuel atmospheric pressure composition), associated data are physically not present in these
Usually occur the relevance of height on the model that big data is established, big data analysis personnel is allowed to be mistakenly considered the presence of association, this
Kind of the association for violating physical law so that the False Rate of fault diagnosis increases, cause production technician to fault detection generation not
Trust, brings certain resistance to promoting industry 4.0 and intelligent plant to build.
Summary of the invention
In view of the above problems, it is an object of that present invention to provide a kind of based on the change of physics-big data mixed model
Work produces online fault detection and diagnosis technology.
In order to achieve the above object, The technical solution adopted by the invention is as follows: a kind of be based on physics-big data mixed model
The online fault detection and diagnosis technology of Chemical Manufacture, the operating method of the fault detection and diagnosis technology is as follows: selecting
After taking object run unit, all historical datas of the unit will be scanned, after checking parameter, establish accident knowledge
Library and parameter model;During subsequent on-line checking, online data is importing directly into parameter model, after scanning
Fault data in online data out, sounds an alarm, and fault data and data in accident knowledge base are compared, must be out of order
Reason.
It is of the present invention based on the online fault detection and diagnosis technology of physics-big data mixed model Chemical Manufacture,
Its detailed operating method the following steps are included:
1) training data of the data composition modeling of factory level process is collected using real-time data base (such as IP21, PI)
Sample set: XRn×m.Wherein, n is the number of sample data set, and m is the variable number of sample data set;
2) it is directed to factory level process data collection XRn×m, the production unit Q of determining physical model is extracted, this is established
The physical model of production unit collects this production unit process variable in data subset Xq ∈ Rn×q, q is to have determining physics
The variable number of the unit operation of model.
Production unit Q herein is either one or more, The more the better under normal conditions, can more react chemical industry
The physical laws of production process and engineering knowledge is made full use of, rather than depends on analysis of statistical data merely.Factory level crosses number of passes
According to collection XRn×m, in the production unit for having determining physical model, the key performance of the production unit is extracted using physical model
Index is actually based on physical model and carries out dimensionality reduction, i.e., according to the chemical composition of the production unit, material, heat are so that change
Balance each other, chemical reaction equilibrium model, calculates the unit to the key contributions variable of target capabilities index, plays physics drop
The effect of dimension.
3) give up the original variable with the production unit for determining physical model, and use the pass extracted according to physical model
Unit operation subset Xq is reduced to k dimension from q dimension by key index k;
The original variable having in the production unit for determining physical model is removed in training sample, and instead from physics
The Key Performance Indicator extracted in model, by its with have neither part nor lot in physical modeling its under remaining performance variable reconfigure, shape
Training sample Xnew, the Xnew ∈ R of Cheng Xinn×p, data set determines the new variables number after production unit physical model dimensionality reduction.
4) Xnew is normalized;Wherein xu_i is the average value that Xnew is respectively arranged, and std_i is the mark that Xnew is respectively arranged
Quasi- poor, Xnorm=(Xnew_i-xu_i)/std_i after normalization obtains being no more than by the absolute value of transformed Xnorm
3.5, judge absolute value be more than 3.5 Xnorm for exceptional value.
5) the main metadata direction of factory level process, the monitoring and statistics amount and its monitoring limit of the grade that sets up a factory are extracted.
The pivot analysis or partial least square model of full level of factory are established with Xnew, wherein the index for the purpose of stability
It refers to only one group of operation operating condition, can not determine the quantification target of production process in real time;Wherein performance indicator refers to producing
Process has the relevant numerical indication of online or offline quality, cost.
If what is monitored is the index for the purpose of stability, pivot procedure fault detection model is established, forms factory
The Fault Model of grade range, calculates the control model in pivot load, SPE, Hotelling T^2 under different confidence levels
It encloses.
If monitoring is performance indicator, the specific steps of partial least square model are used are as follows: assuming that target capabilities become
Measure y and operating condition XnewRn×pIt is related, wherein preceding h performance variable x1, x2 ..xh are uncontrollable operating conditions, such as
Atmospheric temperature, the client that must satisfy can not make come to the demand of product, that is, operator to these performance variables
By adjustment;(p-h) a variable is the adjustable performance variable of operator after wherein.
Firstly, first layer simulated target performance variable y_act and uncontrollable operating condition xi (i=1 ... h) are established partially most
Small two multiply model;The predicted value y_cal of computation model, wherein residual error y_res=y_act-y_cal;Second layer model uses y_ again
Res and controllable operating condition xj (j=h+1 ... p) establish partial least square model, the model predication value y_rcl table of y_res
Show, the final predicted value of final goal performance variable can be expressed: y_t cl=y_cal+y_rcl.
According to the practical controllable range and its to the influence of performance indicator of operator, new performance indicator control zone is judged
Between, specific operating procedure are as follows: since the middle mass data that sample used is directly product practice library directly uses,
Therefore the control range of monitor control index uses the concept of big data, rather than is generallyd use using creation data is represented on a small quantity
95% confidence level;Can also be directly for statistical analysis to y_cer, 25% probability for taking its best as fault detection section,
Its physical significance is: historical normal operating level is 50%, and detection interval deviates 25% to the direction of good performance indexes,
That is, the operation operating condition lower than 75 points is accordingly to be regarded as failure if average is 50 points.
Meanwhile introduce production process sustained improvement concept, for the statistics of historical operating data, sample space it is equal
Value is equivalent to the performance level reached under 50% probability, and the Fault Control range of the first step can be energy under 60%-75% probability
The performance level reached, and pass through detection and diagnosis to failure, basic reason analysis and corrective measure, after 6-12 months,
Again historical data is extracted, modeling process more than repetition forms new performance indicator control interval.
6) new process data is collected, and it is pre-processed and is normalized.
7) the key of the unit is calculated according to the production unit model for having determining physical model with new process data
It can index.
8) processing method for pressing training sample, forms new process data collection, wherein there is the production list of determining physical model
Member is expressed by new Key Performance Indicator.
9) it is also calculated then in addition to SPE, Hotelling T^2 if it is using Partial Least Squares calculation of performance indicators
Current residual error (contribution of controlled variable) and its between ideal residual prediction value whether within control range.It will supervise in real time
Critical data is extracted in the data of survey, critical data is applied to step 4) and 5) obtained model or is carried out in calculation method
It calculates, finally judges in the new on-line monitoring process data of its calculating whether up to standard:
A) judgement of the pivot analysis of corresponding process exception;
B) the latent structure projection model of Key Performance Indicator.
The wherein calculation method of the latent structure projection model of Key Performance Indicator, detailed operating method are as follows: according to new
Production operation data, uncontrollable performance variable current value is input in the first layer model, current y ' _ cal is calculated;
According to current controllable operating variable, inputs in the second layer model, current y ' _ res is calculated, current actual performance refers to
Y ' _ act-y ' _ cal=y " _ res is marked, the actual value of as current controllable operating variable is as controllable by comparing y " res
Monitor control index under the influence of performance variable;The target value of the performance of current production process is y ' _ tcl, by comparing y ' _ act
Show performance fault occur if difference exceeds defined monitoring range with this difference of y ' t_tcl.
C) judgement whether broken down in controlled range;
10) on the basis of fault detection, current failure sensitive variable and insensitive variable are found out respectively, obtains the failure
Diagnostic result, instruct operator to carry out safe, high-quality and efficient production and control.
The present invention has the advantages that method of the invention is integrated application physics and big data method, in determining unit
In one chemical engineering unit operation (a certain specific process, equipment), reliable physical model (usually quality and energy balance mould are established
Type, and simple specific chemical reaction, are not related to complicated chemical reaction process), and in an entire production process (such as vehicle
Between multiple equipment) adopt the method (containing pivot analysis, offset minimum binary) for making Chemical Measurement, introduce at big data processing technique
The real time data for managing the several years, finds unknown failure, in the monitoring to performance fault, using the double-deck partial least square model,
It is conceived to the failure of operator's performance indicator in controllable operating range of variables, eliminates the shadow of uncontrollable production process variable
It rings.
Detailed description of the invention
Operating process schematic diagram Fig. 1 of the invention;
Fig. 2 is example heating furnace flow diagram of the invention;
Fig. 3 is that the pivot score of fault detection and diagnosis of the invention judges figure;
Fig. 4 is SPE the and T^2 calculated result figure of fault detection and diagnosis of the invention;
Fig. 5 is the key variables result figure of stable fault point occur in the present invention;
Fig. 6-a is detection and the monitored results figure of performance fault of the invention;
Fig. 6-b is performance position of failure point enlarged drawing in Fig. 6-a of the invention;
Fig. 7-a is the contribution plot of all controlled variables of the invention to performance fault;
Fig. 7-b is that preceding 10 performance fault points influence maximum variable in Fig. 7-a of the invention;
Fig. 8 and Fig. 9 is finally obtained main operating parameters table in the embodiment of the present invention
Figure 10 is the embodiment of the present invention Zhong Quan factory process units schematic diagram.
Specific embodiment
The present invention is described in further detail with specific embodiment for explanation with reference to the accompanying drawing.
Illustrate the validity of the method for the present invention now in conjunction with a specific factory level chemical production process example.The process
Flow chart as shown in Fig. 1 and Figure 10, determine by technical staff, whole process is by (the desalination, just of 6 different operating units
Evaporate, atmospheric pressure kiln, atmospheric tower group, heating under reduced pressure furnace, vacuum tower group) composition;Therefore finally obtained main operating parameters are such as
Shown in the table of Fig. 8 and Fig. 9.
Embodiment 1: detection and diagnosis to the stable fault of the system.
In order to test the validity of new method, it is labelled with regular data in data set and abnormal data set is tested,
Middle normal data includes 61225 data points, and 66 data points of stable fault occur (at 55370-55436
Position),
The first step collects the training sample set of data composition modeling: X using device real-time data baseR61291×112.Its
In, list sees attached list 1.
Second step be directed to determine physical model production unit, and according to physical model extract key performance refer to
Mark.
For feeding heating furnace in production technology, details referring to fig. 2, in heating furnace unit Q, has about 28 processes to become
Amount, XqR61291×28Some process variables be it is uncontrollable, such as heat dissipation capacity usually with the thermal insulation property of equipment itself and atmosphere
Temperature, humidity and wind speed are related, some process variables are controllable, such as the flow of combustion air, material outlet temperature etc..Such as
This 28 process variables are applied directly in failure monitoring by fruit, are not only increased calculation amount, are also brought data to system and make an uproar
Acoustic jamming.For heating furnace, the Key Performance Indicator of the production unit can be expressed there are two key parameter, and is had specific
Physical significance.
The wherein flow of fuel: F_fuel,
The efficiency of heating furnace: Ef, the temperature with heating furnace smoke evacuation, oxygen content contain CO content, and atmospheric conditions have clear
Physical model, can calculate and obtain.
Thus to 28 process variables of heating furnace unit, it now is possible to use F_fule, two new variables expression of Ef.
Using same method, will there are other production units for determining physical model in the process, extract key physical mould
Shape parameter has physics dimensionality reduction into ground, and effectively remains relevant material, heat balance pass system information.It is exemplified below:
The primitive operation parameter of each production unit for having determining physical model is removed into training sample set, and more than generation
The Key Performance Indicator for stating production unit forms the data set Xnew of new factory levelR61291×67。
Third step Xnew for data setsR61291×67It is normalized;Xu_i is the average value that Xnew is respectively arranged, and std_i is
The standard deviation that Xnew is respectively arranged, then Xnorm=(Xnew_i-xu_i)/std_i after normalizing, for statistical angle, warp
Cross transformation after absolutely mostly sample data value should between [- 3.5 ,+3.5], also more convenient interpretation beyond one of range it is different
Constant value.
It after dimensionality reduction and include that the Xnew of related physical information then crosses the method for carrying out conventional polytomy variable pivot analysis,
The covariance matrix of sample Xnew is after standardization
Xnew is decomposed by pivot:
Xnew=TP
T
+E
T=XP
Wherein P is load matrix, is made of into the preceding q feature vector of S, T is score matrix.
4th step establishes data statistics monitoring model
By directly observing two column of the starting of score matrix T, i.e. observational variable obtaining in two maximum principal component vectors
Point, fault condition can be monitored, since normal pivot mentions between point generally also [- 3.5,3.5] after normalization, deviates significantly from this
Can be considered for one section is abnormal, as shown in figure 3, the second pivot score is obviously beyond normal control when failure occurs in the figure
Range processed, fault detection success.
By constructing monitoring and statistics amount, i.e. square prediction error (SPE) and Hotelling ' s T^2 is realized to process
Stable state has multivariate statistics monitoring.SPE and T^2 has the upper control limit based on historical data, under different level of confidence.
Aspects SPE different from monitoring process.The degree that correlation is changed between SPE principal measure normal processes variable.T^2 degree
Measure distance of the current working apart from principal component subspace origin.
Wherein control limit
Hotelling ' s T^2 is calculated as follows:
The monitoring figure of SPE and T^2 works as system as shown in figure 4, its lower control limit in confidence level 95% also marks in figure
Metered amount significantly exceeds control and prescribes a time limit, i.e., it is believed that breaking down.When fault condition, the normal relation between usual variable is broken
Bad or variable on the whole far from nominal situation, or both occur simultaneously.If confidence level improves, (more sure judgement is different
Often), then red line in Fig. 4, wherein red line refers to that position will be moved up in the dotted line of the top in Fig. 4.
Using SPE and T^2 statistic another advantage is that by different variables to " contribution " ratio of SPE and T^2, base
It is as caused by which parameter when may determine that appearance is abnormal on this.It is certain point when occurring abnormal in upper page in left figure
Contribution of the variable to SPE and T^2, it can be seen which variable is unusual service condition be mainly caused by from left figure.It can help to grasp
Make personnel and screen influence factor rapidly, and makes correct judgement;Concrete outcome is as shown in Figure 5.
The judgement of the new process data of 5th step, such as obtains from the real-time data base of production process in real time, then failure is examined
Survey can be with real time implementation.
6th step calculates the physical model of production unit, extracts Key Performance Indicator, and merge other production operation variables,
It is normalized using the average value and standard deviation of sample obtained in third step.
Step 7: calculating current operation variable and in production unit physics mould by the pivot detection model that the 4th step obtains
The statistical monitoring statistic of the key physical index extracted in type.
8th step calculates each monitoring and statistics magnitude, forms the failure detection result of factory level process, judges active procedure
Operating status.
Embodiment 2: detection and diagnosis to the performance fault of the system:
Next combine the detailed process that implementation steps of the invention are set forth:
First step is the same as the first step in example 1.
Second step;With second step in example 1.
Third step carries out the double-deck regression analysis to performance batch mark, obtains the monitoring model of performance fault.
Extracted physical model key index dimensionality reduction obtains that (same example is normalized containing the Xnew of related physical information
1) and regression analysis Xnew and factory level Key Performance Indicator relationship, whether there is performance fault with monitoring process.
By taking the production process in attached drawing 1 as an example, it is (light that the Key Performance Indicator of factory level can be expressed as high value added product
The yield of matter oil product such as gasoline, diesel oil) and production process energy consumption.Both this need to comprehensively consider, and simplification is expressed as factory herein
Performance indicator, i.e. target value Y is made of following two difference a)-b):
A) unit light oil yield incremental income (petrol and diesel oil averagely repairs price-crude oil price) * (petrol and diesel oil flow/crude stream
Amount)
B) processed in units amount energy cost, i.e. (furnace fuel flow * fuel cost+steam flow * steam cost)/former
Oil stream amount is not counted in the target value of energy cost since water, electric dosage variation are smaller
The regression model of Y ~ Xnew is obtained using the double-deck Partial Least Squares.
Firstly, in XnewRn×pIn, some manufacturing variables are determined by raw material and finished product market, such as different oil varieties,
The content of itself light oil containing high added value is different, directly affects Y, and crude charging capacity comes from supervisory plan, and operates
Personnel can not arbitrarily change, thus first by Y and these uncontrollable factors totally 8 progress first layer Partial Least Squares Regression
The Y_act and uncontrollable operating condition xi (i=1 ... 8) for being calculated and being gone out with creation data establish offset minimum binary mould
Type;The predicted value Y_cal of computation model.
Residual error Y_res=Y_act-Y_cal is calculated again;Second layer model use again Y_res and controllable operating condition xj (j=
H+1 ... p) establishes partial least square model, the model predication value of Y_res is indicated with Y_rcl, then final goal performance variable
Final predicted value can be expressed: Y_tcl=Y_cal+Y_rcl.
Statistically see, when meeting certain recurrence accuracy, the error Y_ of performance indicator actual value and predicted value
Err close to normal distribution, once thus the difference that is believed that between Y_err=Y_act-Y_tcl go out beyond the range between one
Existing performance fault.Control interval can take the control interval under different confidence levels such as general statistical Process Control, usually compared with
Suitable for using on a small quantity having, to represent creation data be 95% confidence level generallyd use.It can also be used close using big data distribution
The concept of degree,;Namely it is directly for statistical analysis to Y_err, 25% probability for taking its best as fault detection section,
Physical significance is: historical normal operating level is 50%, and detection interval deviates 25% to the direction of good performance indexes,
That is, average is 50 points if total score 100 is divided, then the operation operating condition lower than 75 points is accordingly to be regarded as failure.
In order to test the validity of this method, the data set of use example one passes through calculated yield-energy consumption and performance index Y.Point
Y < L1 < L2 control interval is not marked in Fig. 6-a and Fig. 6-b, wherein the line with point in 6-b is current performance, and dotted line is alarm
Section, solid line are the sections that must be taken corrective action;
Finally show that two control lines of wherein figure top respectively correspond L1 and L2, when Y exceeds L1, that is, warning note has can
It can break down, if Y exceeds L2, operator must make corresponding adjustment, so that performance recovery is normal.
On fault point, by each controlled variable to the contribution degree of the predicted value Y_rcl of residual error Y_res, can directly by
The influence factor for influencing the fault point performance indicator is arranged by the sequence for influencing size, is found rapidly convenient for operator
The controllable operating variable for leading to performance fault after making adjustment to it, can disappear and drop performance fault.
Step 5: the judgement of new process data, such as obtains from the real-time data base of production process in real time, then failure is examined
Survey can be with real time implementation.Current operation operating condition is obtained, and calculates yield-energy consumption index Y ' of equivalent.
Step 6: calculating the physical model of production unit, Key Performance Indicator is extracted, and merges other production operations change
Amount, is normalized using the average value and standard deviation of sample obtained in third step.
Step 7: current variable input model is counted respectively by the offset minimum binary bilayer regression model that the 4th step obtains
Calculate Y_cal, Y_res, Y_tcl.
8th step calculates each monitoring and statistics magnitude, forms the failure detection result of factory level process, judges active procedure
Operating status.Y ', Y_tcl+L1, Y_tcl+L2 are got ready on same figure;The Y ' of current point is such as beyond Y_tcl+L1
It is out of order and reports total police, must take action if Y ' is beyond Y_tcl+L2, adjust controlled variable, the priority of adjustment can be with reference to figure
7-a and Fig. 7-b show that influence size of the controlled variable to residual error is sequentially adjusted from two figures, until current performance index
In controlled range.
It should be noted that above-mentioned is only presently preferred embodiments of the present invention, protection model not for the purpose of limiting the invention
It encloses, any combination or equivalents made on the basis of the above embodiments all belong to the scope of protection of the present invention.
Claims (6)
1. a kind of based on the online fault detection and diagnosis technology of physics-big data mixed model Chemical Manufacture, which is characterized in that
The operating method of the fault detection and diagnosis technology is as follows: after choosing object run unit, will own to the unit
Historical data is scanned, and after checking parameter, establishes accident knowledge base and parameter model;In subsequent on-line checking process
In, online data is importing directly into parameter model, the fault data in online data is obtained after scanning, is sounded an alarm,
Fault data and data in accident knowledge base are compared, obtain failure cause;
Its detailed operating method the following steps are included:
1) the training data sample set of the historical data composition modeling in factory level real time data: X ∈ R is utilizedn×m, wherein
N is the number that sample data concentrates sample, and m is the variable number that sample data concentrates each sample;
2) it is directed to factory level process data collection X ∈ Rn×mIn have determining physical model unit operation Q, by this production unit
Process variable is collected in data subset Xq ∈ Rn×q, q is the variable number for having the unit operation of determining physical model;
3) the Key Performance Indicator k having in the production unit for determining physical model is extracted, unit operation subset Xq is tieed up from q and is dropped
For k dimension;
4) original variable having in the production unit for determining physical model is removed in training sample, and instead from physics mould
It is reconfigured with the remaining performance variable for having neither part nor lot in physical modeling, is formed new by the Key Performance Indicator extracted in type
Training sample Xnew, Xnew ∈ Rn×p;Wherein, p is the variable for the participation training being reassembled into;
5) to Xnew ∈ Rn×pIt is normalized;Wherein xu_i is the average value that Xnew is respectively arranged, and std_i is what Xnew was respectively arranged
Standard deviation, Xnorm=(Xnew_i-xu_i)/std_i after normalization obtain not surpassing by the absolute value of transformed Xnorm
Cross 3.5, judge absolute value be more than 3.5 Xnorm for exceptional value;
6) the main metadata direction of factory level process, the monitoring and statistics amount and its detection model of the grade that sets up a factory are extracted;
If what is monitored is the index for the purpose of stability, pivot procedure fault detection model is established, forms factory level model
The Fault Model enclosed calculates the control range of pivot load, SPE, Hotelling T^2 under different confidence levels;
If monitoring is performance indicator, the method returned using dual residual error, the first multiple regression first presses performance indicator and life
Influence of operator's uncontrollable factor to performance indicator during production, the second multiple regression are strain with the residual error of the first multiple regression
Amount, the controllable operating condition of operator are that independent variable carries out Partial Least Squares Regression;
According to the practical controllable range and its to the influence of performance indicator of operator, new performance indicator control interval is judged,
It is obtained using the gap of current operation performance indicator and ideal performance index as failure monitoring index according to process improvement plan
Improvable control range in subsequent operation performance indicator;
7) collect new on-line monitoring process data, using the average of variable and standard deviation obtained in step 5) training sample into
Row pretreatment and normalization;
8) according to the on-line monitoring process data for having the cell processing of determining physical model new, the key performance for calculating the unit refers to
Mark, forms new process data collection, wherein having the production unit of determining physical model by new Key Performance Indicator to express;
9) critical data will be extracted in the data of real-time monitoring, by critical data be applied to step 5) and 6) obtained model or
It is calculated in calculation method, finally judges whether it is up to standard, calculated in new on-line monitoring process data:
The judgement of the pivot analysis of corresponding process exception;
The latent structure projection model of Key Performance Indicator;
The judgement whether broken down in controlled range;
10) monitoring and statistics magnitude is calculated, the failure detection result of factory level process is formed, judges the operating status of active procedure;
11) on the basis of fault detection, each performance variable when failure is found out respectively, the contribution margin of monitoring control amount is obtained
The diagnostic result of the failure.
2. according to claim 1 based on the online fault detection and diagnosis of physics-big data mixed model Chemical Manufacture
Technology, which is characterized in that the step 2) and 3) in determine: factory level process data integrates as X ∈ Rn×m, there is determining object
In the production unit for managing model, the Key Performance Indicator of the production unit is extracted using physical model, i.e., according to the production unit
Chemical composition, material, heat and chemical phase equilibrium, chemical reaction equilibrium model calculate the unit to target capabilities index
Key contributions variable, as based on physical model and carry out dimensionality reduction, indicate are as follows:
Xq∈Rn×q, obtained from physical model: Xq_new=Function (Xq), Xq ∈ Rn×k, meeting the same of key index
When, meet the dimensionality reduction requirement of big data processing.
3. according to claim 1 based on the online fault detection and diagnosis of physics-big data mixed model Chemical Manufacture
Technology, which is characterized in that when what is monitored in the step 6) is performance indicator, using the specific steps of partial least square model
Are as follows:
The target capabilities variable y and operating condition Xnew ∈ R after physical model dimensionality reductionn×pIt is related, wherein preceding h performance variable is
Uncontrollable operating condition, operator can not make free adjustment to these performance variables, wherein rear p-h variable is to grasp
Make the adjustable performance variable of personnel;
First layer simulated target performance variable y_act and uncontrollable operating condition xnewi, i=1 ..., h establish regression model;
The predicted value y_cal of computation model, wherein residual error y_res=y_act-y_cal;
Second layer model y_res and controllable operating condition xnewj, j=h+1 ..., p establish regression model, and model is to y_res's
Predicted value is y_rcl;
The final predicted value of target capabilities variable is expressed are as follows: y_tcl=y_cal+y_rcl;
Model prediction deviation can be expressed as: total deviation y_ter=y_act-y_tcl of first angle, the process of second angle
The deviation of controlled variable: y_cer=y_res- y_rcl.
4. according to claim 3 based on the online fault detection and diagnosis of physics-big data mixed model Chemical Manufacture
Technology, which is characterized in that the detection range to performance indicator failure is by the pre- of the practical controllable operating condition of operator
Survey the statistic of deviation y_cer, specific operating procedure are as follows:
By the step in claim 3, the y_cer of training sample is calculated, y_cer is generally near normal distribution;
Conventional standard deviation control interval can be used to y_cer, take the control range in 95% confidence interval, it can also be directly right
Y_cer is for statistical analysis, and 25% probability for taking its best is as fault detection section;
Sustained improvement is carried out in production process, by the detection and diagnosis to failure, basic reason analysis and corrective measure;In 6-
After 12 months, historical data is extracted again, and modeling process more than repetition forms new performance indicator control interval.
5. according to claim 3 based on the online fault detection and diagnosis of physics-big data mixed model Chemical Manufacture
Technology, which is characterized in that the calculation method of the latent structure projection model of Key Performance Indicator, detailed behaviour in the step 9)
Make method are as follows:
According to new production operation data, uncontrollable performance variable current value is input in the first layer model, is calculated and works as
Preceding model predication value y ' _ cal;
According to current controllable operating variable, inputs in the second layer model, current y ' _ rcl is calculated;
Current actual performance index y ' _ act-y ' _ cal=y ' _ res, y ' _ res are that current controllable operating variable prediction is residual
The actual value of difference, by comparing y ' _ res and y ' _ rcl, comparison result is the monitor control index under the influence of controllable operating variable;
The target value of the performance of current production process are as follows: y ' _ tcl;By comparing the difference between y ' _ act and y ' _ tcl, if
Difference exceeds defined monitoring range, then shows performance fault occur.
6. according to claim 1 based on the online fault detection and diagnosis of physics-big data mixed model Chemical Manufacture
Technology, which is characterized in that the index in step 6) for the purpose of stability refers to only one group of operation operating condition, can not determine in real time
The quantification target of production process;Wherein performance indicator refers to that production process has online or offline quality, cost phase
The numerical indication of pass.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610263402.XA CN105893700B (en) | 2016-04-26 | 2016-04-26 | Based on the online fault detection and diagnosis technology of physics-big data mixed model Chemical Manufacture |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610263402.XA CN105893700B (en) | 2016-04-26 | 2016-04-26 | Based on the online fault detection and diagnosis technology of physics-big data mixed model Chemical Manufacture |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105893700A CN105893700A (en) | 2016-08-24 |
CN105893700B true CN105893700B (en) | 2019-05-31 |
Family
ID=56704549
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610263402.XA Active CN105893700B (en) | 2016-04-26 | 2016-04-26 | Based on the online fault detection and diagnosis technology of physics-big data mixed model Chemical Manufacture |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105893700B (en) |
Families Citing this family (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106682080A (en) * | 2016-11-21 | 2017-05-17 | 云南电网有限责任公司电力科学研究院 | Transformer fault analysis method and system |
CN106872672B (en) * | 2016-12-28 | 2018-04-13 | 中南大学 | A kind of method for being hydrocracked runnability for on-line analysis |
CN106933097B (en) * | 2017-05-15 | 2021-07-20 | 青岛科技大学 | Chemical process fault diagnosis method based on multi-layer optimization PCC-SDG |
CN109213057B (en) * | 2017-06-29 | 2021-10-26 | 上海微电子装备(集团)股份有限公司 | Intelligent diagnosis device and method |
CN109298993B (en) * | 2017-07-21 | 2022-07-29 | 深圳市中兴微电子技术有限公司 | Method and device for detecting fault and computer readable storage medium |
CN107728599B (en) * | 2017-09-01 | 2020-12-18 | 北京中燕信息技术有限公司 | Method and device for determining state of refining device valve |
CN108388221B (en) * | 2018-03-21 | 2019-06-21 | 江南大学 | A kind of multilayered schema monitoring method of process flow industry process |
US11120350B2 (en) | 2018-03-21 | 2021-09-14 | Jiangnan University | Multilevel pattern monitoring method for industry processes |
CN108664696B (en) * | 2018-04-02 | 2023-04-07 | 国家计算机网络与信息安全管理中心 | Method and device for evaluating running state of water chiller |
CN108989083B (en) * | 2018-06-01 | 2021-02-09 | 安徽师范大学 | Fault detection performance optimization method based on hybrid strategy in cloud environment |
CN109327330B (en) * | 2018-09-11 | 2022-02-22 | 南京科思倍信息科技有限公司 | Chemical production abnormal slice management method based on data driving |
CN109816940B (en) * | 2019-03-21 | 2023-05-09 | 北京天诚同创电气有限公司 | Fault alarm method and device for sewage treatment plant |
CN112578740A (en) * | 2019-09-30 | 2021-03-30 | 冯恩波 | Fault diagnosis and processing method and system in industrial production process |
CN110956308A (en) * | 2019-10-29 | 2020-04-03 | 沈阳天眼智云信息科技有限公司 | Chemical equipment health state prediction method based on time sequence |
CN112784862A (en) * | 2019-11-07 | 2021-05-11 | 中国石油化工股份有限公司 | Fault diagnosis and identification method for refining process of atmospheric and vacuum distillation unit |
CN110912775B (en) * | 2019-11-26 | 2021-03-16 | 中盈优创资讯科技有限公司 | Internet of things enterprise network fault monitoring method and device |
CN114906817B (en) * | 2021-02-08 | 2023-12-05 | 中国石油化工股份有限公司 | Memory, hydrogen peroxide extraction process control method, device and equipment |
CN113239187B (en) * | 2021-04-13 | 2024-05-14 | 鹏城实验室 | Monitoring method based on multi-level industrial structure knowledge block division |
CN113620137B (en) * | 2021-06-24 | 2023-01-31 | 浙江力石工程有限公司 | Emergency manual door opening device for elevator door |
CN116107280B (en) * | 2023-04-12 | 2023-07-04 | 南京强思数字科技有限公司 | Digital twin plant control system for multidimensional visual management |
CN116794088B (en) * | 2023-08-03 | 2024-02-09 | 矿冶科技集团有限公司 | Online compensation method for copper flotation foam grade of X fluorescence grade analyzer |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103824137A (en) * | 2014-03-20 | 2014-05-28 | 北京信息科技大学 | Multi-condition fault prediction method for complex mechanical equipment |
CN104062968A (en) * | 2014-06-10 | 2014-09-24 | 华东理工大学 | Continuous chemical process fault detection method |
CN105004542A (en) * | 2015-07-15 | 2015-10-28 | 浙江中烟工业有限责任公司 | Online monitoring and fault diagnosing method for mixing and flavouring process of cigarette filament production based on principal component analysis |
-
2016
- 2016-04-26 CN CN201610263402.XA patent/CN105893700B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103824137A (en) * | 2014-03-20 | 2014-05-28 | 北京信息科技大学 | Multi-condition fault prediction method for complex mechanical equipment |
CN104062968A (en) * | 2014-06-10 | 2014-09-24 | 华东理工大学 | Continuous chemical process fault detection method |
CN105004542A (en) * | 2015-07-15 | 2015-10-28 | 浙江中烟工业有限责任公司 | Online monitoring and fault diagnosing method for mixing and flavouring process of cigarette filament production based on principal component analysis |
Non-Patent Citations (1)
Title |
---|
基于改进多尺度核主元分析的化工过程故障检测与诊断方法研究;许洁 等;《仪器仪表学报》;20100131;第31卷(第1期);第51-55页 |
Also Published As
Publication number | Publication date |
---|---|
CN105893700A (en) | 2016-08-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105893700B (en) | Based on the online fault detection and diagnosis technology of physics-big data mixed model Chemical Manufacture | |
CN106371427B (en) | Industrial process Fault Classification based on analytic hierarchy process (AHP) and fuzzy Fusion | |
CN101169623B (en) | Non-linear procedure fault identification method based on kernel principal component analysis contribution plot | |
CN106447205A (en) | Method for evaluating state of distribution automation terminal based on analytic hierarchy process | |
CN110738274A (en) | nuclear power device fault diagnosis method based on data driving | |
CN101446831B (en) | Decentralized process monitoring method | |
CN111610407B (en) | Cable aging state evaluation method and device based on naive Bayes | |
CN112036089A (en) | Coal mill fault early warning method based on DPC-MND and multivariate state estimation | |
CN104699077B (en) | A kind of failure variable partition method based on nested iterations Fei Sheer discriminant analyses | |
CN105955219A (en) | Distributed dynamic process fault detection method based on mutual information | |
CN101738991B (en) | Method for detecting out-of-specification quality of product and estimating actually measured value of product | |
CN108051709A (en) | Transformer state online evaluation analysis method based on artificial intelligence technology | |
CN105303331A (en) | Transformer repair risk decision-making method | |
CN104536439B (en) | A kind of method for diagnosing faults based on nested iterations Fei Sheer discriminant analysiss | |
CN109540212A (en) | A kind of cable shaft status assessment early warning system and method based on big data | |
CN109583520A (en) | A kind of state evaluating method of cloud model and genetic algorithm optimization support vector machines | |
CN109507972A (en) | Industrial processes fault monitoring method based on layer-stepping non-gaussian monitoring algorithm | |
CN112966990B (en) | Comprehensive state evaluation method for power transformation equipment | |
CN101872182A (en) | Batch process monitoring method based on recursive non-linear partial least square | |
CN108536128A (en) | A kind of machine learning fault diagnosis system of parameter optimization | |
CN108345284A (en) | A kind of quality dependent failure detection method becoming gauge block based on two | |
CN108764305A (en) | A kind of improved colony intelligence machine learning fault diagnosis system | |
CN110009126B (en) | Online alarm analysis method based on fusion of PLS model and PCA contribution degree | |
CN108919755A (en) | A kind of distributed fault detection method based on muti-piece Nonlinear and crossing relational model | |
CN115112978A (en) | Fault detection method for insulating ring main unit of power system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
CB02 | Change of applicant information | ||
CB02 | Change of applicant information |
Address after: Nanjing high tech Zone, Nanjing Software Park (West) unity Road, room 100, No. 227, room 210032 Applicant after: Lu Xinjian Address before: 210036 Gulou District, Jiangsu, Nanjing Blue Sky Park 37-201 Applicant before: Lu Xinjian |
|
GR01 | Patent grant | ||
GR01 | Patent grant |