CN102004486A

CN102004486A - Hybrid fault diagnosis method based on qualitative signed directed graph in petrochemical process

Info

Publication number: CN102004486A
Application number: CN 201010291934
Authority: CN
Inventors: 牟善军; 张卫华; 姜春明; 王春利; 李传坤; 姜巍巍
Original assignee: China Petroleum and Chemical Corp; Sinopec Qingdao Safety Engineering Institute
Current assignee: China Petroleum and Chemical Corp; Sinopec Qingdao Safety Engineering Institute
Priority date: 2010-09-26
Filing date: 2010-09-26
Publication date: 2011-04-06
Anticipated expiration: 2030-09-26
Also published as: CN102004486B

Abstract

The invention relates to a hybrid fault diagnosis method based on qualitative signed directed graph (SDG) in the petrochemical process. The method comprises the following steps: the characteristic data of the key nodes in a fault state of the process flow are extracted to store in a expert knowledge base; the fuzzy logic and the principle component analysis (PCA) are combined to obtain the hybrid algorithm of fault diagnosis based on SDG; for a built SDG model, the hazard and operability analysis (HAZOP) is performed through automated reasoning; and the analysis result is stored in a fault knowledge base in the expert knowledge modes of fault sign, fault cause, propagation path, unfavourable result and treatment measure. A hybrid expert knowledge system is mainly composed of an expert system and HAZOP analysis results. By adopting the method of the invention, the problems of the fault diagnosis technology in fault detection and diagnosis speed, diagnosis completeness and accuracy, diagnosis resolution and robustness and the like can be solved.

Description

In the petrochemical process based on the fault hybrid diagnosis method of qualitative SDG

Technical field

The present invention relates to the method for diagnosing faults in the petrochemical process, particularly a kind of based on the mixing fuzzy logic of qualitative SDG and the method for diagnosing faults of pivot analysis.

Background technology

Petrochemical production process is owing to reasons such as sensor drift, equipment failure, technological fluctuation or operating mistakes, cause often occurring in the production run unusual service condition state, gently then influence product quality, production scheduling plan, heavy then industrial accident can occur, cause casualties and enormous economic loss.

How from the production data of magnanimity, to excavate effective information, in time identify the unusual service condition of existence, find out the reason that causes unusual service condition to produce, predict that this unusual service condition develops down issuable consequence, relative measures is proposed, carrying out effective unusual service condition management, avoid producing serious consequence, is the important step of guaranteeing the enterprise security operation.

The core technology of said process is the fault diagnosis to process.The method that can be used for industrial system Fault Identification and diagnosis at present mainly is divided into two big classes, and a class is based on the method for model, the another kind of method that is based on historical data.Method for diagnosing faults based on model can be divided into quantitative model method and qualutative model method again.

The quantitative model method needs the dynamic perfromance of accurate test process, owing to be difficult to the accurate quantification dynamic model of acquisition process object, the method practicality is restricted, and obtains paying attention to and development based on the method for diagnosing faults of qualitative mathematics model.In the Fault Identification and diagnostic method based on the qualitative mathematics model, Graph-theoretical Approach is the most with practical value a kind of, and wherein signed directed graph (SDG, Signed Directed Graph) method prospect is very good.

To the known priori of procedures system, these class methods can not be left historical data, and emerging fault is not had recognition capability based on the necessary foundation of the method for diagnosing faults of historical data.Method for diagnosing faults based on historical data can be divided into quantivative approach and quilitative method again.The most widely used expert system that is based on rule is called the superficial knowledge expert system again in the quilitative method.

Divide statistical method and non-statistical method two big classes again based on the fault diagnosis quantivative approach of historical data.Two class methods all are that feature extracting methods is carried out in the sampling of real time data.Statistical method comprises pivot analysis method (PCA), partial least square method (PLS) etc.Based on the non-statistical method of the fault diagnosis quantivative approach of historical data neural network method promptly commonly used.

At present, obtain paying attention to and development based on semiquantitative method.So-called sxemiquantitative is in conjunction with random theory, fuzzy set theory, method of weighting etc., adds quantitative information in qualitative method for diagnosing faults, and quantivative approach and quilitative method are had complementary advantages, for example fuzzy SDG method, SDG-PCA method, SDG-PLS method etc.

Contrast above-mentioned method for diagnosing faults, because the quantitative model method is based on mathematical models, so have reasonable early stage perception and resolution, but it is very poor to the robustness of noise and spurious signal, and the accurate model of complication system is difficult to obtain, when technological process changed, adaptive faculty was also very poor.Statistical classification and neural network method are easy to use, and be also better to the robustness of noise, yet powerless for the diagnosis of new fault or unknown failure.Show that with the U.S. and Japanese correlative study can be in petrochemical process real practical method is probably also no more than RBES, PCA method and SDG method.

Because what Fault Identification and diagnosis were faced is changeable, complicated procedures system, at present, does not still have a kind of method can generally be applicable to the needs of the different fault diagnosis of various industry.Therefore it is feasible developing direction that the parallel or mutual fusion of several are practical method is learnt from other's strong points to offset one's weaknesses.At the characteristics of petrochemical process, the present invention is a goal in research with the data processing and the fault diagnosis of complication system, has proposed the malfunction monitoring of suitable petrochemical industry and the key algorithm of diagnosis

Summary of the invention

The present invention is in order to solve the problems of the technologies described above, fault hybrid diagnosis method based on qualitative SDG is provided in a kind of petrochemical process, by proposing a cover, solve the problem of fault diagnosis technology at speed, diagnosis completeness and the aspects such as accuracy, diagnosis resolution and robustness of fault detection and diagnosis based on practical fault diagnosis method qualitative SDG, that merge multiple quantitative fault diagnosis technology; In conjunction with the SDG-HAZOP technology, model and the algorithm that is adopted carried out static state and dynamic check, guarantee fault diagnosis model and algorithm accuracy and completeness, and with the HAZOP analysis result as expert knowledge library, be used for on-line fault diagnosis.

Described fault hybrid diagnosis method has been set up one three layers level diagnostic model:

1) ground floor is an expert system module

Described expert system module is connected with actual flow process in the production run, gathers the real time data from the production scene, the characteristic under the key node of extraction process flow process is nonserviceabled deposits expert knowledge library in.During monitoring in real time, if the state of these nodes is just dropped into the state that knowledge base defines, that just obtains conclusion: entered certain malfunction at present, reason and consequence can be determined.The unusual service condition management system can directly obtain the improper operating mode conclusion of monitored technology like this.This moment, software systems need not enter the reasoning algorithm of back, had significantly reduced the inference time of system.

2) second layer is a comprehensive diagnosis module

Utilize the SDG method to carry out fault diagnosis, diagnostic result completeness height, but because the polysemy of qualitative reasoning causes resolution lower.In conjunction with PCA, fuzzy logic and pivot analysis, obtain hybrid algorithm based on the fault diagnosis of SDG, after entering hybrid algorithm, at first real time data is monitored with the PCA method, when monitoring process generation unusual fluctuations, then utilize the whole bag of tricks to try to achieve deviation point and carry out reasoning for SDG: 1) utilize PCA that real time data is calculated, the residual error method that obtains each point is in the hope of deviation point; 2) real time data is according to the dynamic threshold of SDG model, tries to achieve to reach the point that necessarily departs from.Subsequently, utilization SDG algorithm carries out fault reasoning, obtain compatible path, it is the fault propagation path, successively send into and adopt the SDG inference engine of fuzzy logic to carry out fault diagnosis, compatible degree and sensitivity information according to each travel path sort, and in conjunction with mixed expert knowledge library system, obtain reason, consequence and the treatment measures etc. of fault;

For the SDG model of setting up, automated reasoning carries out HAZOP, adopt key variables " to draw partially ", there is analysis result in the fault knowledge storehouse in search fault propagation path with fault disease million, failure cause, travel path and negative consequence and the treatment measures form with expertise.In practical situation, go to search in the knowledge base processing scheme with failure path among the result of SDG fault diagnosis as keyword, can save operation time greatly, improve arithmetic speed.

Like this, based on the SDG method, merge multiple diagnostic method, it is additional to cooperatively interact, and performance advantage has separately formed the fault diagnosis hybrid algorithm based on SDG, has improved diagnosis efficiency.

3) the 3rd layer is to mix expert knowledge system

This mixing expert knowledge system mainly is made of expert system and HAZOP analysis result.

In the comprehensive diagnosis module of the second layer, the field data that collects is at first delivered to the PCA algoritic module, be used to diagnose the state of the art of whole device or unit whether normal, successively send into the SDG inference engine subsequently and carry out fault diagnosis, the SDG node of this moment and the threshold value of path adopt fuzzy logic algorithm, with a node in the SDG model and a line state obfuscation.

Described PCA algoritic module may further comprise the steps:

A. with the PCA method real time data is monitored, gather real time data, set up principal component model;

B. calculate residual error;

C. set up the PCA-SDG model;

D. carry out bidirection reasoning for the SDG model of assignment, obtain the fault propagation path.

General algorithm based on the SDG fault diagnosis of fuzzy logic:

Fuzzy the fuzzy of threshold value bound that comprise of a.SDG node, gather DCS and go up the high newspaper of field instrument, high newspaper, low newspaper, low count off certificate, obtain the threshold scaling factor of each node among the SDG according to test, based on described data, multiply by the threshold scaling factor, the scope of alarm limit is amplified or dwindle certain multiple, thereby obtain being applicable to the fuzzy threshold value of SDG reasoning;

B.SDG node fuzzy also comprises real-time measurement values fuzzy to threshold value, represents by introducing degree of membership, at a time, obtains the actual measured value of each node in the SDG model, and calculates its degree of membership with respect to fuzzy threshold value;

C. the steady-state gain between system's cause and effect variable is defined as the sensitivity of SDG branch road, by artificial setting.

D. from a certain node that departs from, carry out forward inference and backward reasoning, find out all compatible paths, the sensitivity of the degree of membership of each node in the compatible path of each bar and each branch road is multiplied each other respectively, draw the compatible degree and the sensitivity of the compatible path of whole piece;

E. consider the different influence of interstitial content of compatible path, the compatible degree of every compatible path is got geometrical mean according to the node number, sensitivity is got geometrical mean according to a way;

F. take all factors into consideration the compatible degree and the level of sensitivity of compatible path, carry out priority queueing, and explain reason and the dangerous travel path that causes current warning automatically;

G. repeat above step a-f every a selected time interval, so that real-time follow-up field failure situation.

The function value algorithm of described degree of membership has following two kinds:

A. triangular form subordinate function

Threshold value is thought of as a triangle, widens one section in the position of former bound B and A and depart from, the absolute value that departs from is made as D, the membership function μ of upper limit threshold _i(x) by formula (1) expression, the membership function μ of lower threshold _i(x) by formula (2) expression,

μ_{i} (x) = \{\begin{matrix} 0 & x < B - D \\ \frac{x - B + D}{D} & B - D \leq x < B \\ 1 & x &GreaterEqual; B \end{matrix} - - - (1)

μ_{i} (x) = \{\begin{matrix} 1 & x \leq A \\ \frac{A + D - x}{D} & A < x \leq A + D \\ 0 & x > A + D \end{matrix} - - - (2)

B. quadratic distribution type subordinate function

Threshold value is thought of as a curve, then the membership function μ of upper limit threshold _i(x) by formula (3) expression, the membership function μ of lower threshold _i(x) by formula (4) expression, wherein, B and A represent the position of former bound,

μ_{i} (x) = \{\begin{matrix} 0 & x \leq A \\ 2 {(\frac{x - A}{B - A})}^{2} & A < x < \frac{A + B}{2} \\ 1 - 2 {(\frac{x - B}{B - A})}^{2} & \frac{A + B}{2} \leq x \leq B \\ 1 & x > B \end{matrix} - - - (3)

μ_{i} (x) = \{\begin{matrix} 1 & x \leq A \\ 1 - 2 {(\frac{x - A}{B - A})}^{2} & A < x < \frac{A + B}{2} \\ 2 {(\frac{x - B}{B - A})}^{2} & \frac{A + B}{2} \leq x \leq B \\ 0 & x > B \end{matrix} - - - (4)

With the absolute value of degree of membership value as the node compatibility.

The fuzzy sensitivity by branch road of a line state of SDG model realizes that for each the cause and effect branch road in the SDG model, its sensitivity definition is:

μ _BA(ΔB/ΔA)＝f(ΔB/ΔA) (5)

Wherein, Δ B---the relative departure of consequence node;

Δ A---the relative departure of reason node;

Adopt SDG-HAZOP that model and the algorithm of SDG are carried out static state and dynamic check.

But because the complicacy of actual production is got " height ", " higher ", " on the low side ", " low " four kinds of qualitative states with the sensitivity of branch road, by artificial setting, in computing module inner respectively corresponding " 0.8 ", " 0.6 ", " 0.4 ", " 0.2 ".This is because for two interactional variablees in the production run, its influence degree is constant substantially.

The expert system that it is kernel that described hybrid expert system comprises with general artificial intelligence software Clips and based on the expert knowledge library of HAZOP analysis result.

Utilize the good advantage of SDG completeness, can infer fault source point and fault propagation path; Brought into play the advantage of PCA, can handle mass data efficiently as a kind of statistical method, and to the monitoring reaction sensitivity of fault; In conjunction with the dynamic threshold of fuzzy logic, the sensitivity of branch road, The reasoning results is carried out effective quantification.Because adopt the SDG-HAZOP technology that model and the algorithm that is adopted carried out static state and dynamic check, the completeness and the accuracy of model all improve a lot.Therefore, merge multiple semiquantitative fault diagnosis technology can surmount a kind of method for diagnosing faults of existing simple use on problems such as the resolution of the completeness of the speed of fault detection and diagnosis, fault diagnosis and accuracy, diagnosis and robustness technology based on qualitative SDG.

Description of drawings

Accompanying drawing 1 is the synoptic diagram based on the diagnosis hierarchical model of the different fault diagnosis method fusion of SDG

Accompanying drawing 2 is the synoptic diagram of the fuzzy threshold value of SDG

Accompanying drawing 3 is a simple SDG model

Accompanying drawing 4 is the synoptic diagram based on the algorithm principle of work of the different fault diagnosis method fusion of SDG

Embodiment

The present invention proposes one and set up level diagnostic model quick, that expression is easy, this hierarchical model constitutes by three layers, as shown in Figure 1.

1) ground floor is an expert system module

Sign under the key node of extraction process flow process (drawing according to the manipulation experience) is nonserviceabled deposits expert knowledge library in.During monitoring in real time, if the state of these nodes is just dropped into the state that knowledge base defines, that just obtains conclusion: entered certain malfunction at present, reason and consequence can be determined.The unusual service condition management system can directly obtain the improper operating mode conclusion of monitored technology like this.This moment, software systems need not enter the reasoning algorithm of back, had significantly reduced the inference time of system.

2) second layer is the comprehensive diagnosis module of technology such as utilization SDG, fuzzy logic, PCA:

Utilize the SDG method to carry out fault diagnosis, diagnostic result completeness height, but because the polysemy of qualitative reasoning causes resolution lower.In conjunction with the relative merits of each method such as PCA, fuzzy logic, obtain hybrid algorithm based on the fault diagnosis of SDG.Its essence is, after real time data enters hybrid algorithm, at first process data is monitored with the PCA method, when monitoring process generation unusual fluctuations, then utilize the whole bag of tricks to try to achieve deviation point and carry out reasoning for SDG: 1) utilize PCA that real time data is calculated, the residual error method that obtains each point is in the hope of deviation point; 2) real time data is according to the dynamic threshold of SDG model, tries to achieve to reach the point that necessarily departs from.

Use the SDG algorithm to carry out fault reasoning subsequently, obtain compatible path, be i.e. the fault propagation path.Compatible degree and sensitivity information according to each travel path sort, and in conjunction with mixed expert knowledge library system, obtain reason, consequence and the treatment measures etc. of fault.

For the SDG model of setting up, it is carried out the computing machine automated reasoning carry out HAZOP, adopt key variables " to draw partially ", there is analysis result in the fault knowledge storehouse in search fault propagation path with fault disease million, failure cause, travel path and negative consequence and the treatment measures form with expertise.In practical situation, go to search in the knowledge base processing scheme with failure path among the result of SDG fault diagnosis as keyword, can save operation time greatly, improve arithmetic speed.

3) the 3rd layer is to mix expert knowledge system

It below is algorithm further instruction to merging

One, the algorithm that merges based on the different fault diagnosis method of SDG

1. adopt the SDG algorithm of fuzzy logic

Fuzzy logic is introduced the SDG model, mainly is to utilize fuzzy algorithm with a node in the SDG model and a line state obfuscation, carries out semiquantitative fault diagnosis to utilize quantitative information.

1.1SDG node is fuzzy

SDG node fuzzy comprises fuzzy and real-time measurement values fuzzy of threshold value bound.

1.1.1 the threshold value bound is fuzzy

In actual production, most variablees all have an opereating specification, and in this opereating specification, variable is all thought normally, so can be normal band of node definition to all variablees.Test shows, if directly use the threshold value of the threshold value of DCS as SDG, for some fault, some nodes may play the effect of fault propagation, but owing to do not reach the alarm threshold value of DCS, so in SDG, do not produce deviation, thereby compatible path reasoning is interrupted, and has caused the uncontinuity of compatible path during to these nodes.The DCS threshold value that employing is dwindled can address this problem.This paper proposes to utilize " the threshold scaling factor " to obtain being used for the node threshold value of SDG reasoning.Concrete grammar be based on DCS go up the high newspaper of field instrument, high newspaper, low newspaper, low count off according to, on their basis, multiply by a threshold scaling factor, with the scope amplification of alarm limit or dwindle certain multiple.This factor is got empirical value, and its big I is obtained by test.

As shown in Figure 2: DCS variable threshold zone is dwindled, form SDG node threshold value.Threshold value is divided into three zones: upper limit band, normal band, lower limit band.

1.1.2 instantaneous value is to the degree of membership of threshold value

SDG node instantaneous value is realized by the method for introducing degree of membership the fuzzy of threshold value.When nodal value drops on normal band, degree of membership μ _i(x) value is 0.Exceed normal band, μ _i(x) be ± 1.The threshold value degree of membership μ of certain node i _i(x) value is limited in the closed interval (0,1), and the size of its value has been represented the subjection degree of the currency x of node variable i to threshold limits.Have only as degree of membership μ _i(x)=1 o'clock, x reaches threshold limits.Work as μ _i(x)=0 o'clock, x does not reach threshold limits.For example work as μ _i(x)=0.6 o'clock, x has reached 60% of threshold limits.

The threshold value subordinate function value algorithm that this algorithm is used has following two kinds:

1) triangular form subordinate function

If threshold value is thought of as a triangle, to widen one section in the position of former bound B and A and depart from, the absolute value that departs from is made as D.Upper limit threshold membership function μ then _i(x) describe by formula (1), shown in the following trigonometric expression subordinate function (a).The membership function μ of lower threshold in like manner _i(x) describe by formula (2), shown in the following trigonometric expression subordinate function (b).

μ_{i} (x) = \{\begin{matrix} 0 & x < B - D \\ \frac{x - B + D}{2} & B - D \leq x < B \\ 1 & x &GreaterEqual; B \end{matrix} - - - (1)

μ_{i} (x) = \{\begin{matrix} 1 & x \leq A \\ \frac{A + D - x}{D} & A < x \leq A + D \\ 0 & x > A + D \end{matrix} - - - (2)

The trigonometric expression subordinate function:

2) quadratic distribution type subordinate function

Subordinate function based on triangle is assumed to linear, dull variation with departing from of node.Common rule may be non-linear in real process, therefore under the prerequisite of actual measurement of having ready conditions, should consider that the type of membership function adapts to the Changing Pattern of real process in different range abilities as far as possible.If threshold value is thought of as a curve, the membership function μ of upper limit threshold then _i(x) describe by formula (3), shown in the following quadratic distribution formula subordinate function (a).The membership function μ of lower threshold in like manner _i(x) describe by formula (4), shown in the following quadratic distribution formula subordinate function (b).

μ_{i} (x) = \{\begin{matrix} 0 & x \leq A \\ 2 {(\frac{x - A}{B - A})}^{2} & A < x < \frac{A + B}{2} \\ 1 - 2 {(\frac{x - B}{B - A})}^{2} & \frac{A + B}{2} \leq x \leq B \\ 1 & x > B \end{matrix} - - - (3)

μ_{i} (x) = \{\begin{matrix} 1 & x \leq A \\ 1 - 2 {(\frac{x - A}{B - A})}^{2} & A < x < \frac{A + B}{2} \\ 2 {(\frac{x - B}{B - A})}^{2} & \frac{A + B}{2} \leq x \leq B \\ 0 & x > B \end{matrix} - - - (4)

Quadratic distribution formula subordinate function

After adopting above-mentioned fuzzy membership, need the decision rule of compatible path is changed slightly based on triangle.Introduce the notion of node compatibility, promptly the node compatibility equals the absolute value (a certain value between 0 to 1) of the degree of membership value of current reason node, for compatible fully, is incompatible when compatible degree equals 0 when compatible degree equals 1.Need to introduce the notion of path compatibility simultaneously, as long as each node compatible degree of a certain compatible path is all greater than 0, just as the candidate of compatible path.

Path compatible degree the maximum is that confidence level is the highest in the compatible path of many candidates that reasoning obtains.After adopting said method, threshold range can be widened, the requirement of sensitivity can be taken into account again.

1.2SDG branch road sensitivity is fuzzy

Steady-state gain between system's cause and effect variable (branch road sensitivity) is defined as the fuzzy set of SDG.

That is: for each the cause and effect branch road in the SDG model

Its sensitivity definition is:

μ _BA(ΔB/ΔA)＝f(ΔB/ΔA) (5)

In the formula (5)

Δ B---the relative departure of consequence node

Δ B=(consequence variable measured value-consequence nominal situation value)/(the consequence variable upper limit-consequence variable lower limit)

Δ A---the relative departure of reason node

Δ A=(causal variable measured value-reason nominal situation value)/(the causal variable upper limit-causal variable lower limit)

For the simple SDG model of shown in the accompanying drawing 3, at a time, because the increase of B causes the increase of C, this moment, A also increased, but variation is very little, and A is very little to the influence of C.If employing formula (5) is calculated μ, then calculate μ _CAIn time, can occur than mistake, because A is very little to the influence of C.Can get the sensitivity of branch road " height ", " higher ", " on the low side ", " low " four kinds of qualitative states this moment, by artificial setting, in computing module inner respectively corresponding " 0.8 ", " 0.6 ", " 0.4 ", " 0.2 ".This is because for two interactional variablees in the production run, its influence degree is constant substantially.

In sum, can obtain general algorithm based on the SDG fault diagnosis of fuzzy logic:

(1) warning with each variable among the DCS is limited to the basis, obtains the threshold scaling factor of each node among the SDG according to test, thereby obtains being applicable to the fuzzy threshold value of SDG reasoning;

(2) at a time, obtain the actual measured value of each node in the SDG model, and calculate its degree of membership with respect to fuzzy threshold value;

(3) from a certain node that departs from, carry out forward inference and backward reasoning, find out all compatible paths, the sensitivity of the degree of membership of each node in the compatible path of each bar and each branch road is multiplied each other respectively, draw the compatible degree and the sensitivity of the compatible path of whole piece;

(4) the different influence of interstitial content of the compatible path of consideration is got geometrical mean to the compatible degree of every compatible path according to the node number, and sensitivity is got geometrical mean according to a way.

(5) take all factors into consideration the compatible degree and the level of sensitivity of compatible path, carry out priority queueing, and explain reason and the dangerous travel path that causes current warning automatically;

(6) repeat above step every a selected time interval, so that real-time follow-up field failure situation.

By the combination algorithm of above-mentioned fuzzy logic and SDG as can be known, in SDG, introduce fuzzy logic, make threshold value become an interval, can effectively solve the changeable problem of threshold value by fixed value.Simultaneously, on the node of SDG, introduce compatible degree, in the branch road of SDG, introduce sensitivity, for traditional qualitative SDG has increased quantitative information.Like this, both utilized the good advantage of traditional SDG completeness, and can reject the part deceptive information by quantitative mode again, thereby improve the resolution of carrying out fault diagnosis based on the SDG model greatly.

If adopt fuzzy logic and SDG combination merely, because each all wants the rate of change of computing node and branch gain constantly, calculated amount increases greatly, though improved resolution, has influenced the speed of fault diagnosis.Introduce pca method, can effectively alleviate this problem.

2. the method for diagnosing faults that combines with SDG of pivot analysis

(Principle Component Analysis PCA), as a kind of data-driven method based on signal Processing, has been widely used in the industrial process fault diagnosis pca method.But because the limitation of method itself, the PCA method can not be pointed out the source of trouble accurately, so the combining of PCA and SDG method, and has great significance.

Utilize the PCA residual error to combine to carry out fault diagnosis and mainly be divided into following a few step with SDG:

1) image data is set up principal component model.

2) calculate residual error.Utilize and set up good principal component model, calculate measured value x _NewSub matrix: t _New=x _NewP.Predicted value:

Residual error:

3) set up the PCA-SDG model.According to the residual error that previous step obtains, set threshold residual value, and set up the PCA-SDG model.When finding that residual error surpasses threshold residual value, it is unusual to show that production run produces, and residual error surpasses its threshold value and is positive variable, is designated as "+" on the SDG of correspondence node; Surpass threshold value and variable, on the SDG of correspondence node, be designated as "-" for bearing; Otherwise be designated as " 0 ".The PCA-SDG model has just been set up like this.

4) carry out bidirection reasoning for the SDG model of assignment, obtain the fault propagation path.

In the PCA-SDG method, determining of threshold residual value is very important.Because before carrying out fault diagnosis, all need to determine high and low threshold value for each measurand with SDG.Threshold value upper and lower limit scope is wide, and node variable can be seen normal condition at the threshold value upper and lower limit as with interior its state, and reasoning " engine " can't search out the path of fault propagation, can cause the sensitivity of fault diagnosis and predictability poor.Otherwise threshold value upper and lower limit narrow limits can cause sensitivity too high or pre-alarm is too early, and real process also is in safe range, and promptly fault does not take place.Being set in the PCA-SDG method for diagnosing faults of threshold value is a key, and when this problem of processing, the someone proposes high threshold is existed

With

Between, low threshold value is located at

With

Between (wherein, Q _αFixing computing formula is arranged), also setting threshold rule of thumb.Generally, threshold residual value can be determined by experience.

3. based on the fusion fault diagnosis of SDG

SDG and fuzzy logic combination have been introduced in the front, have solved the fixing problem of threshold value of node, have realized dynamical thresholdization, have improved resolution; SDG and pivot analysis combination can improve inference speed fast.But, still can not solve dynamic threshold problem and quick reasoning problems simultaneously iff combination in twos at all.

Whether as shown in Figure 3, be the fault diagnosis algorithm principle of work, the field data that collects is at first delivered to the PCA algoritic module, be used to diagnose the state of the art of whole device or unit normal, if T ²Value determines then that greater than preset threshold state of the art is unusual, finds out the contribution rate of major parameter, according to the size of contribution rate, as priority, successively sends into the SDG inference engine and carries out fault diagnosis.The node of the SDG of this moment and the threshold value of path adopt fuzzy logic algorithm.Three kinds of method combinations had both solved because interstitial content increases, and the phenomenon that inference speed is slow has solved the problem of resolution again simultaneously.

This based on the SDG method, merge multiple diagnostic method, it is additional to cooperatively interact, and performance advantage has separately formed the fault diagnosis blending algorithm based on SDG, has improved diagnosis efficiency.

Two, SDG model testing

1.SDG fault diagnosis model is unofficially checked

The unofficial check of SDG fault diagnosis model relates generally to the technological principle examination and based on the examination of experience and the content of model simplification three aspects.

Technological principle examination and can be taken on by the other personnel that are familiar with the SDG modeling based on the audit crew of experience, the person's that helps overcoming the model development " blind spot " that thinking inertia caused finished check under the person's that is preferably in the model development certainly the cooperation.Should be during check with reference to the process chart at the band reference mark that is modeled process, relevant technology, equipment and automatic control design data, all associated production data, historical record curve, technological operation explanation etc.The reviewer should check the cause and effect influence relation of SDG model one by one from technology ultimate principle (momentum transfer principle, material transfer principle, heat transferred principle, reaction kinetics principle).This process is equivalent to the repetition modeling process, and each step conclusion of modeling process is carried out the principle examination.

Adopting the SDG technology to carry out the modal problem of fault diagnosis is that the resolution of diagnosis is low, and promptly diagnostic result provides too much less important conclusion or invalid conclusion.Cause the low problem more complicated of SDG fault diagnosis resolution, relate to multiple factor, for example comprise more unobservable node in the SDG model, inference method is improper, the SDG model structure is unreasonable, the threshold setting scope is incorrect, failure symptom is ambiguous etc.Wherein the problem that the unreasonable fault diagnosis resolution that causes of SDG model structure is low can obtain alleviation to a certain degree by the SDG Model Simplification Method based on experience.SDG model simplification principle is as follows.

Following principle not necessarily solves all problems, can improve SDG modeling quality but take in.

1) merging of observable variable node and simplification principle

This node and interference source and fault propagation concern that little person can cancel.

2) merging of unobservable variable node and simplification principle

(1) state the unknown of unobservable variable node should be cancelled in diagnosis application as far as possible.

(2) potential root node, the source of trouble for example, though unobservable, if cancellation may change the structure of correct SDG, then can not arbitrarily cancel.

(3) for the unobservable variable node in centre, must be prudent when many branch roads pass through.For example, cancel the dangerous travel path that this node can be drawn the compatible path of impossible puppet in the reality or disappeared and should not cancel.

3) merging that has multiple branch circuit between two nodes with simplify principle

Find out dominant branch road, the branch road of the non-ascendancy of cancellation.

When (1) being used for fault diagnosis, can be according to the branch road of field data or experience judgement ascendancy.

(2) if the negative feedback branch road belongs to non-Control and Feedback variable (for example from weigh phenomenon), and when being not enough to compensate or offset the propagation of interference, this branch road can be cancelled.

(3) if can obtain related data, satisfy the condition of semi-quantitative analysis, utilize SDG to analyze and find out dominant branch road, the branch road of the non-ascendancy of cancellation.

(5) rationally revise the SDG structure, reduce this node associated branch.

4) introducing of failure factor is added node and branch road by mechanism and historical accident experience that fault takes place.

5) monotonize of non-dull influence; The piece-wise linearization of nonlinear relationship; If non-linear rule is a monotone variation, can regards linear monotonic as and change.

6) condition branch road

During fault diagnosis, can closed Control and Feedback branch road automatically.

When 7) not considering certain fault or danger, the interdependent node or the branch road of cancellation and this variable.

2.SDG fault diagnosis model static check

The fundamental purpose of SDG fault diagnosis model static check is whether check SDG model structure is reasonable.The rational content of SDG model structure comprises: whether node is selected reasonable; Whether cause-effect relationship is described reasonable; Deduce the rationality of judging the SDG model structure by the mathematics that influences equation; Obtain the path of hazard propagation through the SDG reasoning after, analyze the SDG model from The reasoning results and can reach design idea etc.

1) cause-effect relationship is to checking

By checking the cause-effect relationship in the SDG model right one by one, judge whether the cause-effect relationship of all expression is realistic.When can't off-line judging, can adopt the simulation calculation test, can also be when having ready conditions at thread test, but must guarantee not disturb ordinary production.Preferably carry out the nonlinear characteristic test in the checking process, verifying in the cause-effect relationship is to have non-linear effects, is which kind of non-linear effects relation.

2) influence the equation reasonalbeness check

The SDG model can directly be mapped as influences equation, the Algebraic Equation set or the differential equation also can be converted into the qualitative equation that influences, therefore can use relevant mathematical method inspection to influence the reasonable structure of equation, algebraic equation or the differential equation, thus the rationality of inference SDG model structure.For example quantitatively the differential equation can be converted into the qualitative equation that influences by asking local derviation, checks the correctness of local derviation one by one, can guarantee to influence the correctness of equation.

3) check based on the SDG fault diagnosis model of SDG-HAZOP

Adopt the SDG-HAZOP platform to carry out following several check to the static characteristics of SDG model:

(1) operating point accessibility check

In this step, adopt backward inference the departing from of automated reasoning checked operation point in the SDG model by computing machine.The purpose of this step is to check the recognition capability of SDG model to maloperation.

(2) negative consequence (fault) accessibility check

In this step, adopt forward reasoning automated reasoning check consequence in the SDG model whether to take place by computing machine.The purpose of this step is to check the recognition capability of SDG model to failure effect.

(3) story of a play or opera associate feature of node check

Definite definition should be the dangerous story of a play or opera associate feature check of pilot process variable node, promptly checks each pilot process variable node to have how many dangerous story of a play or opera and passes through.The purpose of this step is to select the reasoning starting point of best SDG fault diagnosis.

(4) may observe node optimization configuration check

Every intermediate node with strong dangerous story of a play or opera association should be set at the may observe node.Adopt the SDG method to carry out fault diagnosis, have only the optimization allocation that at first solves the may observe node, just might fundamentally guarantee the completeness and the accuracy of SDG fault diagnosis.

In above-mentioned SDG model static characteristics checking procedure, check of operating point accessibility and the check of negative consequence (fault) accessibility are the most important.

Three, hybrid expert system

Hybrid expert system mainly is made of two parts, and the one, with the general artificial intelligence software Clips expert system that is kernel, the one, based on the expert knowledge library of HAZOP analysis result.The two complements each other, and as hybrid expert system, The reasoning results is provided proper explanations.

Claims

In the petrochemical process based on the fault hybrid diagnosis method of qualitative SDG, at the Central Control Room configuration server, described server links to each other by LAN (Local Area Network) with actual flow process in the production run, collection is from the real time data of production scene, and be connected with client by public network, this method has been set up one three layers level diagnostic model:

1) ground floor is an expert system module

Characteristic under the key node of extraction process flow process is nonserviceabled deposits expert knowledge library in; During monitoring in real time, if the state of key node in the state that expert knowledge library defines, then can be reached a conclusion: entered certain malfunction, reason and consequence can be determined;

2) second layer is a comprehensive diagnosis module

In conjunction with PCA, fuzzy logic and pivot analysis, obtain hybrid algorithm based on the fault diagnosis of SDG, enter hybrid algorithm after, at first real time data is monitored with the PCA method, when monitoring unusual fluctuations, obtain deviation point and carry out fault reasoning for SDG;

Subsequently, utilization SDG algorithm carries out fault reasoning, obtains compatible path, it is the fault propagation path, successively send into subsequently and adopt the SDG inference engine of fuzzy logic to carry out fault diagnosis,, obtain reason, consequence and the treatment measures of fault in conjunction with mixed expert knowledge library system;

For the SDG model of setting up, automated reasoning carries out HAZOP, and there is analysis result in the fault knowledge storehouse in search fault propagation path with fault disease million, failure cause, travel path and negative consequence and the treatment measures form with expertise;

3) the 3rd layer is to mix expert knowledge system

This mixing expert knowledge system mainly is made of expert system and HAZOP analysis result.
2. fault hybrid diagnosis method as claimed in claim 1, it is characterized in that, the field data that collects is at first delivered to the PCA algoritic module, be used to diagnose the state of the art of whole device or unit whether normal, successively send into the SDG inference engine subsequently and carry out fault diagnosis, the SDG node of this moment and the threshold value of path adopt fuzzy logic algorithm, with a node in the SDG model and a line state obfuscation.
3. fault hybrid diagnosis method as claimed in claim 2 is characterized in that,

Described PCA algoritic module may further comprise the steps:

A. with the PCA method real time data is monitored, gather real time data, set up principal component model;

B. calculate residual error;

C. set up the PCA-SDG model;

D. carry out bidirection reasoning for the SDG model of assignment, obtain the fault propagation path.
4. fault hybrid diagnosis method as claimed in claim 2 is characterized in that,

Fuzzy the fuzzy of threshold value bound that comprise of a.SDG node, gather DCS and go up the high newspaper of field instrument, high newspaper, low newspaper, low count off certificate, obtain the threshold scaling factor of each node among the SDG according to test, based on described data, multiply by the threshold scaling factor, the scope of alarm limit is amplified or dwindle certain multiple, thereby obtain being applicable to the fuzzy threshold value of SDG reasoning;

B.SDG node fuzzy also comprises real-time measurement values fuzzy to threshold value, represents by introducing degree of membership, at a time, obtains the actual measured value of each node in the SDG model, and calculates its degree of membership with respect to fuzzy threshold value;

C. the steady-state gain between system's cause and effect variable is defined as the sensitivity of SDG branch road, by artificial setting.

D. from a certain node that departs from, carry out forward inference and backward reasoning, find out all compatible paths, the sensitivity of the degree of membership of each node in the compatible path of each bar and each branch road is multiplied each other respectively, draw the compatible degree and the sensitivity of the compatible path of whole piece;

E. consider the different influence of interstitial content of compatible path, the compatible degree of every compatible path is got geometrical mean according to the node number, sensitivity is got geometrical mean according to a way;

F. take all factors into consideration the compatible degree and the level of sensitivity of compatible path, carry out priority queueing, and explain reason and the dangerous travel path that causes current warning automatically;

G. repeat above step a-f every a selected time interval, so that real-time follow-up field failure situation.
5. fault hybrid diagnosis method as claimed in claim 4 is characterized in that, the function value algorithm of described degree of membership has following two kinds:

A. triangular form subordinate function

Threshold value is thought of as a triangle, widens one section in the position of former bound B and A and depart from, the absolute value that departs from is made as D, the membership function μ of upper limit threshold _i(x) by formula (1) expression, the membership function μ of lower threshold _i(x) by formula (2) expression,

$μ_{i} (x) = \{\begin{matrix} 0 & x < B - D \\ \frac{x - B + D}{D} & B - D \leq x < B \\ 1 & x &GreaterEqual; B \end{matrix} - - - (1)$

$μ_{i} (x) = \{\begin{matrix} 1 & x \leq A \\ \frac{A + D - x}{D} & A < x \leq A + D \\ 0 & x > A + D \end{matrix} - - - (2)$

B. quadratic distribution type subordinate function

Threshold value is thought of as a curve, then the membership function μ of upper limit threshold _i(x) by formula (3) expression, the membership function μ of lower threshold _i(x) by formula (4) expression, wherein, B and A represent the position of former bound,

$μ_{i} (x) = \{\begin{matrix} 0 & x \leq A \\ 2 {(\frac{x - A}{B - A})}^{2} & A < x < \frac{A + B}{2} \\ 1 - 2 {(\frac{x - B}{B - A})}^{2} & \frac{A + B}{2} \leq x \leq B \\ 1 & x > B \end{matrix} - - - (3)$

$μ_{i} (x) = \{\begin{matrix} 1 & x \leq A \\ 1 - 2 {(\frac{x - A}{B - A})}^{2} & A < x < \frac{A + B}{2} \\ 2 {(\frac{x - B}{B - A})}^{2} & \frac{A + B}{2} \leq x \leq B \\ 0 & x > B \end{matrix} - - - (4)$

With the absolute value of degree of membership value as the node compatibility.
6. fault hybrid diagnosis method as claimed in claim 4 is characterized in that, the fuzzy sensitivity by branch road of a line state of SDG model realizes that for each the cause and effect branch road in the SDG model, its sensitivity definition is:

μ _BA(ΔB/ΔA)＝f(ΔB/ΔA) (5)

Wherein, Δ B---the relative departure of consequence node;

Δ A---the relative departure of reason node;
7. as the arbitrary described fault hybrid diagnosis method of claim 1-5, it is characterized in that, adopt SDG-HAZOP that model and the algorithm of SDG are carried out static state and dynamic check.
8. as the arbitrary described fault hybrid diagnosis method of claim 1-5, it is characterized in that the expert system that it is kernel that described hybrid expert system comprises with general artificial intelligence software Clips and based on the expert knowledge library of HAZOP analysis result.