CN106372330A - Application of dynamic Bayesian network to intelligent diagnosis of mechanical equipment failure - Google Patents

Abstract

Application of a dynamic Bayesian network to intelligent diagnosis of a mechanical equipment failure belongs to the field of failure diagnosis. According to the application, the dynamic three-layer Bayesian network is applied to intelligent diagnosis of the mechanical equipment failure for the first time; and a possible failure and the failure probability of related mechanical equipment can be calculated and deduced by acquiring operating state information of the mechanical equipment and setting and controlling related relation among Bayesian network nodes and the probability. The method has the advantages of real-time property, continuity, high accurate rate and the like, and is suitable for failure diagnosis of various kinds of mechanical equipment.

Description

Dynamic application in mechanical equipment fault intelligent diagnostics for the type Bayesian network

Technical field

The invention belongs to equipment fault intelligent diagnostics field, it is related to fault diagnosis for plant equipment and in particular to one Plant the method that dynamic type Bayesian network is applied in intelligent fault diagnosis.

Background technology

In National Industrial production with national life, various big machineries play very important role, especially in stone The industry of the relation economic life line of the country such as oil, chemical industry, gas pipeline industry.Due to complex structure, vibration stimulus source is many, once going out Existing fault, gently then impact produces, heavy then fatal crass, therefore studies the method to these main equipment fault pre-alarmings, accomplish and When note abnormalities, become the focus of current research.

The domestic at present research to heavy mechanical equipment intelligent fault diagnosis is relatively fewer, mainly to prop up in current research Hold vector machine, based on neutral net etc..Although these methods achieve some achievements, also exist much not enough, such as, this A little methods, not well from failure mechanism, make the accuracy rate of early warning relatively low；Sometimes for larger data volume, and a lot Fault data sample shortage etc., never has too big progressive space therefore in early warning field.

In recent years, with the development of machine learning, particularly in terms of state of affairs estimation, Bayesian network is increasingly becoming intelligence Main tool in information processing, therefore Bayesian network is fused in intelligent fault diagnosis, have developed one kind and is based on pattra leaves The method of the equipment fault intelligent diagnostics of this network, the method is applied to the various faults of various main equipments, applied range, Accuracy rate is high.

Content of the invention

The purpose of the present invention is to overcome prior art shortcoming, Bayesian network model is combined with fault diagnosis, provides one Set is brand-new, intelligence, in real time, accurate equipment fault diagnosis method.The present invention is first by three layers of Bayes of dynamic type Network is used in plant equipment intelligent trouble diagnosis, by the information gathering of plant equipment running status, and Bayesian network section The setting of correlative connection and probability and control between point, to reach computational reasoning have the fault that associated machines are likely to occur and The probability that fault occurs.The method has a real-time, persistence, and high accuracy for examination is it is adaptable to the event of various plant equipment Barrier diagnosis.Bayesian network model is used in the intelligent fault diagnosis of large industrialized equipment by the present invention first, sets in monitoring Accomplish the early warning to equipment fault while standby running status, accuracy rate high it is adaptable to industrial every field.

The general idea of the present invention is by building the Bayesian network for fault diagnosis related to machine unit characteristic, defeated Enter various features parameter during unit equipment running status, through network reasoning, calculate what unit equipment was possible to occur Fault and corresponding probability of malfunction.

In order to realize the problems referred to above, the invention discloses a kind of Bayesian network is in mechanical equipment fault intelligent diagnostics Application, concretely comprises the following steps:

1) determine unit equipment fault characteristic, set up device-dependent Bayesian network integral frame, determine Bayes The number of plies of network intelligence fault diagnosis frame structure, meets unit demand.

2) unit equipment fault and fault signature are understood, the Bayesian network determining respectively and setting up for fault diagnosis is each Individual Internet.

3) according to unit equipment relevant information, required for determining and set up Bayesian network each node layer information to set up respectively Each node.

4) determine between adjacent net network layers, interdependent node contacts.

5) by consulting related unit handbook and documents and materials, and the fault data of the statistics passing generation of unit, according to Diagnosis industry experience determines the conditional probability between fault and fault signature, and determines that the reference under unit different characteristic is reported to the police Value.

6) determine and setting network layer in each node suppression probability and leakage probability.

1. make fault signature represent with x, represent i-th kind of fault with subscript i, i takes 1,2 ... ..., i.e. x_iRepresent i-th kind of event Barrier feature.Y is made to represent fault type, then each fault signature x_iPresence depositing of corresponding fault type y can be described Property；And the ability of this effect in this process, can be produced for having independence the reason other.But fault Feature x_iPresence it cannot be guaranteed that fault y is just certain occurs.By p_iCan determine other items of node s conditional probability table, i here =1,2 ... ..., i.e. p_i=p₁, p₂..., h is the set of s father node value, h⁺Subset for h, you can obtain formula (1)

$p(s=t/h)=1-{\∏}_{i:c_i\&element;h^{+}}(1-p_i)---\left(1\right)$

Wherein, p () represents the probability that in bracket, event occurs, p_i=p₁, p₂..., p_nFor each fault and fault signature it Between conditional probability, s represents fault signature x_iOne set, it is true that t represents value, h be s father node value set, c_iFor The father node of arbitrary node s, i=1,2,3 ... n, s=t/h represent that s is the conditional probability that h is during true value, h⁺Subset for h, should Subset is made up of the father node being worth for true value.In formula (1), latter half is to belong to h to all⁺I node, ask 1 to subtract Remove the product of its probability, this simplifies the determination of network condition probability parameter to a certain extent.

2. in order to avoid working as h⁺During for empty set, the situation of p (s=t/h)=0 is in addition it is also necessary to by all unknown combined factors For a factor, it is designated as c_l, it connects probability is p_l.Now calculate Probability p again_iAnd p_lIt is assumed that the father node of s is divided into two Individual: c_iAnd l_all, wherein l_allIt is except c_iThe summation of all factors, corresponding outward, and it connects probability is p_iAnd p_all, can recognize simultaneously For l_allIt is always t value.Had according to formula (1)

p(s/c_l)=p₁+p_all-p_ip_all(2)

$p(s/\stackrel{\&overbar;}{c_l})=p_{all}---\left(3\right)$

And then can obtain

$p_i=\frac{p(s/c_i)-p(s/\stackrel{\&overbar;}{c_l})}{1-p(s/\stackrel{\&overbar;}{c_l})}---\left(4\right)$

HereRepresent non-c₂Event, can calculate the connection Probability p of the father node of egress s by formula (4)₁, p₂..., that is, it has been calculated p_i, that is, suppression probability to be tried to achieve.

Obtaining Probability p_iAfterwards, p can be calculated further_l.

According to known existing formula (5)

$p(\stackrel{\&overbar;}{s}/c_i=f)=\frac{{\σ}_{h/c_{i}eh}-(1-p_l^1){\π}_{c_{j}eh}+(1-p_j){\π}_{c_{j}eh}+\left(c_j\right){\π}_{c_{j}eh}-\left(\stackrel{\&overbar;}{c_j}\right)}{p\left(\stackrel{\&overbar;}{c_l}\right)}---\left(5\right)$

Can be obtained byThe i.e. p of corresponding i-th node_l, here subscript l of p represent that this probability is above-mentioned p_l, subscript I represents that this probability is i-th, i=1,2,3 ....HereRepresent non-s event, f represents that value is false c_jRepresent the father node in s In be different from c_iThe node of sum, p_jRepresent and be different from p_iNode probability,Represent non-c_jEvent, h^-It is expressed as the subset of h, should Subset is made up of the father node being worth for falsity.

ObtainingAfterwards, in order to try to achieve final p_l, then p can be obtained with weighting method_l,

$p_l={\mathrm{\σ}}_i\[1-p\left(c_i\right)\]p_l^i---\left(6\right)$

Just obtain the leakage probability of each node in Internet.

7) pass through the input of unit equipment operating state signal, calculate through dynamic type Bayesian Network Inference, reasoning meter Calculation equipment unit it may happen that fault and its probability.

A first aspect of the present invention, discloses the Bayesian network knot for various large industry equipment intelligent fault diagnosis Structure feature；

A second aspect of the present invention, discloses the correlative connection between each node layer of three layers of Bayesian network；

A third aspect of the present invention, discloses the Bayesian network probability inference for fault diagnosis and computational methods；

A fourth aspect of the present invention, discloses and for the theory of Bayesian network to be used for physical device fault diagnosis concrete application Step.

Accompanying drawing content

Fig. 1 is intelligent diagnostics dynamic type Bayesian network construction step flow chart；

Fig. 2 is dynamic type bayesian network structure schematic diagram；

Fig. 3 is three-layer network node schematic diagram figure；

Fig. 4 is practice flow chart in overall failure diagnosis for the dynamic type Bayesian network；

Fig. 5 is unit vibration Acceleration pulse figure before certain petroleum chemical enterprise's unit connecting rod fault occurs；

Fig. 6 is Bayesian Network Inference result.

Specific embodiment

, below in conjunction with flow chart (Fig. 1) to the present invention taking certain petroleum chemical enterprise's large reciprocating compressor domestic as a example Specific flow process be described further.

1) determine unit equipment fault characteristic, set up device-dependent Bayesian network integral frame: in the present invention Set up three layers of Bayesian network intelligent trouble diagnosis frame structure.

2) unit equipment fault and fault signature are understood, the Bayesian network determining respectively and setting up for fault diagnosis is each Individual Internet: according to equipment fault and corresponding fault signature table, determine three networks of Bayesian network of intelligent diagnostics Layer.In this three layers of bayesian network structures, ground floor is unit information layer, and the node information being comprised was by this unit once Maintenance record, the regular maintenance record of abnormal operation, related components health status and scene at scene etc. provide, Cong Zhongti In taking, available probability reason, exports ground floor in the form of observer state node.This node layer possesses bayesian prior Probability.

The second layer is fault layer, and the various faults being likely to occur with this unit are for each node.Can by the information of ground floor Point to corresponding malfunctioning node in the second layer, this process comprises conditional probability.

Third layer be fault signature layer, using the unit being obtained by the use of suitable fault detection method off-note index as Output, obtains this layer of fault signature node.This layer corresponding fault signature node can be pointed to by the node of second layer fault layer. (Fig. 2)

3) according to unit equipment relevant information, determine and set up Bayesian network each node layer information: in diagnostic network In each layer, set up each required node respectively, as shown in figure (3).In the facility information layer of ground floor, set up and order Name " Literature Consult " node, is designated as node 1；Set up and name " historical failure data " node, be designated as node 2.This layer determines The source mode of diagnostic network |input paramete.In second layer fault layer, set up and name " piston fault " node, be designated as node 3；Set up and name " connecting rod fault " node, be designated as node 4.This layer determines the fault type of unit in diagnostic network.? In three layers of fault signature layer, set up and name " acceleration of vibration " node, be designated as node 5；Set up and name " moment of torsion " node, note For node 6；Set up and name " output " node, be designated as node 7.This layer determines the spy of unit fault in diagnostic network Levy.Third layer node definition is observer nodes, and second layer node definition is destination node.

4) determine between adjacent net network layers, interdependent node contacts: facility information layer and equipment fault layer, and equipment fault Between layer and equipment fault characteristic layer, contact between the node of foundation has: node 1 and node 3, node 1 and node 4, node 2 and section Point 3, node 2 and node 4, node 3 and node 5, node 3 and node 6, node 3 and node 7, node 4 and node 5, node 4 He Node 6, node 4 and node 7, as shown in Figure 3.

5) in whole Bayesian network ground floor, that passes through set up node introduces mode, consults related unit handbook And documents and materials, and the fault data of the statistics passing generation of unit, fault and fault signature are determined according to diagnosis industry experience Between conditional probability, and the reference alarming value under unit different characteristic is it may be assumed that determine piston fault and acceleration of vibration, torsion Conditional probability between square, output, the conditional probability between connecting rod fault and acceleration of vibration, moment of torsion, output, point Do not record as step 6) input numerical value；Determine the acceleration of vibration of unit body with reference to alarming value, as step 7) defeated Enter numerical value.

6) determine and setting network layer in each node suppression probability and leakage probability: determine and set Bayesian network network layers In, related necessary node probability.

1. constructing system Bayesian network model: by step 5), obtain piston fault and acceleration of vibration, moment of torsion, defeated Go out the conditional probability between power, it is assumed that x after the conditional probability between connecting rod fault and acceleration of vibration, moment of torsion, output For a certain node in fault signature layer, i.e. a certain fault, y is a certain node in fault type layer, and that is, a certain fault is special Levy, and the probability of y is not zero, i.e. p (y) ＞ 0.So when known fault y occurs really, then condition under a situation arises in y for the x is general Rate can be expressed as p (x | y), and

$p\left(x\right|y)=\frac{p\left(xy\right)}{p\left(y\right)}=\frac{p\left(x\right)p\left(y\right|x)}{p\left(y\right)}---\left(8\right)$

Here p (xy) expression joint probability, p (xy)=p (y) p (x | y)=p (x) p (y | x).

If setting the corresponding fault type of fault signature x have y₁, y₂... ..., y_nKind, it is designated as y_i, and p (y_i) ＞ 0, i= 1,2 ... n, then now the probability of fault x is:

$p\left(x\right)={\mathrm{\σ}}_{i=1}^{n}p\left(y_i\right)p\left(x\right|y_i)---\left(9\right)$

Then can be calculated under the probability scenarios that certain fault signature occurs by Bayesian formula, corresponding is a certain Plant the probability of possible breakdown type.

2. each fault signature x_iPresence the existence of fault type y can be described；And in this process, energy Enough produce the ability of this effect for having independence the reason other.But fault signature x_iPresence it cannot be guaranteed that fault Y is just certain to be occurred.By p₁, p₂..., p_nCan determine other items of node s conditional probability table, i.e. formula (10)

$p(s=t/h)=1-{\π}_{i:c_i\&element;h}+(1-p_i)---\left(10\right)$

Wherein, p () represents the probability that in bracket, event occurs, p_i=p₁, p₂..., p_nFor each fault and fault signature it Between conditional probability, s represents fault signature x_iOne set, it is true that t represents value, h be s father node value set, c_iFor The father node of arbitrary node s, i=1,2,3 ... n, s=t/h represent that s is the conditional probability that h is during true value, h⁺Subset for h, should Subset is made up of the father node being worth for true value, and in formula (10), latter half is to belong to h to all⁺I node, ask 1 to subtract Remove the product of its probability, this simplifies the determination of network condition probability parameter to a certain extent.

3. in order to avoid working as h⁺During for empty set, the situation of p (s-t/h)=0 is in addition it is also necessary to by all unknown combined factors be One factor, is designated as c_l, it connects probability is p_l.Now calculate Probability p again_iAnd p_lIt is assumed that the father node of s is divided into two: c_lAnd l_all, wherein l_allIt is except c_iThe summation of all factors, corresponding outward, and it connects probability is p_iAnd p_all, simultaneously it is considered that l_allIt is always t value.Had according to formula (10)

p(s/c_i)=p_i+p_all-p_ip_all(11)

$p(s/\stackrel{\&overbar;}{c_l})=p_{all}---\left(12\right)$

And then can obtain

$p_i=\frac{p(s/c_i)-p(s/\stackrel{\&overbar;}{c_l})}{1-p(s/\stackrel{\&overbar;}{c_l})}---\left(13\right)$

HereRepresent non-c_lEvent, can calculate the connection Probability p of the father node of egress s by formula (13)₁, p₂..., obtain final product and be calculated p_i, that is, suppression probability to be tried to achieve.

Obtaining Probability p_iAfterwards, p can be calculated further_l.

According to known existing formula (14)

$p(\stackrel{\&overbar;}{s}/c_i=f)=\frac{{\σ}_{h/c_{i}eh}-(1-p_l^i){\π}_{c_{j}eh}+(1-p_j){\π}_{c_{j}eh}+\left(c_j\right){\π}_{c_{j}eh}-\left(\stackrel{\&overbar;}{c_j}\right)}{p\left(\stackrel{\&overbar;}{c_l}\right)}---\left(14\right)$

Can be calculatedI.e. with respect to the p of i-th node_l, here subscript l of p represent that this probability is above-mentioned p_l, Subscript i represents that this probability is i-th, i=1,2,3 ....HereRepresent non-s event, f represents that value is false, c_jRepresent s's It is different from c in father node_iThe node of sum, p_jRepresent and be different from p_iNode probability,Represent non-c_jEvent, h^-It is expressed as the son of h Collection, this subset is made up of the father node being worth for falsity.In order to try to achieve final p_l, then p can be obtained with weighting method_l,

$p_l={\mathrm{\σ}}_i\[1-p\left(c_i\right)\]p_l^i---\left(15\right)$

Obtain the leakage probability of each node in Internet.

The suppression probability of each observer nodes can be drawn and reveal probability by calculating us, respectively (take arithmetic point Three is significant digits afterwards):

7) pass through set up Bayesian network, calculate and verify equipment it may happen that fault and probability of happening: as flow Shown in journey figure (Fig. 4), the diagnostic cast the reasoning results using the present invention are compared with physical fault case result, to detect The accuracy of diagnosis.

In the daily production of selected unit, once there is a piston fault in this equipment, leads to unit cannot normally run, Have to stop production carrying out maintenance of equipment, suffer heavy losses.Fig. 5 is the oscillogram of unit vibration acceleration before fault generation.Logical Cross step 5) can obtain this unit vibration acceleration reference alarming value be 45m/s², it can be seen that sending out in fault Before life, unit equipment acceleration of vibration peak value is repeatedly more than 50m/s².

After through above steps, obtaining each item data, calculated using genie software.To exceed in software With reference to alarming value 45m/s²Acceleration of vibration node as observer nodes, being finally calculated this equipment unit using software can The fault that can occur and its probability, for unit it is possible that piston fault, the probability breaking down is 73.1% to result, belongs to Higher rate of breakdown, as shown in Figure 6.

As can be seen that fault diagnosis is carried out to plant equipment unit by the dynamic type Bayesian network proposing in this patent Reasoning, is coincide with physical fault situation, can make the higher equipment fault diagnosis result of accuracy rate.

Claims (2)

1. application in mechanical equipment fault intelligent diagnostics for the dynamic type Bayesian network is it is characterised in that comprise the following steps:

1) determine unit equipment fault characteristic, set up device-dependent Bayesian network integral frame:

2) understand unit equipment fault and fault signature, determine and set up each net of Bayesian network for fault diagnosis respectively Network layers: according to equipment fault and corresponding fault signature table, determine three Internets of Bayesian network of intelligent diagnostics, the One layer is facility information layer, and the second layer is equipment fault layer, and third layer is equipment fault characteristic layer；

3) according to unit equipment relevant information, determine and set up Bayesian network each node layer information: each in diagnostic network In layer, set up each required node respectively, and determine nodename；

4) determine between adjacent net network layers, interdependent node contacts: facility information layer and equipment fault layer, and equipment fault layer and Between equipment fault characteristic layer, build correlative connection between corresponding node；

5) by consulting related unit handbook and documents and materials, and the fault data of the statistics passing generation of unit, determine fault Reference alarming value under conditional probability and fault signature between, and unit different characteristic；

6) determine and setting network layer in each node suppression probability and leakage probability: determine and set the section in Bayesian network network layers Point probability；

7) calculate equipment it may happen that fault and probability of happening: by the input of unit equipment operating state signal, pass through Dynamic type Bayesian Network Inference calculates, equipment unit it may happen that fault and its probability dynamic show.

2. the method for claim 1 is it is characterised in that described step 6) particularly as follows:

1. make fault signature represent with x, represent i-th kind of fault with subscript i, i takes 1,2 ... ..., i.e. x_iRepresent i-th kind of fault spy Levy；Y is made to represent fault type, then each fault signature x_iPresence illustrate fault type y existence；By p_iDetermine node s Other items of conditional probability table, i=1 here, 2 ... ..., i.e. p_i=p₁, p₂..., i.e. formula (1)

$p(s=t/h)=1-{\mathrm{\π}}_{i:c_i\&element;h^{+}}(1-p_i)---\left(1\right)$

Wherein, p () represents the probability that in bracket, event occurs, p_i=p₁, p₂..., p_nFor between each fault and fault signature Conditional probability, s represents one of fault signature xi set, and it is true that t represents value, and h is the set of s father node value, c_iFor The father node of arbitrary node s, i=1,2,3 ... n, s=t/h represent that s is the conditional probability that h is during true value, h⁺Subset for h, should Subset is made up of the father node being worth for true value；In formula (1), latter half is to belong to h to all⁺I node, ask 1 to subtract Remove the product of its probability；

2. in order to avoid working as h⁺During for empty set, the situation of p (s=t/h)=0 is in addition it is also necessary to be one by all unknown combined factors Factor, is designated as c_l, it connects probability is p_l；Now calculate Probability p again_iAnd p_lIt is assumed that the father node of s is divided into two: c_iWith l_all, wherein l_allIt is except c_iThe summation of all factors, corresponding outward, and it connects probability is p_iAnd p_all, think l simultaneously_allIt is always T value；Had according to formula (1)

p(s/c_i)=p_i+p_all-p_ip_all(2)

$p\left(s/\stackrel{\&overbar;}{c_l}\right)=p_{all}---\left(3\right)$

And then obtain

$p_i=\frac{p\left(s/c_i\right)-p\left(s/\stackrel{\&overbar;}{c_l}\right)}{1-p\left(s/\stackrel{\&overbar;}{c_l}\right)}---\left(4\right)$

HereRepresent non-c_lEvent, calculates the connection Probability p of the father node of egress s by formula (4)₁, p₂..., obtain final product It has been calculated p_i, that is, suppression probability to be tried to achieve；

Obtaining Probability p_iAfterwards, calculate p further_l；

According to known existing formula (5)

$p\left(\stackrel{\&overbar;}{s}/c_i=f\right)=\frac{{\mathrm{\σ}}_{h/c_i\&element;h}-(1-p_l^i){\mathrm{\π}}_{c_j\&element;h}+(1-p_j){\mathrm{\π}}_{c_j\&element;h}+\left(c_j\right){\mathrm{\π}}_{c_j\&element;h}-\left(\stackrel{\&overbar;}{c_j}\right)}{p\left(\stackrel{\&overbar;}{c_l}\right)}---\left(5\right)$

Just obtainI.e. with respect to the p of i-th node_l, here subscript l of p represent that this probability is above-mentioned p_l, subscript i represents This probability is i-th, i=1,2,3 ...；HereRepresent non-s event, f represents that value is false, c_jRepresent in the father node of s not It is same as c_iThe node of sum, p_jRepresent and be different from p_iNode probability,Represent non-c_jEvent, h^-Be expressed as the subset of h, this subset by It is worth the father node for falsity to constitute；

ObtainingAfterwards, in order to try to achieve final p_l, then obtain p with weighting method_l,

$p_l={\mathrm{\σ}}_i\[1-p\left(c_i\right)\]p_{\underset{\&overbar;}{l}}^i---\left(6\right)$

Obtain the leakage probability of each node in Internet.