CN102879677A - Intelligent fault diagnosis method based on rough Bayesian network classifier - Google Patents

Abstract

The invention provides an intelligent fault diagnosis method based on a rough Bayesian network classifier, which comprises the following steps: using standard fault feature data as a fault diagnosis condition attribute set, using a standard fault mode as a fault diagnosis decision attribute set, and adopting a rough set principle to construct an original fault diagnosis information table T1; adopting the minimum entropy method to carry out discrete processing on various continuous fault diagnosis condition attribute values in the T1, so as to form a discretization fault diagnosis information table T2; using a rough set discernable matrix and a nuclear theory to carry out attribute reduction and optimal feature selection on the T2, so as to form a reduction fault diagnosis information table T3; and using the T3 to establish the Bayesian network classifier, so as to realize efficient and quick intelligent fault diagnosis. The intelligent fault diagnosis method avoids the 'curse of dimensionality' problem existed in a Bayesian network diagnostic method, overcomes weaknesses of rigid reasoning and critical misjudgment in a rough set diagnostic method, and greatly improves the efficiency and accuracy of fault diagnosis.

Description

Intelligent failure diagnosis method based on coarse BAYESIAN NETWORK CLASSIFIER

Technical field

The present invention relates to the mechanical ﹠ electrical technology field, especially a kind of intelligent failure diagnosis method.

Background technology

In failure diagnostic process, because the mechanism that fault produces is unclear, the form of expression of fault is not unique, when extracting the various parameter of describing fault signature also usually with certain blindness, thereby to have caused between the malfunction be fuzzy.Rough set theory can under the prerequisite that guarantee information is not lost, carry out effective yojan to decision system from describing the raw data of malfunction, removes redundant information, can carry out Rule Extraction and yojan simultaneously, thereby can effectively address the above problem.But if use separately rough set to carry out fault diagnosis, often have following two problems: 1. many based on the knowledge base logical theory to the processing of rule, efficient is not high; 2. do not provide the relation between rule, the rule that obtains is the reasoning of production to the reasoning of malfunction, and diagnosis speed is not high.

Because Bayesian network can effectively represent heuristic knowledge, knowledge base system is simple in structure, standard, space search efficient are high, is easy to realize parallel inference, and the reasoning results is easy to explain, so it can effectively overcome the above-mentioned deficiency of rough set.But when directly using Bayesian network model, often exist to hinder for some reason the erroneous judgement problem that feature too much causes.

Foundation and the study of BAYESIAN NETWORK CLASSIFIER that at present numerous scholar's research arranged.Riccardi G, Hakkani-Tur D.Active learning:Theory and applications to automatic speech Recognition[J] .IEEE Transactions on Speech and Audio Processing, 2005,13 (4): 504-510 initiatively Bayes classifier is used for speech recognition.But the method also has the following disadvantages: initiatively Bayes classifier is only limited to regular length to the extraction of feature, and verification and measurement ratio and the precision of model have much room for improvement, and calculated amount is large when setting up model, and the telephone expenses time is many.Huang Wei, Dai Beiqian, Li Hui. based on the Speaker Identification [J] of characteristic of division space Gaussian Mixture Model and neural network fusion. electronics and information journal, 2004,26 (10): Xue 1607-1612. is upright, Fang Shuai etc. the Opponent Modeling [J] in the emulation of multirobot countermeasure system. Journal of System Simulation, 2005,9:2138-2141 is applied to the integrated Bayesian network of multimode in Speaker Identification and the modeling of multirobot countermeasure system.But in use still there is following problem in the method: when the example that comprises in dividing was very few, the estimation about conditional probability in these little divisions was insecure; In training set, for the situation of data defect, all to have taked to regard as a special value or before training, just will contain the method that the example of damaged value removes, this will can cause losing of information inevitably.Liu Dayou, Wang Fei etc. based on the Bayesian web frame study research [J] of genetic algorithm. Journal of Computer Research and Development, 2001,20 (5): 916-922 has carried out theoretic research to the BAYESIAN NETWORK CLASSIFIER based on genetic algorithm, comprises the structure and parameter study of Bayesian network, the realization of algorithm and the exploitation of system platform.But the method has only been considered Category Attributes and desirable data in the learning algorithm of BAYESIAN NETWORK CLASSIFIER, has taked the method cast out for the processing of obliterated data, and has not considered that there is the situation of hidden variable in data centralization; In addition, the learning algorithm of the method has adopted the way that generates at random and the structure that produces is repaired in the generation of initial population, do not generate initial population in conjunction with priori, the fitness function of algorithm needs the calculating of a large amount of conditional mutual informations, it is larger to calculate the time overhead that needs, and it is relatively poor to carry out efficient.

Summary of the invention

Not high and cause easily the deficiency of erroneous judgement in order to overcome prior art efficient, the invention provides a kind of intelligent failure diagnosis method based on coarse BAYESIAN NETWORK CLASSIFIER, not only can from diagnostic data, extract the key condition attribute, and can reduce the scale of Bayesian network model, shorten Bayesian Network Inference computing time, thereby " dimension disaster " problem the problem includes: of having avoided problem in the Bayesian network diagnosis, overcome the weakness of the rigidity reasoning of rough set diagnosis and critical erroneous judgement, greatly improved efficient and the accuracy of fault diagnosis.

The technical solution adopted for the present invention to solve the technical problems may further comprise the steps:

(1) with standard fault signature data as fault diagnosis conditional attribute collection, as the Fault Tree Diagnosis Decision property set, adopt the rough set principle to make up primary fault diagnosis information system T with the standard fault mode ₁

(2) adopt the minimum entropy method to T ₁In each continuous fault conditions for diagnostics property value carry out discrete processes, form discretize failure diagnosis information table T ₂, specifically may further comprise the steps:

(2a) with continuous fault conditions for diagnostics property value by sorting from small to large, and calculate mean value and the information entropy thereof of the conditional attribute value that links to each other in twos;

The mean value of (2b) selecting the information entropy minimum is as first threshold value PRI, and with interval division be two sub-ranges [0, PRI] and [PRI ,+∞);

(2c) for each mean value less than PRI, recomputate its in the sub-range corresponding information entropy in [0, PRI], select the mean value of information entropy minimum as second threshold value SEC ₁

(2d) for each mean value greater than PRI, recomputate its in the sub-range [PRI ,+∞) in corresponding information entropy, select the mean value of information entropy minimum as the 3rd threshold value SEC ₂

(2e) calculate each conditional attribute value degree of membership based on three threshold values, and carry out discretize by maximum membership grade principle, thereby form discretize failure diagnosis information table T ₂

(3) adopt rough set discrimination matrix and nuclear theory to T ₂It is preferred to carry out attribute reduction and optimal characteristics, forms yojan failure diagnosis information table T ₃, specifically may further comprise the steps:

(3a) calculate discretize failure diagnosis information table T ₂Discrimination matrix M;

(3b) calculate discretize failure diagnosis information table T ₂The nuclear attribute, and generate new discrimination matrix M ₁

(3c) calculate discrimination matrix M ₁Resolution function;

(3d) resolution function is turned to the disjunctive normal form form;

(3e) will examine in the attribute all properties and join in the disjunctive normal form in each conjunction expression, each conjunct of disjunctive normal form is just corresponding to the result of an attribute reduction;

(3f) calculate the clustering precision of each yojan, and form yojan failure diagnosis information table T according to the yojan with maximum clustering precision ₃

(4) adopt T ₃Thereby set up BAYESIAN NETWORK CLASSIFIER and realize efficiently fast intelligent trouble diagnosis, specifically may further comprise the steps:

(4a) by yojan failure diagnosis information table T ₃Set up the diagnostic reasoning Bayesian network model;

(4b) calculate each fault new samples with respect to the posterior probability of each standard fault mode;

(4c) determine the corresponding fault type of this new fault sample based on the maximum a posteriori probability principle.

The invention has the beneficial effects as follows:

The advantage of rough set diagnosis is by the yojan conditional attribute, can compress the diagnostic characteristic rule, rejects unwanted feature, and diagnostic rule is effectively simplified, and can improve the efficient of fault diagnosis.But when adopting rough set to carry out reasoning, if the fault signature data volume is larger, the search rule of tabling look-up is a process that calculated amount is very large; In addition, it requires to process the discrete magnitude property value, is a kind of qualitative analysis in essence, critical problem occurs easily, thereby causes classification error or erroneous judgement.

The advantage of Bayesian network diagnosis is standard simple in structure, and space search efficient is high, is easy to realize parallel quantitative reasoning.The priori but the inferential capability of Bayesian network model places one's entire reliance upon, and the ability of knowledge not being simplified, and often have certain redundancy in the priori, this causes the scale of Bayesian network model too huge, the work of obtaining of fault signature is heavy, directly affects accuracy and the efficient of reasoning.

It is integrated that the present invention has done serial with the function of rough set and Bayesian network, the robust parsing ability of rough set that organically blended and the parallel inference ability of Bayesian network, rule search and the critical erroneous judgement problem of rough set diagnosis have not only been overcome, and having avoided the dimension disaster problem of Bayesian Diagnosis method, simultaneously the present invention also combines the qualitative and quantitative analysis.Therefore, the present invention is better than existing rough set diagnosis and Bayesian network diagnosis.

Description of drawings

Fig. 1 is the systems function diagram of intelligent trouble diagnosis of the present invention;

Fig. 2 is the key diagram of minimum entropy method discretize process of the present invention, wherein, is to divide for the first time (a), (b) is to divide for the second time;

Fig. 3 is the key diagram of Bayesian network diagnostic model of the present invention.

Embodiment

The present invention is further described below in conjunction with drawings and Examples.

One, acquisition characteristics data and make up the failure diagnosis information decision table based on rough set

The present invention carries out fault diagnosis take certain complex rotor system as example.According to collecting the fault signature data, can set up primary fault diagnostic message decision table T ₁As shown in table 1, wherein seven conditional attributes are respectively: C ₁Expression 0.01～0.40f, C ₂Expression 0.41～0.50f, C ₃Expression 0.51～0.99f, C ₄Expression 1f, C ₅Expression 2f, C ₆Expression 3～5f, C ₇Expression 5f; Five kinds of fault types are respectively: D ₁The expression rotor unbalance, D ₂The expression rotor misalignment, D ₃Expression oil whip, D ₄The expression surge, D ₅The expression collision.

Certain rotor-support-foundation system primary fault diagnostic message decision table of table 1

Two, adopt the minimum entropy method to realize the processing of continuous fault conditions for diagnostics property value discretize

Be located at primary fault diagnosis decision table T ₁=＜U, C, D〉in, total n fault sample, m fault condition attribute.For each continuous condition attribute C _i(i=1,2 ..., m), its discretize step is:

(1) with continuous condition attribute value c _I1, c _I2..., c _Ij..., c _InBy sorting from small to large, and the conditional attribute value after will sorting is designated as c ' _I1, c ' _I2..., c ' _Ij..., c ' _In

(2) calculate the c ' that links to each other in twos _I1, c ' _I2..., c ' _Ij..., c ' _InMean value, and mean value is designated as

, wherein ${\stackrel{\‾}{c}}_{\mathrm{ij}}^{\′}=(c_{\mathrm{ij}}^{\′}+c_{i(j+1)}^{\′})/2;$

(3) for each

, calculate its interval [c ' _I1, c ' _In] interior corresponding information entropy;

S(x)＝p(x)·S _p(x)+q(x)·S _q(x)

S _p(x)＝-[p ₁(x)lnp ₁(x)+p ₂(x)lnp ₂(x)]

S _q(x)＝-[q ₁(x)lnq ₁(x)+q ₂(x)lnq ₂(x)]

p _k(x)＝(n _k+1)/(n(x)+1)

q _k(x)＝(N _k+1)/(N(x)+1)

p(x)＝n(x)/n

q(x)＝1-p(x)

In the formula, p _k(x) and q _k(x) be respectively the conditional probability of k class sample on p interval and q interval; P (x) and q (x) are respectively the probability of all samples on p interval and q interval; n _k(x) and N _k(x) be respectively the number of samples of k class sample on p interval and q interval; N (x) and N (x) are respectively the total sample number on p interval and the q interval; N is interval [x ₁, x ₂] on total sample number; K is sample type; M is the sample type sum.

(4) select the information entropy minimum

As first threshold value PRI, and with the interval [c ' _I1, c ' _In] be divided into two sub-ranges [c ' _I1, PRI] and [PRI, c ' _In].

(5) for each

Recalculate its in this sub-range [c ' _I1, PRI] and interior corresponding information entropy, selection information entropy minimum

As threshold value SEC ₁

(6) for each

Recalculate it at this sub-range [PRI, c ' _In] interior corresponding information entropy, selection information entropy minimum

As threshold value SEC ₂

(7) based on three eigenwerts and maximum membership grade principle, with each conditional attribute value c _I1, c _I2..., c _Ij..., c _InCarry out discretize, and the conditional value after the discretize is designated as

Wherein

$c_{\mathrm{ij}}^d=\left\{\begin{array}{ccc}1& \mathrm{if}& c_{\mathrm{ij}}\&le;\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA00002184525300012.png"\; state="\{\{state\}\}">P</figure-callout>}1\\ 2& \mathrm{if}& \mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HDA00002184525300012.png"\; state="\{\{state\}\}">P</figure-callout>}1<c_{\mathrm{ij}}\&le;P2\\ 3& \mathrm{if}& c_{\mathrm{ij}}>P2\end{array}\right.$

P in the formula ₁And P ₂Be break value, and P ₁=(SEC ₁+ PRI)/2, P ₂=(SEC ₂+ PRI)/2.

Fig. 2 has provided continuous condition attribute C _iThe key diagram processed of discretize, if with original decision table T ₁In after all conditional attributes all carry out discretize, just can obtain the decision table T of discretize ₂Table 2 and table 3 have provided respectively break value and the discretize failure diagnosis information decision table of each conditional attribute discretize of this rotor-support-foundation system.

The break value of each conditional attribute discretize of this rotor-support-foundation system of table 2

Conditional attribute	SEC 1	PRI	SEC 2	Break value P 1	Break value P 2
						C 1	0.05165	0.08610	0.24790	0.06888	0.16700
C 2	0.13125	0.18155	0.42620	0.15640	0.30388
						C 3	0.04250	0.05590	0.12320	0.04920	0.08955
C 4	0.27190	0.56080	0.66630	0.41635	0.61355
						C 5	0.14620	0.24475	0.48955	0.19548	0.36715
C 6	0.01615	0.03795	0.32395	0.02705	0.18095
						C 7	0.05035	0.05455	0.19930	0.05245	0.12692

This rotor-support-foundation system discretize failure diagnosis information decision table of table 3

Ob	C 1	C 2	C 3	C 4	C 5	C 6	C 7	D	Ob	C 1	C 2	C 3	C 4	C 5	C 6	C 7	D
																		X 1	2	1	3	3	1	2	1	D 1	X 10	1	2	2	1	1	2	2	D 3
X 2	2	1	3	3	1	2	1	D 1	X 11	3	3	3	1	1	2	2	D 3
																		X 3	1	1	3	3	1	3	1	D 1	X 12	1	2	1	2	1	2	2	D 4
X 4	1	1	2	2	1	3	1	D 1	X 13	2	2	2	1	1	2	3	D 4
																		X 5	2	1	3	1	3	2	2	D 2	X 14	2	3	1	1	1	2	3	D 4
X 6	2	1	3	2	2	1	3	D 2	X 15	2	1	1	3	1	2	2	D 5
																		X 7	2	1	2	1	3	2	2	D 2	X 16	2	1	2	2	1	2	2	D 5

X 8	2	1	2	2	2	2	1	D 2	X 17	2	1	3	2	1	3	2	D 5
																		X 9	3	2	2	1	1	2	2	D 3	X 18	2	1	3	1	1	3	3	D 5

Three, adopt rough set discrimination matrix and nuclear theory is carried out attribute reduction and optimal characteristics is preferred

Reduction of condition attributes is exactly to remove unnecessary property value.Because rough set discrimination matrix and nuclear theory simple, intuitive, the present invention adopts the reduction of condition attributes algorithm based on discrimination matrix, and its concrete steps are:

(1) calculate discretize Fault Tree Diagnosis Decision table S '=＜U, C ', D '〉discrimination matrix M;

(2) calculate nuclear attribute and generate new discrimination matrix: be that 1 attribute is listed core attribute set in and closed with combinations of attributes number in the discrimination matrix, obtain CORE _D(M).If Elements C among the M _KiComprise element among the Cr, then make C _Ki=0; Otherwise remain unchanged, thereby obtain a new discrimination matrix M '.

(3) resolution function of calculating discrimination matrix M ': claim

Be the resolution function of S, wherein ∧ represents the computing of extracting, and ∨ represents the conjunction computing.

(4) changing resolution function is the disjunctive normal form form;

(5) will examine that all properties joins each conjunction expression in the disjunctive normal form in the attribute, just corresponding to an attribute reduction;

(6) calculate the cluster rate of each minimum yojan collection, then make the minimum yojan of cluster rate maximum integrate as best induct, just can get optimum yojan failure diagnosis information decision table T ₃

$R_C=\frac{N_0-N_R}{N_0-1}$

In the formula, N ₀Be decision table T ₂In object number; N _RDecision table T after the yojan ₃In object number.

Table 4 and table 5 have provided respectively each yojan collection of this rotor-support-foundation system and optimum yojan failure diagnosis information decision table.

Each reduction of condition attributes collection of this rotor-support-foundation system of table 4 and cluster rate thereof

Yojan	The cluster rate	Yojan	The cluster rate
				{C 1,C 2,C 4,C 7}	0.1765	{C 1,C 4,C 6,C 7}	0.2353
{C 1,C 3,C 4,C 5}	0.0588	{C 2,C 4,C 5,C 7}	0.2941
				{C 1,C 3,C 4,C 7}	0.0588	{C 1,C 2,C 3,C 5,C 6}	0.1176

{C 1,C 3,C 5,C 7}	0.0588	{C 2,C 4,C 6,C 7}	0.1765
				{C 2,C 3,C 5,C 7}	0.1765	{C 3,C 4,C 5,C 7}	0.1765

The optimum yojan failure diagnosis information of this rotor-support-foundation system of table 5 decision table

Ob	Number	C 2	C 4	C 5	C 7	D	Ob	Number	C 2	C 4	C 5	C 7	D
														X 1	3	1	3	1	1	D 1	X 8	1	2	2	1	2	D 4
X 2	1	1	2	1	1	D 1	X 9	1	2	1	1	3	D 4
														X 3	2	1	1	3	2	D 2	X 10	1	3	1	1	3	D 4
X 4	1	1	2	2	3	D 2	X 11	1	1	3	1	2	D 5
														X 5	1	1	2	2	1	D 2	X 12	2	1	2	1	2	D 5
X 6	2	2	1	1	2	D 3	X 13	1	1	1	1	3	D 5
														X 7	1	3	1	1	2	D3

Four, adopt BAYESIAN NETWORK CLASSIFIER to realize efficiently fast intelligent trouble diagnosis

If yojan failure diagnosis information table T ₃Object among the=＜U ", C ', D 〉, wherein, U " be X ' ₁, X ' ₂..., X ' _n, the number of times that they occur separately is respectively

And satisfy

Conditional attribute C ' is the optimum yojan of determining according to maximum cluster ratio, be designated as C ' ₁, C ' ₂... C ' _m; Fault type among the decision attribute D still is D ₁, D ₂..., D _k, the number of times that they occur separately is respectively

And satisfy Make conditional attribute C ' ₁, C ' ₂..., C ' _mBe the attribute variable of example, fault type D ₁, D ₂..., D _KBe the class variable of example, suppose that all properties is separate, and each attribute variable with class variable as unique father node, just can set up fault diagnosis Bayesian network model as shown in Figure 3.

Suppose a given fault sample X={C ' ₁, C ' ₂..., C ' _m, each property value is respectively C ' 1=a ₁, C ' ₂=a ₂..., C ' _m=a _m, this fault sample belongs to certain failure classes D in the failure classes variables D _kProbability be

P(D＝D _k|C′1＝a ₁,C′ ₂＝a ₂,…,C′ _m＝a _m)

According to Bayes' theorem:

P(D _k|C′)＝P(C′|D _k)P(D _k)P(C′)

Because P (C ') is constant for all failure classes, so only need P (C ' | D _k) P (D _k) get final product.P (D wherein _k) for prior probability be:

$P\left(D_k\right)=N_{D_k}/N,(k=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,K)$

When the number of attribute is very many, maximum a posteriori probability P (D _k| C ') calculated amount very large, for reducing calculated amount, Bayesian network model is supposed each attribute C ' ₁, C ' ₂..., C ' _mBetween separate, only relevant with fault type D, then its computing formula is:

$P\left(C^{\&prime;}\right|D_k)=\underset{i=1}{\overset{m}{\mathrm{\&Pi;}}}P\left(C_i^{\&prime;}\right|D_k)$

Wherein probability P (C ' _i| D _k) be

$P\left(C_i^{\&prime;}\right|D_k)=\left\{\begin{array}{cc}{N}_{D_k}^{C_i^{\&prime;}}/N_{D_k}& \mathrm{if}{N}_{C_k}^{x_i}\&NotEqual;0\\ 1/(N\&CenterDot;N_{D_k}+N_{C_i^{\&prime;}})& \mathrm{if}{N}_{C_k}^{x_i}=0\end{array}\right.$

In the formula,

Be fault type D in the training sample _kWith conditional attribute value C ' _iThe number of samples that occurs simultaneously.Posterior probability P that can be corresponding according to each fault type (C ' | D _k) P (D _k) carry out the fault decision-making, its primitive rule be the fault mode judged should have maximum P (C ' | D _k) P (D _k) value.

Claims (1)

1. the intelligent failure diagnosis method based on coarse BAYESIAN NETWORK CLASSIFIER is characterized in that comprising the steps:

(1) with standard fault signature data as fault diagnosis conditional attribute collection, as the Fault Tree Diagnosis Decision property set, adopt the rough set principle to make up primary fault diagnosis information system T with the standard fault mode ₁

(2) adopt the minimum entropy method to T ₁In each continuous fault conditions for diagnostics property value carry out discrete processes, form discretize failure diagnosis information table T ₂, specifically may further comprise the steps:

(2a) with continuous fault conditions for diagnostics property value by sorting from small to large, and calculate mean value and the information entropy thereof of the conditional attribute value that links to each other in twos;

The mean value of (2b) selecting the information entropy minimum is as first threshold value PRI, and with interval division be two sub-ranges [0, PRI] and [PRI ,+∞);

(2c) for each mean value less than PRI, recomputate its in the sub-range corresponding information entropy in [0, PRI], select the mean value of information entropy minimum as second threshold value SEC ₁

(2d) for each mean value greater than PRI, recomputate its in the sub-range [PRI ,+∞) in corresponding information entropy, select the mean value of information entropy minimum as the 3rd threshold value SEC ₂

(2e) calculate each conditional attribute value degree of membership based on three threshold values, and carry out discretize by maximum membership grade principle, thereby form discretize failure diagnosis information table T ₂

(3) adopt rough set discrimination matrix and nuclear theory to T ₂It is preferred to carry out attribute reduction and optimal characteristics, forms yojan failure diagnosis information table T ₃, specifically may further comprise the steps:

(3a) calculate discretize failure diagnosis information table T ₂Discrimination matrix M;

(3b) calculate discretize failure diagnosis information table T ₂The nuclear attribute, and generate new discrimination matrix M ₁

(3c) calculate discrimination matrix M ₁Resolution function;

(3d) resolution function is turned to the disjunctive normal form form;

(3e) will examine in the attribute all properties and join in the disjunctive normal form in each conjunction expression, each conjunct of disjunctive normal form is just corresponding to the result of an attribute reduction;

(3f) calculate the clustering precision of each yojan, and form yojan failure diagnosis information table T according to the yojan with maximum clustering precision ₃

(4) adopt T ₃Set up BAYESIAN NETWORK CLASSIFIER, specifically may further comprise the steps:

(4a) by yojan failure diagnosis information table T ₃Set up the diagnostic reasoning Bayesian network model;

(4b) calculate each fault new samples with respect to the posterior probability of each standard fault mode;

(4c) determine the corresponding fault type of this new fault sample based on the maximum a posteriori probability principle.

Images