CN102509116A - Fault diagnosis knowledge acquisition method for support vector machine and rough set - Google Patents

Abstract

The invention discloses a fault diagnosis knowledge acquisition method for a support vector machine and a rough set. The fault diagnosis knowledge acquisition method comprises the following steps of: firstly, obtaining a support vector by using the support vector machine; then generating a certain amount of samples nearby the support vector by using a Monte Carlo simulation method; and finally obtaining a rule by using a rough set method and storing the obtained rule into a knowledge base. According to the fault diagnosis knowledge acquisition method, small samples are processed by using the support vector machine, and the support vector containing a lot of information is obtained, therefore, the sample randomness is guaranteed, the number of the samples is increased at the same time so that the obtained rule has broader applicability, not only the opaqueness of using the support vector machine directly is avoided, but also the objectivity of the obtained rule is ensured.

Description

The fault diagnosis knowledge acquisition method of a kind of SVMs and rough set

The method field

The invention belongs to artificial intelligence field, relate to method for diagnosing faults, knowledge acquisition method, reach fault tree analysis method, rough set theory, fault diagnosis knowledge acquisition method.

Background method

Along with the progress of science and technology, the complexity of equipment is increasingly high, and the thing followed is that the safety issue of equipment becomes increasingly conspicuous.Utilize servicing manual and expertise at present, carrying out fault diagnosis is a kind of main diagnostic method.But this diagnostic method can not demonstrate fully the internal system failure mechanism, also often has subjectivity.Therefore need make full use of the case (sample) of daily accumulation, therefrom excavate and to embody the causal knowledge of logic between failure symptom and the fault mode, set up the KBS of fault diagnosis, utilize these knowledge that equipment is carried out fault diagnosis then.

From sample, obtain the problem of fault diagnosis knowledge, be equivalent to the data qualification problem in the artificial intelligence, therefore can be by the knowledge head it off of artificial intelligence field.But sample size is a main method difficult problem that runs into when utilizing sample acquiring knowledge less.Support vector machine classification method based on statistical theory is applicable to the small sample classification of Data problem of handling just.SVMs relates to that the scale parameter of gaussian kernel function is chosen, penalty factor is chosen, the problem such as find the solution of quadratic programming problem.Selection of parameter problem for SVMs; At present both at home and abroad existing achievements in research, utilizing of wherein proposing in the document " the Determination of the spread parameter in the Gaussian kernel for classification and regression " method that the Fisher statistic calculates the calculating penalty factor that proposes in method and the document " Fast and efficient strategies for model selection of Gaussian Support vector Machine " of scale parameter of gaussian kernel function are univocal but also be easy to realize not only.

Utilize the number of the support vector that support vector machine method obtains fewer; One-sidedness for fear of the rule that produces; Make again that simultaneously the sample that is used for extracting rule can provide more information; Can utilize the Monte Carlo simulation method around support vector, to produce the sample point of some, utilize the automatic knowledge acquisition method of introducing among the patent ZL200910081793.3,200910236241.5 then, from the sample point of new generation, obtain rule based on rough set.Rough set theory relates to problems such as discretize, attribute reduction, value yojan.Document " expert system knowledge acquisition methods research with use " but in use the discretization method based on profile exponent, all be the method that is easy to Project Realization based on the didactic old attribute reduction algorithms of attribute importance degree and based on the value Algorithm for Reduction of identification matrix.

Summary of the invention

Can not the good treatment small sampling condition in order to overcome existing method; The fault diagnosis knowledge acquisition method that the present invention provides a kind of SVMs to combine with rough set; Can handle the fault diagnosis knowledge acquisition under the small sample situation, make the result have objectivity and higher reliability and accuracy.

The present invention solves the method scheme that its method problem adopted: at first utilize SVMs to obtain support vector; Utilize near a certain amount of sample of method generation support vector of Monte-Carlo Simulation then; Utilize rough set method to obtain rule at last; And the rule that will obtain deposits knowledge base in, specifically may further comprise the steps:

(1) calculates support vector

If sample set is T={ (x ₁, y ₁) ..., (x _l, y _l) ∈ (X * Y) ^l, x _i∈ X=R ⁿ, y _i∈ Y={1 ,-1}, i=1 ..., l, wherein x _iThe sign information of representing i sample, the conditional attribute of i sample just, y _iI sample of=1 expression is the fault sample, y _iI sample of=-1 expression is the non-fault sample.

Utilize SVMs to set up sorter f:R ⁿ→ { 1,1} calculates the quantity of support vector simultaneously, promptly finds the solution

$\max \underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i-\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_{j}K(x_i,x_j)$

s.t.0≤α _i≤C，i＝1，...，l (1)

$\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i=0$

Wherein C is the punishment parameter,

σ is the scale parameter of gaussian kernel function; a _i(i=1,2, Λ l) is known variables, α _i(i=1,2, Λ is not that 0 pairing sample is called support vector l).

(2) produce new samples

Being the center, be in the circle at center with δ respectively, utilize DSMC to produce n sample point, 0≤δ≤0.5,30≤n≤50 with the support vector of being tried to achieve.

(3) extracting rule may further comprise the steps:

(1) set up decision table, decision table is made up of sample, and the value of decision attribute is 1 or-1.

(2) discretize of carrying out of conditional attribute in the sample is handled.

(3) adopt the conditional attribute of removing redundancy in the decision table based on the heuristic Algorithm for Reduction of attribute importance degree.

(4) but adopted based on the value Algorithm for Reduction of identification matrix and removed case redundant in the decision table.

(5) rule that obtains in the decision table is stored in the knowledge base.

The invention has the beneficial effects as follows:

1. utilize SVMs to handle the small sample situation, obtain to have contained the support vector of bulk information;

2. utilize monte carlo simulation methodology, generate support vector great amount of samples on every side, not only guaranteed the randomness of sample, increased the quantity of sample simultaneously, make the rule of obtaining have wider applicability.

3. utilize the rough set principle from newly-generated sample, to obtain diagnostic rule, not only avoided directly utilizing the opacity of SVMs, the objectivity of the rule that has also guaranteed simultaneously to obtain.

Description of drawings

Fig. 1 is a fault diagnosis knowledge acquisition process flow diagram of the present invention.

Embodiment

Below in conjunction with accompanying drawing the fault diagnosis knowledge acquisition method based on SVMs of the present invention is elaborated.

The present invention includes following steps:

(3) calculate support vector

Fault diagnosis knowledge acquisition method based on SVMs as shown in Figure 1, that the present invention proposes at first utilizes SVMs to find out the support vector in the sample point.

If sample set is T={ (x ₁, y ₁) ..., (x _l, y _l) ∈ (X * Y) ^l, x _i∈ X=R ⁿ, y _i∈ Y={1 ,-1}, i=1 ..., l, wherein x _iThe sign information of representing i sample, the conditional attribute of i sample just, y _iI sample of=1 expression is the fault sample, y _iI sample of=-1 expression is the non-fault sample.

Given sample set T={ (x ₁, y ₁) ..., (x _l, y _l) ∈ (X * Y) ^lSituation under, the extraction problem of Failure Diagnostic Code just can be converted into calculates a sorter f:R ⁿ→ { 1,1} can utilize SVMs to set up sorter f:R ⁿ→ { 1,1} calculates the quantity of support vector simultaneously.

Support vector machine method is equivalent to find the solution a following quadratic programming problem:

$\max \underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i-\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_{j}K(x_i,x_j)$

s.t.0≤α _i≤C，i＝1，...，l (1)

$\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i=0$

Wherein C is the punishment parameter, and the effect of C is exactly to coordinate the model ability of classifier and nicety of grading degree;

Wherein σ is the scale parameter of gaussian kernel function; a _i(i=1,2, Λ l) is known variables, α _i(i=1,2, Λ is not that 0 pairing sample is called support vector l).

(1) calculating of the scale parameter of gaussian kernel function

The selection of gaussian kernel function scale parameter directly influences the number of support vector, and the Fisher statistic method of selecting among the present invention to propose in the document " Determination of the spread parameter in the Gaussian kernel for classification and regression " of utilizing is calculated the scale parameter of gaussian kernel function.This kind method has not only made full use of the range performance that gaussian kernel function can guarantee former data space; And utilize people always to hope that the distance between the similar sample is as far as possible little, and this big as far as possible general knowledge of the distance between the inhomogeneity sample has been constructed the Fisher statistic.

(2) calculating of penalty factor

The size of C also has certain influence to having influenced support vector, adopts the method that proposes in the document " Fast and efficient strategies for model selection of Gaussian Support vector Machine " to calculate penalty factor among the present invention.The thinking that the method is chosen C is an at first given smaller value C ₀, increase C with certain step-length then ₀Value, up to reaching optimal value.

(3) find the solution support vector

(1) finding the solution of formula can be summed up as quadratic programming (QP; Quadratic Programming) finds the solution problem; The training algorithm that has proposed at present is about to original QP PROBLEM DECOMPOSITION and becomes the less QP problem solving of plurality of scales as selecting block algorithm, decomposition algorithm and sequence minimum optimization algorithm etc. most of based on decomposing iterative idea.The SMO algorithm decomposes minimum with sub-optimization problem, in each iterative process, only need resolve the optimization subproblem of two variablees and find the solution, and does not have matrix operation, realize easily, and be to use algorithm the most widely.Adopt the minimized method of sequence to find the solution support vector in the present invention.)

(4) produce new samples

Being the center with the support vector of being tried to achieve respectively, is in the circle at center with δ (0≤δ≤0.5), utilizes DSMC to produce n (30≤n≤50) sample point.

(3) extracting rule

Utilize the process of rough set method extracting rule following:

(1) sets up decision table;

Decision table is made up of sample, and the value of decision attribute is 1 or-1.

Table 1 decision table

(2) attribute discretize;

When the utilization rough set theory obtained knowledge, requiring the value in the decision table was discrete data.Therefore need the discretize of carrying out of conditional attribute in the sample be handled.The present invention has adopted the discretization method based on profile exponent, and this is a kind of have supervision, local discretization method, and it can obtain rational breakpoint according to the actual distribution situation of data.

(3) decision table attribute reduction;

The conditional attribute that is comprised in the case be not be equal to important, some or even redundant, therefore need to remove conditional attribute redundant in the decision table, to obtain more easy rule.Design realizes attribute reduction module, the heuristic Algorithm for Reduction that is based on the attribute importance degree of employing.This algorithm is a starting point with the relative nuclear of decision table, and the importance degree size according to attribute joins it in yojan set.Then, remove each unnecessary attribute more successively, finally obtain the yojan property set.

(4) decision table property value yojan;

Still there is redundancy in case in the decision table through behind the attribute reduction, and the rule that therefrom obtains is not letter rule, therefore also need further be worth yojan to decision table, removes the case of redundancy.During design implementation value yojan functional module, but adopted value Algorithm for Reduction based on the identification matrix.The fresh information table that this value Algorithm for Reduction obtains, all properties value are the nuclear value of this table, and all records are the rule of this information table, for the conversion of rule is provided convenience.

(5) rale store.

Decision table is simplified to step (4) through above-mentioned steps (1), finally from decision table, obtained rule, be stored in the knowledge base.

Claims (1)

1. the fault diagnosis knowledge acquisition method of SVMs and rough set is characterized in that comprising the steps:

(1) calculates support vector

If sample set is T={ (x ₁, y ₁) ..., (x _i, y _i) ∈ (X * Y) ^l, x _i∈ X=R ⁿ, y _i∈ Y={1 ,-1}, i=1 ..., l, wherein x _iThe sign information of representing i sample, the conditional attribute of i sample just, y _iI sample of=1 expression is the fault sample, y _iI sample of=-1 expression is the non-fault sample;

Utilize SVMs to set up sorter f:R ⁿ→ { 1,1} calculates the quantity of support vector simultaneously, promptly finds the solution

$\max \underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i-\frac{1}{2}\underset{i=1}{\overset{l}{\mathrm{\Σ}}}\underset{j=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_{j}K(x_i,x_j)$

s.t.0≤α _i≤C，i＝1，...，l

$\underset{i=1}{\overset{l}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i=0$

Wherein C is the punishment parameter, and

σ is the scale parameter of gaussian kernel function;

a _i(i=1,2, Λ l) is known variables, α _i(i=1,2, Λ is not that 0 pairing sample is called support vector l);

(2) produce new samples

Being the center, be in the circle at center with δ respectively, utilize DSMC to produce n sample point, 0≤δ≤0.5,30≤n≤50 with the support vector of being tried to achieve;

(3) extracting rule may further comprise the steps:

(1) set up decision table, decision table is made up of sample, and the value of decision attribute is 1 or-1;

(2) discretize of carrying out of conditional attribute in the sample is handled;

(3) adopt the conditional attribute of removing redundancy in the decision table based on the heuristic Algorithm for Reduction of attribute importance degree;

(4) but adopted based on the value Algorithm for Reduction of identification matrix and removed case redundant in the decision table;

(5) rule that obtains in the decision table is stored in the knowledge base.

Images