CN105530122A - Network failure diagnosis method based on selective hidden Naive Bayesian classifier - Google Patents

Abstract

The invention discloses a network failure diagnosis method based on a selective hidden Naive Bayesian classifier, comprising: (1), obtaining history data from a network history database, wherein the history data comprise a symptom variable set and a failure class variable set; (2), constructing a selective hidden Naive Bayesian classifier prediction model, determining corresponding most related symptom variable set according to every symptom variable in the symptom variable set; (3), automatically learning classifier parameters by the selective hidden Naive Bayesian classifier through training the history data; (4), in failure diagnosis, estimating the test data by using the selective hidden Naive Bayesian classifier so as to obtain corresponding final failure diagnosis result. Through executing the network failure diagnosis method of the invention, the problems in the existing network failure diagnosis that the operation complexity is high and the network diagnosis result is great in deviation are effectively solved; the network diagnosis accuracy is greatly improved; the operation complexity is further reduced, and better learning capability and fault-tolerant character are kept at the same time.

Description

A kind of network fault diagnosis method based on the hidden Naive Bayes Classifier of selectivity

Technical field

The present invention relates to technical field of communication network, more specifically, relate to a kind of network fault diagnosis method based on the hidden Naive Bayes Classifier of selectivity.

Background technology

It is large that modern network has scale, the feature that network topology complexity is high.In network, certain part breaks down, and can cause series of symptoms.If fault fails to be diagnosed in time, the systemic-function of so whole network, reliability service, safety in production all can be affected, even can network paralysis, and therefore, failure diagnosis is particularly important to network.

Early stage failure diagnosis relies on expertise completely, but is difficult to the reliability and the stability that maintain network running.Therefore, in large complicated network, intelligent diagnostics is widely applied, and certain methods wherein is also suggested, such as Bayesian network, artificial intelligence.

Naive Bayes Classifier (NB) is the simplest Bayesian network, but the prerequisite that it is set up is separate between each attribute, make Naive Bayes Classifier can not utilize Dependency Specification between attribute variable, be not therefore suitable for most of data set in reality.Follow-up, the Naive Bayes Classifier (TAN) improved has put forward, it is by finding that the dependence between attribute reduces the hypothesis of attribute independent in Naive Bayes Classifier, but its attribute node can only have at most an attribute father node, make TAN grader can only utilize the Dependency Specification of part, this can not play good effect.For making full use of the dependence between attribute node, hidden Naive Bayes Classifier (HNB) has put forward, this model represents dependence between this attribute and other each attributes by introducing a hiding father node for each attribute node on the basis of Naive Bayes Classifier, in so avoiding the complicated calculations learning optimum bayesian network structure, the independence assumption between attribute can be weakened again.

Hidden Naive Bayes Classifier HNB structure as shown in Figure 1, contains 3 category nodes: category node, attribute node, hiding father node, uses C respectively, A, A _hprepresent.In HNB, the dependence between attribute is exactly represented by the hiding father node of attribute, and the method for expressing hiding father node just can determine the power of the dependence between attribute.

Though Naive Bayes Classifier HNB has considered the dependence between attribute, its computational complexity is not satisfactory, still can the expensive testing time, is especially used on high-dimensional data set.In addition, if consider all properties of sample, may noise jamming be added, and reduce grader efficiency.Therefore, this area is needed badly and is found the network fault diagnosis method that a kind of accuracy is higher, computation complexity is lower.

Summary of the invention

For above defect or the Improvement requirement of prior art, the invention provides a kind of network fault diagnosis method based on the hidden Naive Bayes Classifier of selectivity, effectively can solve network failure quick diagnosis and the lower problem of accuracy rate of diagnosis accordingly, thus be particularly useful for the network fault diagnosis that topological complexity is high.

For achieving the above object, the present invention proposes a kind of network fault diagnosis method based on the hidden Naive Bayes Classifier of selectivity, it is characterized in that, described method comprises:

(1) from network history data storehouse, obtain historical data, comprise symptom variables collection and failure classes variables set;

(2) build selectivity hidden Naive Bayes Classifier forecast model, each symptom variables concentrated according to symptom variables determines corresponding most related symptoms variables collection;

(3) the hidden Naive Bayes Classifier of described selectivity is by training historical data automatic learning to classifier parameters;

(4), when carrying out failure diagnosis, the hidden Naive Bayes Classifier of above-mentioned selectivity is utilized to estimate to obtain corresponding final fault diagnosis result to test data.

As further preferred, describedly determine corresponding most related symptoms variables collection for each symptom variables, be specially: for symptom variables A _i, estimate A _iand A _jbetween mutual information, wherein i, j ∈=1,2 ..., n, i ≠ j}, line ordering of going forward side by side, chooses front m maximum symptom variables of mutual information as symptom variables A _imost related symptoms variables collection.

As further preferred, the quantity of described most related symptoms variable is determined according to reality test.

As preferred further, described selectivity hidden Naive Bayes Classifier forecast model F is specially:

$F=\arg \underset{c\⋐C}{max}P\left(c\right)\underset{i=1}{\overset{n}{\Π}}P\left(a_i\right|a_{phi},c)$

Wherein, C is failure classes variables set, and c is failure classes variable, and P (c) is corresponding failure classes variable probability, P (a _i| a _phi, c) be symptom variables a _iwith its most related symptoms variables collection a _phiconditional probability, i=1,2 ..., n, n are symptom variables number.

As preferred further, described symptom variables a _iwith its most related symptoms variables collection a _phiconditional probability P (a _i| a _phi, c) be specially:

$P\left(a_i\right|a_{phi},c)=\underset{k=1}{\overset{m}{\Σ}}{W}_{i-ik}*P\left(a_i\right|a_{ik},c)$

Wherein, W _i-ikrepresent symptom variables a _iwith its most related symptoms variable a _ikbetween conditional mutual information, P (a _i| a _ik, c) be symptom variables a _iwith its most related symptoms variable a _ikconditional probability, a _ik∈ a _phi, k=1,2 ..., m, m are most related symptoms variable number.

As preferred further, described symptom variables a _iwith its most related symptoms variable a _ikbetween conditional mutual information W _i-ikbe specially:

$W_{i-ik}=\frac{\underset{a_i,a_{ik},c}{\Σ}P(a_i,a_{ik},c)log\frac{P(a_i,a_{ik},c)}{P\left(a_i|c\right)P\left(a_{ik}|c\right)}}{{\mathrm{\Σ}}_{k=1}^m\underset{a_i,a_{ik},c}{\Σ}P\left(a_i,a_{ik},c\right)log\frac{P\left(a_i,a_{ik},c\right)}{P\left(a_i|c\right)P\left(a_{ik}|c\right)}}$

Wherein, P (a _i, a _ik, c) be a _iand a _ikjoint probability, P (a _i| c) be a _iconditional probability, P (a _ik| c) be a _ikconditional probability.

As preferred further, in step (3), described classifier parameters comprises symptom variables conditional probability P (a _i| a _j, c) and failure classes variable probability P (c), be specially:

$P\left(a_i\right|a_j,c)=\frac{F(a_i,a_j,c)+1/n}{F(a_j,c)+1}$

$P\left(c\right)=\frac{F\left(c\right)+1/s}{t+1}$

Wherein, F (a _i, a _j, c) symptom variables a in historical data _iand a _jthe number of times simultaneously occurred, F (a _j, c) be symptom variables a in historical data _jthe number of times occurred, F (c) is the number of times of failure classes occurrences in historical data, and t is the quantity of historical data, and s is failure classes variable quantity, and n is symptom variables number.

In general, the above technical scheme conceived by the present invention compared with prior art, mainly possesses following technological merit:

1. the network fault diagnosis method based on the hidden Naive Bayes Classifier of selectivity of the present invention's proposition, can accuracy rate of diagnosis is low and failure diagnosis computational complexity is high in corresponding solution existing network method for diagnosing faults problem, by the improvement to Bayes classifier, further increase the accuracy rate of failure diagnosis, thus make the method have lower computational complexity, shorten the diagnostic test time;

2. the feature that the present invention is especially large for existing communication network of network scale and network topology complexity is high, devise the hidden Naive Bayes Classifier of a kind of selectivity, by this grader, failure diagnosis is carried out to network, substantially increase learning ability and the failure tolerance of grader self.

Accompanying drawing explanation

Fig. 1 is hidden Naive Bayes Classifier structure chart;

Fig. 2 is failure diagnosis protocol procedures figure;

Fig. 3 is the hidden Naive Bayes Classifier structure chart of selectivity;

Fig. 4 is the historical data schematic diagram obtained in example;

Fig. 5 is failure classes variable and symptom variables statistical chart in example;

Fig. 6 is the accuracy rate that under experiment condition, different m value is corresponding;

Fig. 7 is the selectivity hidden Naive Bayes Classifier forecast model structure chart built in example;

Fig. 8 is the failure diagnosis accuracy rate statistical chart adopted in example under different forecast model;

Fig. 9 is the Diagnostic Time statistical chart of the failure diagnosis adopted when being 300 for historical data quantity in example under different forecast model;

Figure 10 adopts the failure diagnosis accuracy rate statistical chart under different forecast model when being historical data loss.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.In addition, if below in described each execution mode of the present invention involved technical characteristic do not form conflict each other and just can mutually combine.

The place that the present invention is different from existing fault diagnosis technology is just on choosing of grader forecast model, the present invention proposes the network fault diagnosis method based on the hidden Naive Bayes Classifier of selectivity (SHNB), by performing the technical scheme in the present invention, failure diagnosis accuracy rate can be significantly improved, reduce the computation complexity of network fault diagnosis.

As shown in Figure 2, said method comprising the steps of:

(1) from network history data storehouse, obtain historical data, comprise symptom variables collection and failure classes variables set;

(2) build selectivity hidden Naive Bayes Classifier forecast model, each symptom variables concentrated according to symptom variables determines corresponding most related symptoms variables collection;

(3) the hidden Naive Bayes Classifier of described selectivity is by training historical data automatic learning to classifier parameters;

(4), when carrying out failure diagnosis, the hidden Naive Bayes Classifier of above-mentioned selectivity is utilized to estimate to obtain corresponding final fault diagnosis result to test data.

As shown in Figure 3, wherein C is category node to SHNB grader forecast model structure, i.e. failure classes variables set, points to all attribute A ₁, A ₂..., A _n, i.e. symptom variables collection, to each attribute A _i(i=1,2 ..., n) have hiding father node S _i(i=1,2 ..., n), namely corresponding with each symptom variables most related symptoms variables collection, it comprises and A _ia maximally related m symptom variables, m is empirical value, and the value for different historical data m is not quite similar, and that is, the size m of described most related symptoms variables collection determines according to reality test.

Concerning SHNB, how to select most related symptoms variables set particularly crucial as hiding father node, the present invention adopts the mutual information between symptom variables.To symptom variables A _i(i=1,2 ..., n), by calculating A _i(i=1,2 ..., n) and A _j(j=1,2 ..., n, j ≠ i) between mutual information, select the maximum m of mutual information most related symptoms variable as hiding father node, this m most related symptoms variable joined S _i(i=1,2 ..., n), i.e. S _i={ A _i1, A _i2..., A _ik..., A _im.

In step (2), the described selectivity hidden Naive Bayes Classifier forecast model F of structure is specially:

$F=\arg \underset{c\⋐C}{max}P\left(c\right)\underset{i=1}{\overset{n}{\Π}}P\left(a_i\right|a_{phi},c)$

Wherein, C is failure classes variables set, and c is failure classes variable, and P (c) is corresponding failure classes variable probability, P (a _i| a _phi, c) be symptom variables a _iwith its most related symptoms variables collection a _phiconditional probability, i=1,2 ..., n, n are symptom variables number.

Described symptom variables a _iwith its most related symptoms variables collection a _phiconditional probability P (a _i| a _phi, c) be specially:

$P\left(a_i\right|a_{phi},c)=\underset{k=1}{\overset{m}{\Σ}}{W}_{i-ik}*P\left(a_i\right|a_{ik},c)$

Wherein, W _i-ikrepresent symptom variables a _iwith its most related symptoms variable a _ikbetween conditional mutual information, P (a _i| a _ik, c) be symptom variables a _iwith its most related symptoms variable a _ikconditional probability, a _ik∈ a _phi, k=1,2 ..., m, m are most related symptoms variable number.

Described symptom variables a _iwith its most related symptoms variable a _ikbetween conditional mutual information W _i-ikbe specially:

$W_{i-ik}=\frac{\underset{a_i,a_{ik},c}{\Σ}P(a_i,a_{ik},c)log\frac{P(a_i,a_{ik},c)}{P\left(a_i|c\right)P\left(a_{ik}|c\right)}}{{\mathrm{\Σ}}_{k=1}^m\underset{a_i,a_{ik},c}{\Σ}P\left(a_i,a_{ik},c\right)log\frac{P\left(a_i,a_{ik},c\right)}{P\left(a_i|c\right)P\left(a_{ik}|c\right)}}$

Wherein, P (a _i, a _ik, c) be a _iand a _ikjoint probability, P (a _i| c) be a _iconditional probability, P (a _ik| c) be a _ikconditional probability.

In addition, in step (3), described classifier parameters comprises symptom variables conditional probability P (a _i| a _j, c) and failure classes variable probability P (c), be specially:

$P\left(a_i\right|a_j,c)=\frac{F(a_i,a_j,c)+1/n}{F(a_j,c)+1}$

$P\left(c\right)=\frac{F\left(c\right)+1/s}{t+1}$

Wherein, F (a _i, a _j, c) symptom variables a in historical data _iand a _jthe number of times simultaneously occurred, F (a _j, c) be symptom variables a in historical data _jthe number of times occurred, F (c) is the number of times of failure classes occurrences in historical data, and t is the quantity of historical data, and s is failure classes variable quantity, and n is symptom variables number.

In instances, the present invention have chosen 350 cases from network history data, and sums up some symptom variables and integrate as attribute, and some failure classes variables sets are class, and details as shown in Figure 4.The corresponding relation of these symptoms and fault as shown in Figure 5.In Fig. 5, A ₁=0 and A ₁=1 represents that interface management state is normal and disconnects, A respectively ₂=0 and A ₂=1 represents that interface operation state is normal and disconnects, A respectively ₃, A ₄, A ₅, A ₆, A ₇, A ₈, A ₉, A ₁₀, these values are that 0 expression property value is normal, are there is fault in 1 expression network.

SHNB most critical, the place that is different from other is most that in the selection of most related symptoms variables set size, the i.e. value of m, m is relevant with historical data.Therefore, the present invention selects different m values to test on failure classes variables set, and Fig. 6 is the accuracy rate that under experiment condition, different m value is corresponding.Time m is 2 as seen from the figure, accuracy rate is the highest, therefore gets m=2 in this example, is each symptom variables and determines two most related symptoms variablees.

The present invention goes out to comprise the most related symptoms variables set of two most related symptoms variablees as hiding father node by calculating cross-correlation Information Selection.Be exemplified below, for symptom variables A ₁, calculate I _p(A ₁| A _j, C), then utilize the algorithm of selected and sorted to sort to mutual information, select two A that mutual information is maximum ₂and A ₃, therefore, symptom variables A ₁corresponding hiding father node S ₁comprise A ₂and A ₃.Like this, actual SHNB model is as shown in Figure 7 obtained when implementing this method.

Realistic model according to Fig. 7, the present invention's test data carries out failure diagnosis, and and NB, TAN, HNB tri-kinds of graders carry out longitudinal comparison, obtain result as shown in Figure 8.When training data is increased to after 300, the accuracy rate of often kind of grader all tends towards stability, but the accuracy rate of SHNB is the highest, reaches 95%, is method the highest in longitudinal comparison.

Meanwhile, the Diagnostic Time of the various method of test further, training data selects 300, and the result obtained as shown in Figure 9, can be known and see, the Diagnostic Time of SHNB is also shorter.

But in actual conditions, often have sample attribute and lose incomplete situation, therefore, the present invention also discuss this situation, and result as shown in Figure 10.Under loss symptom variables or the increasing situation of failure classes variable quantity, the accuracy rate of various algorithm is all in reduction, but it is the fastest that the accuracy rate of TAN declines, contrary, the accuracy rate decline situation of Bayes classifier is much better, therefore, Bayes classifier comprises the hidden Naive Bayes Classifier of selectivity that the present invention mentions good failure tolerance and learning ability.

Those skilled in the art will readily understand; the foregoing is only preferred embodiment of the present invention; not in order to limit the present invention, all any amendments done within the spirit and principles in the present invention, equivalent replacement and improvement etc., all should be included within protection scope of the present invention.

Claims (7)

1. based on a network fault diagnosis method for the hidden Naive Bayes Classifier of selectivity, it is characterized in that, described method comprises:

(1) from network history data storehouse, obtain historical data, comprise symptom variables collection and failure classes variables set;

(2) build selectivity hidden Naive Bayes Classifier forecast model, each symptom variables concentrated according to symptom variables determines corresponding most related symptoms variables collection;

(3) the hidden Naive Bayes Classifier of described selectivity is by training historical data automatic learning to classifier parameters;

(4), when carrying out failure diagnosis, the hidden Naive Bayes Classifier of above-mentioned selectivity is utilized to estimate to obtain corresponding final fault diagnosis result to test data.

2. the method for claim 1, is characterized in that, describedly determines corresponding most related symptoms variables collection for each symptom variables, is specially: for symptom variables A _i, estimate A _iand A _jbetween mutual information, wherein i, j ∈=1,2 ..., n, i ≠ j}, line ordering of going forward side by side, chooses front m maximum symptom variables of mutual information as symptom variables A _imost related symptoms variables collection.

3. method as claimed in claim 2, is characterized in that, the quantity of described most related symptoms variable is determined according to reality test.

4. the method as described in any one of claim 1-3, is characterized in that, described selectivity hidden Naive Bayes Classifier forecast model F is specially:

$F=\arg \underset{c\⋐C}{max}P\left(c\right)\underset{i=1}{\overset{n}{\Π}}P\left(a_i\right|a_{phi},c)$

Wherein, C is failure classes variables set, and c is failure classes variable, and P (c) is corresponding failure classes variable probability, P (a _i| a _phi, c) be symptom variables a _iwith its most related symptoms variables collection a _phiconditional probability, i=1,2 ..., n, n are symptom variables number.

5. method as claimed in claim 4, is characterized in that, described symptom variables a _iwith its most related symptoms variables collection a _phiconditional probability P (a _i| a _phi, c) be specially:

$P\left(a_i\right|a_{phi},c)=\underset{k=1}{\overset{m}{\Σ}}{W}_{i-ik}*P\left(a_i\right|a_{ik},c)$

Wherein, W _i-ikrepresent symptom variables a _iwith its most related symptoms variable a _ikbetween conditional mutual information, P (a _i| a _ik, c) be symptom variables a _iwith its most related symptoms variable a _ikconditional probability, a _ik∈ a _phi, k=1,2 ..., m, m are most related symptoms variable number.

6. method as claimed in claim 5, is characterized in that, described symptom variables a _iwith its most related symptoms variable a _ikbetween conditional mutual information W _i-ikbe specially:

$W_{i-ik}=\frac{\underset{a_i,a_{ik},c}{\Σ}P\left(a_i,a_{ik},c\right)\log \frac{P\left(a_i,a_{ik},c\right)}{P\left(a_i|c\right)P\left(a_{ik}|c\right)}}{{\Σ}_{k=1}^m\underset{a_i,a_{ik},c}{\Σ}P\left(a_i,a_{ik},c\right)\log \frac{P\left(a_i,a_{ik},c\right)}{P\left(a_i|c\right)P\left(a_{ik}|c\right)}}$

Wherein, P (a _i, a _ik, c) be a _iand a _ikjoint probability, P (a _i| c) be a _iconditional probability, P (a _ik| c) be a _ikconditional probability.

7. the method as described in any one of claim 1-3, is characterized in that, in step (3), described classifier parameters comprises symptom variables conditional probability P (a _i| a _j, c) and failure classes variable probability P (c), be specially:

$P\left(a_i|a_j,c\right)=\frac{F(a_i,a_j,c)+1/n}{F(a_j,c)+1}$

$P\left(c\right)=\frac{F\left(c\right)+1/s}{t+1}$

Wherein, F (a _i, a _j, c) symptom variables a in historical data _iand a _jthe number of times simultaneously occurred, F (a _j, c) be symptom variables a in historical data _jthe number of times occurred, F (c) is the number of times of failure classes occurrences in historical data, and t is the quantity of historical data, and s is failure classes variable quantity, and n is symptom variables number.