CN106130780A - A kind of IP network Fault Locating Method based on static Bayesian model - Google Patents

Abstract

The invention discloses a kind of IP network Fault Locating Method based on static Bayesian model, on the one hand by newly-increased suspected malfunctions filtering module, eliminate the network noise impact on result of detection, be greatly promoted fault location accuracy；On the other hand by newly-increased fault pretreatment module, calculate optimum failure collection, greatly reduce the complexity of existing algorithm, thus be applicable to large-scale network topological.

Description

A kind of IP network Fault Locating Method based on static Bayesian model

Technical field

The present invention relates to network failure locating method field, the event of a kind of IP network based on static Bayesian model Barrier localization method.

Background technology

Existing FLT mainly has definitiveness inference technology and uncertain inference technology, definitiveness reasoning skill Art refers to that fault inevitably results in the generation of some symptom, mainly has rule-based, based on model etc.；And it is the most true Qualitative inference technology refers to that the generation of fault causes the generation of some symptom with certain probability, the most popular is based on The FLT of Bayesian network, including FLT based on static Bayesian Network with based on Dynamic Bayesian The FLT of network, it is adaptable to the application scenarios of heterogeneous networks scale.The present invention mainly studies static Bayesian Network Fault Locating Method, current technology implementation is as it is shown in figure 1, mainly include 4 steps:

Step one mainly obtains needs to carry out the objective network topology situation of fault location；Step 2 is to send out in a network Send end-to-end detection and receive result of detection；Step 3 is to set up Bayesian model for result of detection to carry out fault location and divide Analysis；Step 4 is output fault location result.

Prior art mainly has 2 deficiencies:

(1) in real network, there is random noise, the fault location of the result of detection in step 2, step 4 can be tied The accuracy of fruit output causes certain impact.I.e. may may only be affected by influence of noise by some node being judged as fault, And untrue fault；Some is judged as normal node, it is also possible to the detected discovery by influence of noise.

(2) when the Bayesian analysis model that step 3 is set up, all nodes are set all it may happen that fault, and simultaneously The number broken down does not limits, and the probability that actually more than 4 nodes break down simultaneously is less than 0.5%, thus causes existing The computation complexity having the fault location of technology is higher, it is impossible to be applicable to large-scale network topological.

Summary of the invention

It is an object of the invention to provide a kind of IP network Fault Locating Method based on static Bayesian model, existing to solve There is the problem that technical network Fault Locating Method is vulnerable to influence of noise, computation complexity is high.

In order to achieve the above object, the technical solution adopted in the present invention is:

A kind of IP network Fault Locating Method based on static Bayesian model, it is characterised in that: comprise the following steps:

(101), the topology acquisition of objective network: the Fault Management System on upper strata is by equipment in interface collecting device webmaster Between interconnection link information, again set up need to carry out fault location objective network topology；

(102), in objective network, choose the probe node that can send detection bag, and mutually send out between probe node Send end-to-end detection bag, if detection bag can successfully arrive at, then change labelling 0 on path；If can not successfully arrive at, Ze Gai road Labelling 1 on footpath, sends detection bag between all end-to-end paths, and obtains the result of detection in all paths；

(103), according to result of detection obtaining suspected malfunctions set, and filter inaccurate result of detection, detailed process is as follows:

(3.1), due to the existence of IP network noise, detection often presents the result of contradiction, in order to solve asking of network noise Topic, first definition suspected malfunctions threshold value, be shown below:

${\mathrm{\α}}_{X_i}=\frac{\underset{T_j\∈T_{obs}}{\Σ}P\left(T_j\right|X_i)}{\underset{T_j\∈T}{\Σ}P\left(T_j\right|X_i)},$

Wherein T represents the set of whole detective path, T_obsRepresent the set of the detective path having been observed that result of detection, X_iRepresent the state (X of i-th node_i=0 represents normal, X_i=1 represents fault), P (T_j|X_i) it is to work as X_iVisit during normal or fault Survey path T_jThe probability of result of detection success or failure,For nodes X_iIf,Then think nodes X_i For suspected malfunctions, wherein α is the constant between 0 to 1, is generally chosen for 0.5；

(3.2), find out all suspected malfunctions nodes by calculating formula after, only retain at least through a suspected malfunctions node Detection, other detections will be filtered, and the most both filter inaccurate result of detection, greatly reduce again system scale；

(104), calculating optimum failure collection, detailed process is as follows:

(4.1), using the output result of step (3) as one of step initial conditions, optimum failure collection is calculated: assume each The probability of node failures is p, and in network, in N number of node, faulty nodes number is α, the most now nodes X={ X₁, X₂,...,X₂₀The probability distribution of state is represented by P (X)=p^α(1-p)^N-α, variable X_iIt is to obey binomial distribution, it is assumed that F is The node broken down, | F | is the node number that breaks down, thenN is node number；Making great efforts experiment according to shellfish can ?Assume Probability p=10 of nodes break down^-3, network can be calculated and occur The probability of one fault is 0.368063, and the trouble-proof probability of network is 0.367695, and more than 4 general occurs simultaneously Rate is:

If therefore it can be concluded that limit the node number broken down less than 5 simultaneously, then thus bring Error less than 0.37%, but the computation complexity of reasoning will can be substantially reduced；

(4.2), alternative failure collection is produced: setting fault generation number as k, from the beginning of k=1, producing fault number is k Fault combination, to k equal to maximum fault number；

(4.3), fault is selected to combine from alternative set one by one；

(4.4), judge whether current failure combination can solve all result of detections: assume that current failure combination is fault joint Point, it may be judged whether the return result of current detection can be met；

(4.5), failure collection H is added；

(4.6), increase fault number, fault number k is added 1；

(4.7), judge whether to exceed maximum fault number, if current failure combination can explain all result of detections, then This fault is added in combination failure collection H；

(4.8), output set H；

(105), set up Bayesian model algorithm, utilize Bayesian formula to calculate each Hi probability, wherein maximum probability Combine for most possible fault；

H gathers by multiple hypothesis failure collection H_iComposition, each H_iIt is all suspected malfunctions set F_sA subset, and All T can be explained_lIn failed detection；First each F is initialized_sIn node be alternative failure collection h；Secondly, T can be explained_lIn the h of most failed detection join in H as a H_i, simultaneously by this set h from F_sIn remove；So After, expand to alternative failure collection h comprise two F_sInterior joint；

Repeat above procedure, until h comprises k suspected malfunctions node, finally, if being still empty, then by alternative fault in H Set h is directly added in H,

Each H_iThe probability occurred is calculated by below equation:

$\begin{array}{c}P\left(H_i\right)=compute\mathrm{Pr}obability(H_i,T_{obs})\\ =\underset{F_i\∈F}{\Π}P\left(F_i\right)\underset{T_j\∈T}{\Π}P\left(T_j\right|pa\left(T_j\right))\\ =\underset{F_i\∈H_i}{\Π}P(F_i=1)\underset{F_i\∈F\backslash H_i}{\Π}P(F_i=0)\end{array}$

$\underset{T_j\∈T_l}{\Π}P(T_j=1|pa\left(T_j\right))\underset{T_j\∈T\backslash T_l}{\Π}P(T_j=0|pa\left(T_j\right))$

Failure collection H that probability of happening is maximum_iIt is considered as most possible failure collection, if the fault of maximum of probability Gather and be more or less the same with time big probability failure collection probability of happening, then two set are incorporated as most possible failure collection；

The calculating of Bayesian model algorithm includes two parts: (1) finds out suspected malfunctions set F_s, filter falseness detection knot Really；(2) selecting most possible failure collection, in reality, the number of detection is as what network node number linearly increased, because of The computation complexity of this Part I is approximately O (N²), wherein N is the number of network node；Assume suspected malfunctions set F_sIn doubt It is n like the number of malfunctioning node, it is clear that n≤N；And the number that fault occurs is no more than k, therefore, the meter of Part II Calculate complexity and be approximately O (n^k+1), then the time complexity that new algorithm is total is O (N²+n^k+1)；In practical situation, n is to be far smaller than N's, compare O (N²), in most cases O (n^k+1) negligible, therefore Algorithms T-cbmplexity can be approximately O (N²)；

(106), fault location result, the most most possible failure collection are exported.

Described a kind of based on static Bayesian model IP network Fault Locating Method, it is characterised in that: in step (4) (4.3) (4.7) in, for reducing the computation complexity of former Bayesian inference algorithm, the most incrementally selecting several groups, have most can Energy failure collection, calculates the probability of happening often organizing fault, and one group of last select probability maximum is root fault.

The present invention when carrying out the analysis of result of detection, shadow result of detection caused by algorithm screen noise Ring, abandon inaccurate result of detection, promote fault location accuracy, only retain the spy at least through a suspected malfunctions node Survey result, also reduce systematic analysis scale.

The present invention introduces the link of " limiting the nodes simultaneously broken down " in conventional failure positioning analysis, at band The complexity of computational analysis it is substantially reduced in the case of carrying out less error.Such as limit the node number simultaneously broken down not surpass Crossing 5, the error thus brought is less than 0.37%, is substantially reduced computation complexity simultaneously.

Present invention fault reasoning based on greedy strategy algorithm, calculates optimum failure collection, greatly reduces existing algorithm Complexity, strengthen tenability to large-scale network topological fault location.

The invention provides a kind of IP network Fault Locating Method based on static Bayesian model and device, the present invention changes Enter related algorithm and the flow process of existing fault location, had the advantage that

1) the anti-network noise ability of fault diagnosis is enhanced.For the feature of noise network, efficient algorithm mistake is proposed The impact that result of detection is caused by filter network noise, abandons inaccurate result of detection.

2) by reducing the calculating scale of fault diagnosis, reduce the Algorithms T-cbmplexity of fault diagnosis, improve quickly The ability of locating network fault.Introduce the link of " limiting the nodes simultaneously broken down " after analyzing fault signature, save Data are inputted in a large number, it is ensured that a large amount of Reduction Computation scales under conditions of less error during Analysis on Fault Diagnosis.The proposition time Complexity is approximately O (N²) fault reasoning algorithm, optimize the calculation process of optimum failure collection, support at large scale network Fault fast positioning under Tuo Pu.

Accompanying drawing explanation

Fig. 1 is the Fault Locating Method FB(flow block) of prior art static Bayesian Network.

Fig. 2 is the inventive method FB(flow block).

Fig. 3 is step (103) FB(flow block) in the specific embodiment of the invention.

Fig. 4 is step (103) exemplary plot in the specific embodiment of the invention.

Fig. 5 is specific embodiment of the invention step (104) FB(flow block).

Detailed description of the invention

The invention provides a kind of IP network Fault Locating Method based on static Bayesian model, as in figure 2 it is shown, the party Method process is as follows:

(101) the topology acquisition of objective network.The Fault Management System on upper strata can be by interface collecting device webmaster The link information of equipment room interconnection, sets up the objective network topology that need to carry out fault location again.

(102) send end-to-end detection and receive result of detection.

(103) as it is shown on figure 3, by detection filtering module, obtain suspected malfunctions set according to result of detection, and filter not Accurately result of detection, detailed process is as follows:

301) due to the existence of IP network noise, detection often presents the result of contradiction.Such as through the detection of malfunctioning node A T₁The failure of return result, and another is through the detection T of node A₂But return successfully.This phenomenon is likely due to T₂Route Change and cause.If these noises are not processed, then will will necessarily affect final diagnostic result.

In order to solve the problem of network noise, the present invention proposes a kind of method filtering false result of detection.First define Suspected malfunctions threshold value, is shown below.

${\mathrm{\α}}_{X_i}=\frac{\underset{T_j\∈T_{obs}}{\Σ}P\left(T_j\right|X_i)}{\underset{T_j\∈T}{\Σ}P\left(T_j\right|X_i)},0\≤{\mathrm{\α}}_{X_i}\≤1$

For nodes X_iIf,Then think nodes X_iFor suspected malfunctions, wherein α is the constant between 0 to 1.Assume α Be 0.5, each node failure cause process detect failed probability be 0.9,

Example: if in Fig. 4 nodes X₁₀It is out of order, then detection T₂, T₈And T₉Failed knot can be returned with the biggest probability Really.If only T in reality₂Return the result of failure, T₈And T₉All return successfully, thenX₁₀Node is considered It it is spurious glitches.

302), after finding out all suspected malfunctions nodes by calculating formula, only retain at least through a suspected malfunctions node Detection, other detections will be filtered.The most both filter inaccurate result of detection, greatly reduce again system scale.

(104) as it is shown in figure 5, pass through fault pretreatment module, the optimum failure collection of calculating:

401.1) the output result of step 103 it is equal to, as one of the initial conditions of step 402.

401.2) by fault pretreatment module, optimum failure collection is calculated.

The probability assuming each node failures is p, and in network, in N number of node, faulty nodes number is α, the most now Nodes X={ X₁,X₂,...,X₂₀The probability distribution of state is represented by P (X)=p^α(1-p)^N-α, variable X_iIt is to obey binomial to divide Cloth, it is assumed that F is the node broken down, | F | is the node number that breaks down, thenN is node number.According to Shellfish is made great efforts experiment and can obtainAssume Probability p=10 of nodes break down^-3, can count The probability calculating network one fault of generation is 0.368063, and the trouble-proof probability of network is 0.367695, and sends out simultaneously The probability of raw more than 4 is:

If therefore it can be concluded that limit the node number broken down less than 5 simultaneously, then thus bring Error less than 0.37%, but the computation complexity of reasoning will can be substantially reduced.

402) alternative failure collection is produced.

If fault generation number is k, from the beginning of k=1, produce the fault combination that fault number is k, to k equal to maximum Till fault number.For example, it is assumed that suspected malfunctions collection is combined into H={H₁,H₂,...,H_n, then as k=1, alternative failure collection For: { { H₁},{H₂},...{H_n}}；As k=2, alternative failure collection is: { { H₁,H₂},{H₁,H₃},...{H_n-1,H_n}}。

403) fault is selected to combine from alternative set one by one.

404) judge whether current failure combination can solve all result of detections.

Assume that current failure combination is malfunctioning node, if the return result of current detection can be met.

405) failure collection H is added.

406) increase fault number, fault number k is added 1.

407) maximum fault number whether is exceeded

If current failure combination can explain all result of detections, then this fault is added in combination failure collection H.

In step 403-407, for reducing the computation complexity of former Bayesian inference algorithm, it is proposed that a kind of based on greediness The fault reasoning algorithm of strategy.New algorithm the most incrementally selects several groups of most possible failure collection, calculates and often organizes sending out of fault Raw probability, one group of last select probability maximum is root fault.

The false code of algorithm is as follows:

Input: suspected malfunctions set F_s, detection T after filtration_l, fault restriction number k

Output: most possible failure collection H^*

408) output set H

(105) set up Bayesian model, utilize Bayesian formula to calculate each Hi probability, wherein maximum probability for having most Possible breakdown combines.

H gathers by multiple hypothesis failure collection H_iComposition, each H_iIt is all suspected malfunctions set F_sA subset, and All T can be explained_lIn failed detection.First algorithm initializes each F_sIn node be alternative failure collection h. Secondly, T can be explained_lIn the h of most failed detection join in H as a H_i, simultaneously by this set h from F_sMiddle shifting Remove.Then, expand to alternative failure collection h comprise two F_sInterior joint.Repeat to call above step, doubt until h comprises k Like malfunctioning node.Finally, if H being still empty, then alternative failure collection h is directly added in H.

Each H_iThe probability occurred is calculated by below equation:

$\begin{array}{c}P\left(H_i\right)=compute\mathrm{Pr}obability(H_i,T_{obs})\\ =\underset{F_i\∈F}{\Π}P\left(F_i\right)\underset{T_j\∈T}{\Π}P\left(T_j\right|pa\left(T_j\right))\\ =\underset{F_i\∈H_i}{\Π}P(F_i=1)\underset{F_i\∈F\backslash H_i}{\Π}P(F_i=0)\end{array}$

$\underset{T_j\∈T_l}{\Π}P(T_j=1|pa\left(T_j\right))\underset{T_j\∈T\backslash T_l}{\Π}P(T_j=0|pa\left(T_j\right))$

Failure collection H that probability of happening is maximum_iIt it is considered as most possible failure collection.If the fault of maximum of probability Gather and be more or less the same with time big probability failure collection probability of happening, then two set are incorporated as most possible failure collection.

The calculating of new algorithm mainly includes two parts: (1) finds out suspected malfunctions set F_s, filter false result of detection；(2) Select most possible failure collection.In reality, the number of detection is as what network node number linearly increased, and therefore first The computation complexity of part is approximately O (N²), wherein N is the number of network node.Assume suspected malfunctions set F_sMiddle suspected malfunctions The number of node is n, it is clear that n≤N.And the number that fault occurs is no more than k, therefore, the calculating of Part II is complicated Degree is approximately O (n^k+1).The time complexity that then new algorithm is total is O (N²+n^k+1).In practical situation, n is far smaller than N, phase Than O (N²), in most cases O (n^k+1) negligible, therefore Algorithms T-cbmplexity can be approximately O (N²)。

(106) output fault location result, the most most possible failure collection.

Claims (2)

1. an IP network Fault Locating Method based on static Bayesian model, it is characterised in that: comprise the following steps:

(101), the topology acquisition of objective network: the Fault Management System on upper strata is mutual by equipment room in interface collecting device webmaster The link information of connection, sets up the objective network topology that need to carry out fault location again；

(102), in objective network, the probe node that can send detection bag, and mutual transmitting terminal between probe node are chosen To end detection bag, if detection bag can successfully arrive at, then changing labelling 0 on path；If can not successfully arrive at, then changing on path Labelling 1, sends detection bag between all end-to-end paths, and obtains the result of detection in all paths；

(103), according to result of detection obtaining suspected malfunctions set, and filter inaccurate result of detection, detailed process is as follows:

(3.1), due to the existence of IP network noise, detection often presents the result of contradiction, in order to solve the problem of network noise, first First definition suspected malfunctions threshold value, is shown below:

${\mathrm{\α}}_{X_i}=\frac{\underset{T_j\∈T_{obs}}{\Σ}P\left(T_j\right|X_i)}{\underset{T_j\∈T}{\Σ}P\left(T_j\right|X_i)},$

Wherein T represents the set of whole detective path, T_obsRepresent the set of the detective path having been observed that result of detection, X_iRepresent State (the X of i-th node_i=0 represents normal, X_i=1 represents fault), P (T_j|X_i) it is to work as X_iDetective path during normal or fault T_jThe probability of result of detection success or failure,For nodes X_iIf,Then think nodes X_iFor doubtful Fault, wherein α is the constant between 0 to 1, is generally chosen for 0.5；

(3.2), find out all suspected malfunctions nodes by calculating formula after, only retain the spy at least through a suspected malfunctions node Surveying, other detections will be filtered, and the most both filter inaccurate result of detection, greatly reduce again system scale；

(104), calculating optimum failure collection, detailed process is as follows:

(4.1), using the output result of step (3) as one of step initial conditions, optimum failure collection is calculated: assume each node Out of order probability is p, and in network, in N number of node, faulty nodes number is α, the most now nodes X={ X₁,X₂,..., X₂₀The probability distribution of state is represented by P (X)=p^α(1-p)^N-α, variable X_iIt is to obey binomial distribution, it is assumed that F is for occurring event The node of barrier, | F | is the node number that breaks down, thenN is node number；Make great efforts experiment according to shellfish can obtainAssume Probability p=10 of nodes break down^-3, network can be calculated and occur one The probability of individual fault is 0.368063, and the trouble-proof probability of network is 0.367695, and the probability of more than 4 occurs simultaneously For:

If therefore it can be concluded that limit the node number broken down less than 5 simultaneously, then the mistake thus brought The computation complexity of reasoning less than 0.37%, but can will be substantially reduced by difference；

(4.2), alternative failure collection is produced: set fault generation number as k, from the beginning of k=1, produce the event that fault number is k Barrier combination, being equal to maximum fault number to k；

(4.3), fault is selected to combine from alternative set one by one；

(4.4), judge whether current failure combination can solve all result of detections: assume that current failure combination is malfunctioning node, Judge whether to meet the return result of current detection；

(4.5), failure collection H is added；

(4.6), increase fault number, fault number k is added 1；

(4.7), judge whether to exceed maximum fault number, if current failure combination can explain all result of detections, then this Fault is added in combination failure collection H；

(4.8), output set H；

(105), setting up Bayesian model algorithm, utilize Bayesian formula to calculate each Hi probability, wherein maximum probability is Likely fault combination；

H gathers by multiple hypothesis failure collection H_iComposition, each H_iIt is all suspected malfunctions set F_sA subset, and permissible Explain all T_lIn failed detection；First each F is initialized_sIn node be alternative failure collection h；Secondly, by energy Enough explain T_lIn the h of most failed detection join in H as a H_i, simultaneously by this set h from F_sIn remove；Then, will Alternative failure collection h expands to comprise two F_sInterior joint；

Repeat above procedure, until h comprises k suspected malfunctions node, finally, if being still empty, then by alternative failure collection in H H is directly added in H,

Each H_iThe probability occurred is calculated by below equation:

$\begin{array}{c}P\left(H_i\right)=compute\mathrm{Pr}obability(H_i,T_{obs})\\ =\underset{F_i\∈F}{\Π}P\left(F_i\right)\underset{T_j\∈T}{\Π}P\left(T_j\right|pa\left(T_j\right))\\ =\underset{F_i\∈H_i}{\Π}P(F_i=1)\underset{F_i\∈F\backslash H_i}{\Π}P(F_i=0)\end{array}$

$\underset{T_j\∈T_l}{\Π}P(T_j=1|pa\left(T_j\right))\underset{T_j\∈T\backslash T_l}{\Π}P(T_j=0|pa\left(T_j\right))$

Failure collection H that probability of happening is maximum_iIt is considered as most possible failure collection, if the failure collection of maximum of probability It is more or less the same with secondary big probability failure collection probability of happening, then two set is incorporated as most possible failure collection；

The calculating of Bayesian model algorithm includes two parts: (1) finds out suspected malfunctions set F_s, filter false result of detection；(2) Selecting most possible failure collection, in reality, the number of detection is as what network node number linearly increased, and therefore first The computation complexity of part is approximately O (N²), wherein N is the number of network node；Assume suspected malfunctions set F_sMiddle suspected malfunctions The number of node is n, it is clear that n≤N；And the number that fault occurs is no more than k, therefore, the calculating of Part II is complicated Degree is approximately O (n^k+1), then the time complexity that new algorithm is total is O (N²+n^k+1)；In practical situation, n is far smaller than N, phase Than O (N²), in most cases O (n^k+1) negligible, therefore Algorithms T-cbmplexity can be approximately O (N²)；

(106), fault location result, the most most possible failure collection are exported.

A kind of IP network Fault Locating Method based on static Bayesian model the most according to claim 1, its feature exists In: in (4.3) (4.7) of step (4), for reducing the computation complexity of former Bayesian inference algorithm, the most incrementally select Going out several groups of most possible failure collection, calculate the probability of happening often organizing fault, a group of last select probability maximum is root event Barrier.