CN103533571B - Fault-tolerant event detecting method based on temporal voting strategy - Google Patents

Abstract

The invention discloses a kind of fault-tolerant event detecting method based on temporal voting strategy, including the following steps that order performs: (1) each node obtains the detection data of self; (2) based on rectangular histogram, detection data are carried out fault-tolerant processing; (3) based on the dependency of attribute, detection data are carried out local error detection, distinguish LF node and preliminary Normal node; (4) preliminary Normal node is carried out local event decision-making, distinguish LE node and Normal node; (5) to LF node and LE node, the credibility of its neighbor node is calculated; (6) vote based on credibility and distance weighting, it is judged that node is Event node or Normal node. Fault-tolerant event detecting method based on temporal voting strategy provided by the invention, has higher event detection rate and error correction rate, has relatively low event rate of false alarm and mistake introducing rate, and has less ability consumption.

Description

Fault-tolerant event detecting method based on temporal voting strategy

Technical field

The present invention relates to fault-tolerant event detecting method in a kind of wireless sensor network, particularly relate to a kind of fault-tolerant event detection based on temporal voting strategy (Fault-tolerantEventDetectionAlgorithmbasedonVoting is called for short FEDAV) method.

Background technology

Wireless sensor network (WirelessSensorNetworks, it is called for short WSN) it is the wireless self-organization network being made up of one group of sensor node, it is used for gathering in real time data and the information of various monitored object, is analyzed and processed, testing result is supplied to the user of network.

Event detection in WSN, it is simply that refer to region or border that the event that detects occurs. For having the event (such as, when forest fire occurs, the high temperature of regional area, dry etc.) of special characteristic, when this kind of event occurs, the sampled value of respective sensor can deviate the sampled value of normal condition. Therefore, if the sampled value of sensor reaches an event threshold, it is believed that the monitored area of sensor there occurs event, is called event area; Otherwise it is assumed that this region occurs without event, it is referred to as normal region.

But, the unreliability of rugged environment and sensor itself, often lead to sensor and produce wrong data or noise data, thus affecting the detection to event area of the whole network. Therefore event detection must take into fault-tolerant processing, and namely when node makes a mistake, or under the interference of noise data, incident Detection Algorithm still has and detects performance preferably. Because the event that environmental factors causes or Deviant Behavior are space correlation, the fault of node or noise data are then stochastic independences, so the fault-tolerant processing of existing incident Detection Algorithm, mostly depend on the spatial coherence of sensing data, undertaken fault-tolerant by the swapping data of neighbor node, as based on theoretical fault-tolerant of Bayesian, fault-tolerant based on mobile median, based on moving the fault-tolerant of average, based on the tolerant fail algorithm etc. of ballot.

Tolerant fail algorithm (KrishnamachariB based on Bayesian, IyengarS.DistributedOTDSAlgorithmsforFault-TolerantEvent RegionDetectioninWirelessSensorNetworks [J] .IEEETrans, Computers, 2004, 53 (3): 241-250) it is the tolerant fail algorithm of a kind of classics, although having good fault freedom, but only only in accordance with whether the neighbours number consistent with oneself event court verdict reaches half, or the size of posterior probability revises the event decision result of oneself, accuracy in detection is not high, because when spatial redundancy information deficiency, such as the node for event boundaries place, fault freedom is poor. and based on the tolerant fail algorithm (ChenJ of temporal voting strategy, KherS, SomaniA.DistributedFaultDetectionofWirelessSensorNetwork s [C] .Procofthe2006WorkshoponDependabilityIssuesinWirelessAdH ocNetworksandSensorNetworks, 2006:65-71) do not consider the distance factor between neighbor node, and the fault-tolerant effect of Centroid is different by different far and near neighbor node.

Fault-tolerant processing based on spatial coherence has good fault freedom, but generally requires frequent exchange data between node, and energy expenditure is very big. And, if the data of detection itself comprise mistake, based on local event decision-making and neighbours' collaborative process, the testing result of mistake can be produced. It should also be seen that, in wireless sensor and actor networks, each node has different duties and reliability, and therefore, in ballot processes, the ballot weight of each neighbor node also should difference.

Summary of the invention

Goal of the invention: in order to overcome the deficiencies in the prior art, the present invention is from the temporal correlation of node its data, it is considered to the fault-tolerant processing in node, adopts the fault-tolerant processing based on rectangular histogram and Attribute Correlation; On this basis, based on the spatial coherence of internodal data, provide a kind of fault-tolerant incident Detection Algorithm based on temporal voting strategy, have that event detection is effective, fault-tolerant processing is effective and the advantage such as energy expenditure is low.

Technical scheme: for achieving the above object, the technical solution used in the present invention is:

Based on the fault-tolerant event detecting method of temporal voting strategy, the state of definition node is as follows:

Normal: represent node S_localMonitored area in without event occur, node S_localDo not break down and mistake;

LF: represent node S_localIt is likely to occur fault or wrong data, for node S_localLocal event judge in;

LE: represent node S_localEvent it is likely to occur, for node S in monitored area_localLocal event judge in;

Event: represent node S_localEvent occurs, for node S in monitored area_localIn ballot decision-making;

Fault: represent node S_localBreak down event, and can not correct, for node S_localIn ballot decision-making;

Including the following steps that order performs:

(1) each node obtains the detection data of self;

(2) based on rectangular histogram, detection data are carried out fault-tolerant processing;

(3) based on the dependency of attribute, detection data are carried out local error detection, distinguish LF node and preliminary Normal node;

(4) preliminary Normal node is carried out local event decision-making, distinguish LE node and Normal node;

(5) to LF node and LE node, the credibility of its neighbor node is calculated;

(6) vote based on credibility and distance weighting, it is judged that node is Event node or Normal node.

Concrete, in described step (2), when detection data being carried out fault-tolerant processing based on rectangular histogram, the acquisition methods of data center is as follows:

(21) for data x₁,x₂,��,x_n, according to each data genaration histogram structure, obtain k packet g₁,g₂,��,g_i,��,g_k;

(22) size of definition i-th packet | g_i| the number of the data comprised in being grouped for i-th, obtain g according to the descending arrangement of the size of each packet_k',��,g_i',��,g₂',g₁';

(23) a given mean data coverage m₀, each group of data of descending accumulation, until group g_r' meet following formula:

$\frac{\underset{i=r}{\overset{k}{\mathrm{\Σ}}}{g}_i^{\′}}{\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{g}_i^{\′}}\≥m_0$

m₀Value is relevant with the signal to noise ratio of data, and signal to noise ratio is more little, m₀Value is more little, to filter more noise data; If the generation probability of noise data is p, then m₀Should be less than 1-p;

(24) to organize g_r',��,g_k' the data that comprise, calculate data center

$p=\underset{i=r}{\overset{k}{\mathrm{\Σ}}}\left|g_i^{\′}\right|$

${\stackrel{\‾}{x}}_h=\frac{1}{p}\mathrm{\Σ}{x}_j$

Hereini=r,r+1,��,k��

Node wrong data and noise data are small probability event. This method filters out small probability data, has both decreased the impact of noise data, considers again more normal data in the calculation.

Concrete, in described step (3), detection data are carried out local error detection by the dependency based on attribute, distinguish LF node and preliminary Normal node, method particularly includes: the initial relevance of two strong correlation attribute X and Y of note node is R₀(X, Y), real-time dependency is R (X, Y), for a dependency irrelevance threshold xi, if the intensity of variation of Attribute Correlation meets following formula:

$\left|\frac{R(X,Y)-R_0(X,Y)}{R_0(X,Y)}\right|<\mathrm{\&xi;}$

Then thinking that the dependency of X and Y is not changed in, node is preliminary Normal node; Otherwise it is assumed that the dependency of X and Y there occurs change, node is Event node.

If the expected value of E (X), E (Y) respectively two attribute X and Y, the variance that D (X), D (Y) are attribute X and Y, the linear dependence definition of X and Y is as follows:

$R(X,Y)=\frac{E\left\{\right[X-E\left(X\right)\left]\right[Y-E\left(Y\right)\left]\right\}}{\sqrt{D\left(X\right)}\sqrt{D\left(Y\right)}}$

For node data, following formula is adopted to carry out approximate calculation:

$R(X,Y)=\frac{\mathrm{\&Sigma;}(X-\stackrel{\&OverBar;}{X})(Y-\stackrel{\&OverBar;}{Y})}{\sqrt{\mathrm{\&Sigma;}{(X-\stackrel{\&OverBar;}{X})}^2\&times;\sqrt{\mathrm{\&Sigma;}{(Y-\stackrel{\&OverBar;}{Y})}^2}}}$

Consider that node storage space is extremely limited, adopt cumulative mode to calculate the dependency of X and Y:

$R(X,Y)=\frac{\mathrm{\&Sigma;XY}-\frac{\mathrm{\&Sigma;X\&Sigma;Y}}{n}}{\sqrt{{\mathrm{\&Sigma;X}}^2-\frac{{\left(\mathrm{\&Sigma;X}\right)}^2}{n}}\&times;\sqrt{\mathrm{\&Sigma;}{Y}^2-\frac{{\left(\mathrm{\&Sigma;Y}\right)}^2}{n}}}$

Concrete, in described step (4), preliminary Normal node is carried out local event decision-making, distinguishes LE node and Normal node, method particularly includes:

If the average value that attribute X is under normal circumstances is M_n, the average value when event occurs is M_e, M_e>M_n, the threshold value x of definition event generation_thFor:

$x_{\mathrm{th}}=\frac{1}{2}(M_n+M_e)$

Real-time value x for attribute X_i, the binary judgment rule that event occurs is:

$B_e=\left\{\begin{array}{cc}1& x_i>x_{\mathrm{th}}\\ 0& x_i\&le;x_{\mathrm{th}}\end{array}\right.$

Wherein, B_e=1 represents that node is LE node, B_e=0 represents that node is Normal node.

The credibility of node, the confidence level of representation node current decision, reflect node current operating state. The concordance that node credibility detects data by node current detection data and history embodies. Concrete, in described step (5), the credibility C of node is calculated by following formula:

$C=1-\frac{|R-R_0|}{\left|R\right|+\left|R_0\right|}$

Wherein, R₀(X, Y) represents attribute X and the initial relevance of attribute Y, and R (X, Y) represents the real-time dependency of attribute X and attribute Y. C is the closer to 1, it was shown that the dependency between attribute X and Y changes more little as time goes by, and the probability of this one malfunctions is more low, thus the credibility of node is more high; Otherwise, C is the closer to 0, and node credibility is more low.

Concrete, in described step (6), vote based on credibility and distance weighting, it is judged that node is Event node or Normal node, method particularly includes:

(61) definition S_localFor present node, S_iFor S_localA hop neighbor node, then:

Distance present node S_localNearer node is more similar to its state, and ballot weight also should be more big, therefore adopts anti-distance weighting rule to vote. Node S_iRelative to node S_localAnti-distance weighting InvDistWeight_ilFor:

${\mathrm{InvDist}}_{\mathrm{il}}=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\mathrm{Dis}\tan \mathrm{ce}(S_i,S_{\mathrm{local}})-\mathrm{Dis}\tan \mathrm{ce}(S_i,S_{\mathrm{local}})$

${\mathrm{InvDistWeight}}_{\mathrm{il}}=\frac{{\mathrm{InvDist}}_{\mathrm{il}}}{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{InvDist}}_{\mathrm{il}}}$

Wherein, Distance (S_i,S_local) it is node S_iWith node S_localBetween the Euclidean distance of coordinate;

Node S_iTo node S_localBallot value V_iFor:

V_i=C_i��InvDistWeight_il

Wherein, C represents the credibility of node;

(62) at present node S_localAll hop neighbor node S_iIn, calculate and there is identical local event result of decision ballot value sum, remember that the ballot value sum that all local decision-makings are LE node is V_e, the ballot value sum that all local decision-makings are Normal node is V_n, calculate V_eAnd V_nBetween difference degree V_diffFor:

$V_{\mathrm{diff}}=\frac{|V_e-V_n|}{|V_e+V_n|}$

(64) by V_diffCompare with the event boundaries threshold �� set:

Table 1FEDAV algorithm ballot processes decision rule

According to upper table decision node S_localEnd-state.

As node S_localWhen being positioned at event boundaries region and cause information of neighbor nodes deficiency, it is impossible to node S is corrected in the ballot relying on neighbours_localError condition. For this situation, according to Tobler First Law, different from the dependency of present node apart from different neighbor nodes, therefore adopt the space interpolation algorithm based on inverse distance weight, by the data estimation node S of neighbor node_localDetected value

Wherein, d_iFor node S_iTo node S_localBetween distance, p is the values of powers of distance;

Then withAs node S_localReal-time value, adopt the method for step (4) to carry out local event decision-making, and result step (4) distinguished be as node S_localFinal decision result.

For the node S not having effective neighbor node to vote_local(such as, node S_localThe local event result of decision of all neighbor nodes be LF), if the end-state processed through ballot is unsettled condition, then node S is set_localState be Fault.

Beneficial effect: the fault-tolerant event detecting method based on temporal voting strategy provided by the invention, has higher event detection rate and error correction rate, has relatively low event rate of false alarm and mistake introducing rate, and has less ability consumption.

Accompanying drawing explanation

Fig. 1 is the flow chart of the present invention;

Fig. 2 is interior joint state transition diagram of the present invention;

Fig. 3 is Intel's Berkeley laboratory event simulated scenario, first group of data source;

Fig. 4 is the event simulation scene that 1024 nodes constitute in sensor network, second group of data source;

Fig. 5 is the contrast (adopting first group of data) of FEDAV, OTDS and SFEDA event recognition rate;

Fig. 6 is the contrast (adopting second group of data) of FEDAV, OTDS and SFEDA event recognition rate;

Fig. 7 is the contrast (adopting first group of data) of FEDAV, OTDS and SFEDA event rate of false alarm;

Fig. 8 is the contrast (adopting second group of data) of FEDAV, OTDS and SFEDA event rate of false alarm;

Fig. 9 is the contrast (adopting first group of data) of FEDAV, OTDS and SFEDA error correction rate;

Figure 10 is the contrast (adopting second group of data) of FEDAV, OTDS and SFEDA error correction rate;

Figure 11 is the contrast (adopting first group of data) of FEDAV, OTDS and SFEDA mistake introducing rate;

Figure 12 is the contrast (adopting second group of data) of FEDAV, OTDS and SFEDA mistake introducing rate;

Figure 13 is the contrast (adopting first group of data) of FEDAV, OTDS and SFEDA information exchange frequency;

Figure 14 is the contrast (adopting second group of data) of FEDAV, OTDS and SFEDA information exchange frequency.

Detailed description of the invention

Below in conjunction with accompanying drawing, the present invention is further described.

Being a kind of fault-tolerant event detecting method based on temporal voting strategy as shown in Figure 1 and Figure 2, the state of definition node is as follows:

Normal: represent node S_localMonitored area in without event occur, node S_localDo not break down and mistake;

LF: represent node S_localIt is likely to occur fault or wrong data, for node S_localLocal event judge in;

LE: represent node S_localEvent it is likely to occur, for node S in monitored area_localLocal event judge in;

Event: represent node S_localEvent occurs, for node S in monitored area_localIn ballot decision-making;

Fault: represent node S_localBreak down event, and can not correct, for node S_localIn ballot decision-making;

Including the following steps that order performs:

(1) each node obtains the detection data of self;

(2) based on rectangular histogram, detection data are carried out fault-tolerant processing;

(3) based on the dependency of attribute, detection data are carried out local error detection, distinguish LF node and preliminary Normal node;

(4) preliminary Normal node is carried out local event decision-making, distinguish LE node and Normal node;

(5) to LF node and LE node, the credibility of its neighbor node is calculated;

(6) vote based on credibility and distance weighting, it is judged that node is Event node or Normal node.

Concrete, in described step (2), when detection data being carried out fault-tolerant processing based on rectangular histogram, the acquisition methods of data center is as follows:

(21) for data x₁,x₂,��,x_n, according to each data genaration histogram structure, obtain k packet g₁,g₂,��,g_i,��,g_k;

(22) size of definition i-th packet | g_i| the number of the data comprised in being grouped for i-th, obtain g according to the descending arrangement of the size of each packet_k',��,g_i',��,g₂',g₁';

(23) a given mean data coverage m₀, each group of data of descending accumulation, until group g_r' meet following formula:

$\frac{\underset{i=r}{\overset{k}{\mathrm{\&Sigma;}}}{g}_i^{\&prime;}}{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{g}_i^{\&prime;}}\&GreaterEqual;m_0$

m₀Value is relevant with the signal to noise ratio of data, and signal to noise ratio is more little, m₀Value is more little, to filter more noise data; If the generation probability of noise data is p, then m₀Should be less than 1-p;

(24) to organize g_r',��,g_k' the data that comprise, calculate data center

$p=\underset{i=r}{\overset{k}{\mathrm{\&Sigma;}}}\left|g_i^{\&prime;}\right|$

${\stackrel{\&OverBar;}{x}}_h=\frac{1}{p}\mathrm{\&Sigma;}{x}_j$

Hereini=r,r+1,��,k��

Node wrong data and noise data are small probability event. This method filters out small probability data, has both decreased the impact of noise data, considers again more normal data in the calculation.

Concrete, in described step (3), detection data are carried out local error detection by the dependency based on attribute, distinguish LF node and preliminary Normal node, method particularly includes: the initial relevance of two strong correlation attribute X and Y of note node is R₀(X, Y), real-time dependency is R (X, Y), for a dependency irrelevance threshold xi, if the intensity of variation of Attribute Correlation meets following formula:

$\left|\frac{R(X,Y)-R_0(X,Y)}{R_0(X,Y)}\right|<\mathrm{\&xi;}$

Then thinking that the dependency of X and Y is not changed in, node is preliminary Normal node; Otherwise it is assumed that the dependency of X and Y there occurs change, node is Event node.

If the expected value of E (X), E (Y) respectively two attribute X and Y, the variance that D (X), D (Y) are attribute X and Y, the linear dependence definition of X and Y is as follows:

$R(X,Y)=\frac{E\left\{\right[X-E\left(X\right)\left]\right[Y-E\left(Y\right)\left]\right\}}{\sqrt{D\left(X\right)}\sqrt{D\left(Y\right)}}$

For node data, following formula is adopted to carry out approximate calculation:

$R(X,Y)=\frac{\mathrm{\&Sigma;}(X-\stackrel{\&OverBar;}{X})(Y-\stackrel{\&OverBar;}{Y})}{\sqrt{\mathrm{\&Sigma;}{(X-\stackrel{\&OverBar;}{X})}^2\&times;\sqrt{\mathrm{\&Sigma;}{(Y-\stackrel{\&OverBar;}{Y})}^2}}}$

Consider that node storage space is extremely limited, adopt cumulative mode to calculate the dependency of X and Y:

$R(X,Y)=\frac{\mathrm{\&Sigma;XY}-\frac{\mathrm{\&Sigma;X\&Sigma;Y}}{n}}{\sqrt{{\mathrm{\&Sigma;X}}^2-\frac{{\left(\mathrm{\&Sigma;X}\right)}^2}{n}}\&times;\sqrt{\mathrm{\&Sigma;}{Y}^2-\frac{{\left(\mathrm{\&Sigma;Y}\right)}^2}{n}}}$

Concrete, in described step (4), preliminary Normal node is carried out local event decision-making, distinguishes LE node and Normal node, method particularly includes:

If the average value that attribute X is under normal circumstances is M_n, the average value when event occurs is M_e, M_e>M_n, the threshold value x of definition event generation_thFor:

$x_{\mathrm{th}}=\frac{1}{2}(M_n+M_e)$

Real-time value x for attribute X_i, the binary judgment rule that event occurs is:

$B_e=\left\{\begin{array}{cc}1& x_i>x_{\mathrm{th}}\\ 0& x_i\&le;x_{\mathrm{th}}\end{array}\right.$

Wherein, B_e=1 represents that node is LE node, B_e=0 represents that node is Normal node.

The credibility of node, the confidence level of representation node current decision, reflect node current operating state. The concordance that node credibility detects data by node current detection data and history embodies. Concrete, in described step (5), the credibility C of node is calculated by following formula:

$C=1-\frac{|R-R_0|}{\left|R\right|+\left|R_0\right|}$

Wherein, R₀(X, Y) represents attribute X and the initial relevance of attribute Y, and R (X, Y) represents the real-time dependency of attribute X and attribute Y. C is the closer to 1, it was shown that the dependency between attribute X and Y changes more little as time goes by, and the probability of this one malfunctions is more low, thus the credibility of node is more high; Otherwise, C is the closer to 0, and node credibility is more low.

Concrete, in described step (6), vote based on credibility and distance weighting, it is judged that node is Event node or Normal node, method particularly includes:

(61) definition S_localFor present node, S_iFor S_localA hop neighbor node, then:

Distance present node S_localNearer node is more similar to its state, and ballot weight also should be more big, therefore adopts anti-distance weighting rule to vote. Node S_iRelative to node S_localAnti-distance weighting InvDistWeight_ilFor:

${\mathrm{InvDist}}_{\mathrm{il}}=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}\mathrm{Dis}\tan \mathrm{ce}(S_i,S_{\mathrm{local}})-\mathrm{Dis}\tan \mathrm{ce}(S_i,S_{\mathrm{local}})$

${\mathrm{InvDistWeight}}_{\mathrm{il}}=\frac{{\mathrm{InvDist}}_{\mathrm{il}}}{\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{InvDist}}_{\mathrm{il}}}$

Wherein, Distance (S_i,S_local) it is node S_iWith node S_localBetween the Euclidean distance of coordinate;

Node S_iTo node S_localBallot value V_iFor:

V_i=C_i��InvDistWeight_il

Wherein, C represents the credibility of node;

(62) at present node S_localAll hop neighbor node S_iIn, calculate and there is identical local event result of decision ballot value sum, remember that the ballot value sum that all local decision-makings are LE node is V_e, the ballot value sum that all local decision-makings are Normal node is V_n, calculate V_eAnd V_nBetween difference degree V_diffFor:

$V_{\mathrm{diff}}=\frac{|V_e-V_n|}{|V_e+V_n|}$

(64) by V_diffCompare with the event boundaries threshold �� set, according to table 1 decision node S_localEnd-state.

As node S_localWhen being positioned at event boundaries region and cause information of neighbor nodes deficiency, it is impossible to node S is corrected in the ballot relying on neighbours_localError condition. For this situation, according to Tobler First Law, different from the dependency of present node apart from different neighbor nodes, therefore adopt the space interpolation algorithm based on inverse distance weight, by the data estimation node S of neighbor node_localDetected value

Wherein, d_iFor node S_iTo node S_localBetween distance, p is the values of powers of distance;

Then withAs node S_localReal-time value, adopt the method for step (4) to carry out local event decision-making, and result step (4) distinguished be as node S_localFinal decision result.

For the node S not having effective neighbor node to vote_local(such as, node S_localThe local event result of decision of all neighbor nodes be LF), if the end-state processed through ballot is unsettled condition, then node S is set_localState be Fault.

Below in conjunction with experiment, the present invention is further illustrated.

Experiment have employed two groups of data and algorithm carried out simulation analysis. First group, we adopt the real wireless sensor network data of the offer of Intel's Berkeley laboratory; These real node data have good temporal correlation, have strong correlation between part physical amount, can substantially embody the Detection results of algorithm. But, because the sensor node that this laboratory is disposed only has 54, and in practical application, the deployment density of WSN is very big. For this, we adopt second group of data, simulate the sensor network that 1024 nodes are constituted, better to embody the behavior pattern of algorithm.

First group of data are to be produced by 54 nodes being deployed in Intel's Berkeley laboratory, laboratory be sized to 42m �� 32m, node data collection be spaced apart 31 seconds, the physical quantity of collection includes temperature, relative humidity, illumination, node voltage. There is stronger linear dependence between temperature and relative humidity in physical quantity, therefore, the analysis of Attribute Correlation can be carried out accordingly.

In simulation, we adopt the normal data of some day to process, and noise data and event data adopt the mode of synthesis to add. The probability of happening of noise data and wrong data assumes conformance with standard normal distribution. Event data, by simulating, at certain point A, (x, y) produces an event, and centered by this point, with R for the circle of radius, for event area, the detection data in this region are apparently higher than other regions. Modeling event occur scene as it is shown on figure 3, wherein, the center that node 1 occurs for event, with curve line around region be event area.

Second group of data, we simulate the sensor network that 1024 nodes are constituted, and are distributed in the spatial dimension of 32 �� 32. Adopt a space time correlation model, generate analog data. The data that each node produces include two attributes, and they have strong linear dependence. The analog form of noise data and event data is with first group of data. Fig. 4 is the simulated scenario that an event occurs, and wherein * represents node mistake, represents that node is normal ,+represent that Node Events occurs.

Experiment adopts FEDAV, OTDS(OptimalThresholdDecisionScheme, optimal threshold decision method) and SFEDA(Spatio-temporalbasedFault-tolerantEventDetectionAl gorithm, fault-tolerant incident Detection Algorithm based on temporal correlation) above-mentioned two groups of data carry out fault-tolerant event detection by three kinds of detection methods, and result is such as shown in Fig. 5��14.

Fig. 5, Fig. 6 show when node error probability changes, the contrast situation of the event recognition rate of FEDAV, OTDS and SFEDA algorithm. As seen from the figure, when node error rate is relatively low, the event recognition rate difference of three kinds of algorithms is little. But along with the increase of error rate, the event recognition hydraulic performance decline of OTDS and SFEDA algorithm is quickly; And the detection Performance comparision of FEDAV algorithm is stable. This is because along with the increase of error rate, adopt OTDS and SFEDA algorithm, at the borderline region of event to inside event area, increasing node can not meet the condition of half ballot, thus the generation of event can not effectively be identified.

And that FEDAV algorithm have employed based on Attribute Correlation is fault-tolerant, error node can be distinguished from normal node and event node, be not involved in the ballot to other nodes, thus avoiding the error node impact on ballot; Furthermore, when ballot, it is contemplated that the credibility of node decision-making, make voting results truly reflect the duty of node; Finally, in event boundaries region, FEDAV algorithm depends on the local decision-making of safe node, does not rely on ballot and processes, so on the whole, the event detection Performance comparision of FEDAV algorithm is stable. Just because of SFEDA algorithm is also not entirely dependent on most ballot, so event recognition rate is slightly better than OTDS algorithm.

In Fig. 6, the change ratio of curve is shallower, illustrates along with the increase of node deployment density, the redundancy of space nodes increases, the event detection performance of three kinds of algorithms all becomes more stable, it can be seen that the event detection rate of FEDAV algorithm is still substantially better than OTDS and SFEDA algorithm.

Fig. 7, Fig. 8 show when node error probability changes, the contrast situation of the event rate of false alarm of FEDAV, OTDS and SFEDA algorithm, and Fig. 7 adopts first group of data, and Fig. 8 adopts second group of data to realize. In event area, the decision-making of node meets the condition of most voting mechanism, can effectively identify event node, and rate of false alarm is only small, and therefore, the rate of false alarm of three kinds of algorithms is all relatively low.

But in event boundaries region, OTDS and SFEDA algorithm adopts majority vote rule, it is easy to the ambiguousness of decision-making occurs, because the ballot value of normal node and event node is more or less the same, is unsatisfactory for the condition that majority is voted. And FEDAV algorithm does not adopt majority to vote, the present node that credibility is higher, rely on the judged result of self; For the boundary node already at error condition, according to the data estimation measured value of neighbor node, then carry out event detection. So, just decrease the generation of erroneous judgement, reduce rate of false alarm.

Fig. 9 and Figure 10 shows when node error probability changes, the contrast of FEDAV, OTDS and SFEDA algorithmic error adjusted rate. It can be seen that for error correction rate, the error correction rate of FEDAV algorithm is more stable, even if error rate increases, still maintain good fault freedom. This is due to the fault-tolerant processing based on Attribute Correlation, decreases the impact of error node; And in voting process, it is contemplated that the credibility of node decision-making, so, in ballot, the node that majority is with a high credibility determine the result of ballot, can correctly detect node mistake and carry out error correction. And OTDS and SFEDA algorithm determines the state of present node according only to the vote state of node of majority, fault freedom Shortcomings, increase along with error rate, there is mistake in more neighbor node, thus determined the result of decision of present node by the node of mistake, cause the mistake that can not repair local decision-making.

Figure 10 strengthens the performance comparison result of FEDAV algorithm and OTDS, SFEDA algorithm, also illustrate that the deployment density increasing node, it is possible to make the fault freedom of algorithm become stable.

The new mistake that mistake introducing rate and algorithm introduce, Figure 11 and Figure 12 is the contrast of the mistake introducing rate of FEDAV algorithm and OTDS, SFEDA algorithm.

For the algorithm based on temporal voting strategy, the introducing of new mistake occurs mainly in event boundaries region, because spatial redundancy information is not enough, is unsatisfactory for most voting mechanism. And FEDAV algorithm is in event boundaries region, ballot is not adopted to process: if node state is credible, it does not have make a mistake, then using the local event decision-making state of node as final detection result; If node state is LF, then carry out local decision-making by the estimated value that neighbor data calculates. So, so that it may avoid introducing new mistake.

And OTDS algorithm, in whole network, adopt temporal voting strategy, therefore mistake introducing rate is higher. And SFEDA algorithm is only when detecting data significant change, ballot is adopted to process; In other situations, depend on the local decision-making of node, so mistake introducing rate is lower than OTDS algorithm.

In Figure 11, when node error probability is relatively low, mistake introducing rate is also significantly high, being because first group of data, node density is less, and therefore spatial redundancy information is not enough, mistake introducing rate now is not caused by node mistake, and it is because the voting results to event boundaries node, there is ambiguousness, it is easy to produce new mistake.And the node density of Figure 12 is relatively big, embody the duty in practical application, therefore, mistake introducing rate coincidence theory value.

Owing to sensor node calculates the energy that the energy consumed to consume much smaller than communication, therefore, we represent the energy expenditure of node by the traffic of node. Consider that FEDAV, SFEDA and OTDS algorithm is all based on the space voting mechanism of the node result of decision, between node, the data of exchange mostly are decision data, therefore, for ease of analyzing, we adopt information exchange times to represent the energy expenditure of algorithm, and the exchange of each information includes sending solicited message to a neighbor node and accepting the feedback information of this neighbor node. And then, our definition information exchange frequency, it is the information exchange times by algorithm, is removed by the information exchange times of OTDS algorithm and obtain.

When Figure 13, Figure 14 sets forth two groups of data of employing, the energy expenditure contrast situation of FEDAV, SFEDA and OTDS algorithm. As seen from the figure, the energy expenditure of OTDS algorithm is the highest, because no matter node is normal condition or abnormality, is required for the ballot decision-making of neighbor node, therefore, needs frequent exchange data between node, and energy expenditure is very big. And SFEDA algorithm is only when data generation significant change, just carrying out ballot process, accordingly, it is capable to consume less, and the frequency of its information exchange is relevant with the probability of data generation significant change.

And FEDAV algorithm is because only when Node Events and node mistake, just needing node collaborative process, therefore information exchange frequency is less. But along with the rising of node error rate, the increased frequency of node cooperation, so information exchange frequency also raises. By scheming, when node error probability is relatively low, the energy expenditure of FEDAV algorithm is less; And node error probability higher time, the energy consumption of SFEDA algorithm is less.

Through experiment above, it can be deduced that FEDAV algorithm is all substantially better than OTDS algorithm and SFEDA algorithm in event detection effect, fault-tolerant processing effect and energy expenditure three, and this also demonstrates the analysis result of comparator algorithm.

The above is only the preferred embodiment of the present invention; it is noted that, for those skilled in the art; under the premise without departing from the principles of the invention, it is also possible to make some improvements and modifications, these improvements and modifications also should be regarded as protection scope of the present invention.

Claims (3)

1. based on the fault-tolerant event detecting method of temporal voting strategy, it is characterised in that:

The state of definition node is as follows:

Normal: represent node S_localMonitored area in without event occur, node S_localDo not break down and mistake;

LF: represent node S_localIt is likely to occur fault or wrong data, for node S_localLocal event judge in;

LE: represent node S_localEvent it is likely to occur, for node S in monitored area_localLocal event judge in;

Event: represent node S_localEvent occurs, for node S in monitored area_localIn ballot decision-making;

Fault: represent node S_localBreak down event, and can not correct, for node S_localIn ballot decision-making;

Including the following steps that order performs:

(1) each node obtains the detection data of self;

(2) based on rectangular histogram, detection data are carried out fault-tolerant processing; When detection data being carried out fault-tolerant processing based on rectangular histogram, the acquisition methods of data center is as follows:

(21) for data x₁,x₂,��,x_n, according to each data genaration histogram structure, obtain k packet g₁,g₂,��,g_i,��,g_k;

(22) size of definition i-th packet | g_i| the number of the data comprised in being grouped for i-th, obtain g according to the descending arrangement of the size of each packet_k',��,g_i',��,g₂',g₁';

(23) a given mean data coverage m₀, each group of data of descending accumulation, until group g_r' meet following formula:

$\frac{\underset{i=r}{\overset{k}{\&Sigma;}}{g}_i^{\&prime;}}{\underset{i=1}{\overset{k}{\&Sigma;}}{g}_i^{\&prime;}}\&GreaterEqual;m_0$

(24) to organize g_r',��,g_k' the data that comprise, calculate data center

$p=\underset{i=r}{\overset{k}{\mathrm{\&Sigma;}}}\left|g_i^{\&prime;}\right|$

${\stackrel{\&OverBar;}{x}}_h=\frac{1}{p}{\mathrm{\&Sigma;x}}_j$

X herein_j��g��_i, i=r, r+1 ..., k;

(3) based on the dependency of attribute, detection data are carried out local error detection, distinguish LF node and preliminary Normal node; Method particularly includes: the initial relevance of two strong correlation attribute X and Y of note node is R₀(X, Y), real-time dependency is R (X, Y), for a dependency irrelevance threshold xi, if the intensity of variation of Attribute Correlation meets following formula:

$\left|\frac{R(X,Y)-R_0(X,Y)}{R_0(X,Y)}\right|<\mathrm{\&xi;}$

Then thinking that the dependency of X and Y is not changed in, node is preliminary Normal node; Otherwise it is assumed that the dependency of X and Y there occurs change, node is Event node;

(4) preliminary Normal node is carried out local event decision-making, distinguish LE node and Normal node; Method particularly includes:

If the average value that attribute X is under normal circumstances is M_n, the average value when event occurs is M_e, M_e>M_n, the threshold value x of definition event generation_thFor:

$x_{th}=\frac{1}{2}(M_n+M_e)$

Real-time value x for attribute X_i, the binary judgment rule that event occurs is:

$B_e=\left\{\begin{array}{cc}1& x_i>x_{th}\\ 0& x_i\&le;x_{th}\end{array}\right.$

Wherein, B_e=1 represents that node is LE node, B_e=0 represents that node is Normal node;

(5) to LF node and LE node, the credibility of its neighbor node is calculated; The credibility C of node is calculated by following formula:

$C=1-\frac{|R-R_0|}{\left|R\right|+\left|R_0\right|}$

Wherein, R₀(X, Y) represents attribute X and the initial relevance of attribute Y, and R (X, Y) represents the real-time dependency of attribute X and attribute Y;

(6) vote based on credibility and distance weighting, it is judged that node is Event node or Normal node; Method particularly includes:

(61) definition S_localFor present node, S_iFor S_localA hop neighbor node, then:

Node S_iRelative to node S_localAnti-distance weighting InvDistWeight_ilFor:

${\mathrm{InvDist}}_{i1}=\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}Dis\tan ce(S_i,S_{local})-Dis\tan ce(S_i,S_{local})$

${\mathrm{InvDistWeight}}_{i1}=\frac{{\mathrm{InvDist}}_{i1}}{\underset{i=1}{\overset{k}{\&Sigma;}}{\mathrm{InvDist}}_{i1}}$

Wherein, Distance (S_i,S_local) it is node S_iWith node S_localBetween the Euclidean distance of coordinate;

Node S_iTo node S_localBallot value V_iFor:

V_i=C_i��InvDistWeight_il

Wherein, C represents the credibility of node;

(62) at present node S_localAll hop neighbor node S_iIn, calculate and there is identical local event result of decision ballot value sum, remember that the ballot value sum that all local decision-makings are LE node is V_e, the ballot value sum that all local decision-makings are Normal node is V_n, calculate V_eAnd V_nBetween difference degree V_diffFor:

$V_{diff}=\frac{|V_e-V_n|}{|V_e+V_n|}$

(64) by V_diffCompare with the event boundaries threshold �� set:

If V_n> 0 and V_e=0, then judge that LE node is Normal node, LF node is Normal node;

If V_n-V_e> ��, then judging that LE node is Event node, LF node is Normal node;

If V_diff< ��, then judge that LE node is Event node, and LF node state is indefinite;

If V_e-V_n> ��, then judging that LE node is Event node, LF node is Event node;

If V_n=0 and V_e> 0, then judging that LE node is Event node, LF node is Event node.

2. the fault-tolerant event detecting method based on temporal voting strategy according to claim 1, it is characterised in that: as node S_localThe redundancy of neighbor node not enough, it is impossible to node S is corrected in the ballot relying on neighbours_localError condition time, first pass through the data estimation node S of neighbor node_localDetected value

Wherein, d_iFor node S_iTo node S_localBetween distance, p is the values of powers of distance;

Then withAs node S_localReal-time value, adopt the method for step (4) to carry out local event decision-making, and result step (4) distinguished be as node S_localFinal decision result.

3. the fault-tolerant event detecting method based on temporal voting strategy according to claim 1, it is characterised in that: for the node S not having effective neighbor node to vote_localIf the end-state processed through ballot is unsettled condition, then arrange node S_localState be Fault.