CN104318688A - Multi-sensor fire early-warning method based on data fusion - Google Patents

Abstract

The invention provides a multi-sensor fire early-warning method based on data fusion. The multi-sensor fire early-warning method comprises the steps that firstly, continuous sampling data of all sensor nodes are obtained within a sampling period Tn, and a matrix X of the sampling nodes and the sampling data of the sampling nodes is built; secondary, noise in the sampling data is filtered out through median filtering, sudden change points in the sampling data are found on this basis, and fire information detected by the nodes is obtained; thirdly, maximum sudden change value solving processing is carried out on the sampling data of homogeneous sensors, probability distribution of the sampling data is obtained, secondary fusion is carried out based on the D-S evidence theory, and the final fire probability condition is obtained. According to the method, wrong data and the noise can be effectively filtered out, accuracy and reliability of fire early-warnings are greatly improved, and therefore wrong warnings and missing warnings are avoided.

Description

A kind of multisensor fire alarm method based on data fusion

Technical field

The present invention relates to a kind of fire alarm method, be specifically related to a kind of multisensor fire alarm method based on data fusion.

Background technology

Forest, warehouse, park etc. to take place frequently area for fire, have the condition of a fire of some to occur every year, cause great economic loss and serious environmental pollution.This kind of fire has sudden strong, the destructive feature such as large.

Existing many fire early-warning systems, its traditional unit fire detector many employings unit Detection Techniques, and be single parameter fire detector (comprising threshold triggers formula and nalog quantity type).This detector, to the unevenness of fire characteristic signal response sensitivity, causes it to be restricted the detectivity of actual fire, and rate of false alarm and rate of failing to report are remained high always.Along with the progress of technology, start in fire detector to adopt polytype sensor jointly to detect flame.But still face many difficulties in actual use.One, the physical signalling utilizing sensor to obtain existing with fire condition relation not only changes with the condition of a fire, also can be subject to other factors interference, as installation site and other the human activity of the dust in electronic noise, air, sensor, these change often with condition of a fire feature similarity, therefore easily cause erroneous judgement.Such as photoelectric smoke detector is in ventilation speed place faster, and detection sensitivity can be a greater impact.Therefore, the noise how eliminated in sensor detection data obtains the crucial style that accurate result primarily solves.Two, the different phase induction of different sensors to flame combustion is also different.The temperature rise that the sensor of such as sense temperature only produces naked light is responsive, insensitive to smoldering fire, nor can distinguish the heat causing temperature to rise and produced by fire or produced by cigarette or bonfire etc.; Such as photoelectric smoke detector all has higher sensitivity to the general condition of a fire again, also has fabulous detection to smoldering fire, but to burning produce invisible pyrotechnics (particle diameter is less than 0.4um) or occur that the black smoke of naked light is insensitive.May occur the situation of result of detection conflict between dissimilar sensor like this, now how accepting or rejecting the result of detection of each sensor is the difficult problem that must solve.Due in the actual environments such as forest, park, warehouse, fire often has feature that is sudden, randomness, and sensor wrong report easily occurs and fails to report.Therefore, need a kind of method of data fusion badly, the data obtained comprehensively are analyzed under certain criterion, to improve unreliability and the limitation of the existence of moment sensor fire detecting system.

Summary of the invention

In order to solve the above-mentioned technical matters that existing fire early-warning system exists, the invention provides a kind of multisensor fire alarm method based on data fusion.The valid data of all kinds of homogeneity sensor are carried out pre-service by the present invention, and on this basis in conjunction with the D-S evidence theory of improvement, obtain final probability size, thus reach the object reducing fire wrong report, fail to report.

The technical scheme that the present invention solves the problems of the technologies described above comprises the following steps:

Step 1: obtain temperature sensor, smokescope sensor, carbonomonoxide concentration sensor, the sampled data of ultraviolet flame sensor within the sampling period in bunch territory respectively; Pre-service is carried out to sampled data, judges whether data should be sent to aggregators;

1.1: in bunch territory, arrange z temperature sensor, a z smokescope sensor, a z carbonomonoxide concentration sensor, a z ultraviolet flame sensor node respectively, the node total number of the sensor arranged adds up to 4z;

1.2:4z sampling node is at sampling period T _ninterior equal interval sampling environmental data n time;

Step 2: the X matrix setting up various kinds of sensors node sample data in the sampling period;

X matrix expression is as follows:

$x_A=\left[\begin{array}{cccc}{x}_{11}& x_{12}& \mathrm{\Λ}& x_{1n}\\ \mathrm{\Λ}& \mathrm{\Λ}& \mathrm{\Λ}& \mathrm{\Λ}\\ x_{i1}& x_{i2}& \mathrm{\Λ}& x_{\mathrm{in}}\\ \mathrm{\Λ}& \mathrm{\Λ}& \mathrm{\Λ}& \mathrm{\Λ}\\ x_{z1}& x_{z2}& \mathrm{\Λ}& x_{\mathrm{zn}}\end{array}\right]A\left(\mathrm{1,2,3,4}\right)---\left(1\right)$

Wherein (A=1,2,3,4) represent the kind of sensor, represent temperature sensor, smokescope sensor, carbon monoxide gas concentration sensor, ultraviolet flame sensor respectively; The row of matrix represents that i-th (i=1,2, Λ, z) individual node is at sampling period T respectively _nn sampled data of interior continuous acquisition, is designated as x _i1, x _i2..., x _in;

Step 3: carry out medium filtering process to every data line of X matrix, finds the catastrophe point in sampled data on this basis, obtains the fire information arrived by nodal test.To matrix X _arow X _aithe step of data processing is as follows:

3.1: set up the one dimension median filter that length of window is s (s is odd number, and S can be divided exactly by n).With row X _aifar Left be starting point, intermediate value is got to the data in window, and is designated as Q _ai(r), wherein (A=1,2,3,4), (i=1,2, Λ z), r represents the number of times getting intermediate value, now r=0;

3.2: with step-length s to the right moving window once, make r=r+1, intermediate value Q is got to the data in window _air (), judges twice, front and back intermediate value according to formula (2), that is:

α _ai=[Q _ai(r+1)-Q _ai(r)]-δ _a>0 (A=1,2,3,4; I=1,2, Λ, z) (2) wherein δ _a(A=1,2,3,4) represent the pre-judgment threshold of 4 class sensor, if meet formula (2), i.e. α _abe greater than zero, illustrate that data occur that non-stationary changes, and this point is designated as astable some X from this sampling _air, and jump to step 3.3 continue next step process; Otherwise, abandon data Q _air () also re-executes step 3.2; If same type of sensor has the sampled data of part of nodes can not meet formula (2) in the sampling period, then this cycle abandons the sampled data of this part of nodes;

3.3: the windowing cancelling the sampled data to the r time median operation.To the sampled data remained, with sampled point X _airfor true origin sets up the coordinate system of sampled data about sampling number;

3.4: under this coordinate system, obtain the function F of sampled data about sampling number by curve _a(x _w), and first order derivative is asked to each sampled data relative on this curve.By ask max [F ' _a(x _w)] obtain each sampling node gradient maxima, by the sampled data Y of its correspondence _a(x _i) send into step 3.5, because this value can reflect the catastrophe of fire in actual environment;

3.5: ask for same type of sensor node at T _nby obtaining the maximal value of output valve after medium filtering and differentiate process in cycle length:

m ₁＝max{Y ₁(x ₁)，Y ₁(x ₂)，Λ，Y ₁(x _z)}

m ₂＝max{Y ₂(x ₁)，Y ₂(x ₂)，Λ，Y ₂(x _z)}

m ₃＝max{Y ₃(x ₁)，Y ₃(x ₂)，Λ，Y ₃(x _z)} (3)

m ₄＝max{Y ₄(x ₁)，Y ₄(x ₂)，Λ，Y ₄(x _z)}

Wherein max{} represents and gets maxima operation, and maximal value step 3.5 exported is sent into next link and merged;

Step 4: average is merged according to the D-S evidence theory method improved;

4.1: definition event type Θ={ u ₁, u ₂, u ₃there is fire }={, without fire, uncertain }; Definition list 1 is the event occurrence rate under different sampling fusion value; For the four class sensor sample fusion values that step 3.5 exports, the elementary probability situation obtaining corresponding evidence according to table 1 is:

m ₁(u ₁)＝p _1,1；m ₁(u ₂)＝p _1,2；m ₁(u ₃)＝p _1,3

m ₂(u ₁)＝p _2,1；m ₂(u ₂)＝p _2,2；m ₂(u ₃)＝p _2,3

m ₃(u ₁)＝p _3,1；m ₃(u ₂)＝p _3,2；m ₃(u ₃)＝p _3,3

m ₄(u ₁)＝p _4,1；m ₄(u ₂)＝p _4,2；m ₄(u ₃)＝p _4,3

Wherein, m ₁, m ₂, m ₃, m ₄represent the evidence of four class sensors, i.e. m ₁: carbonomonoxide concentration, m ₂: smokescope, m ₃: temperature, m ₄: flame ultraviolet ray intensity;

4.2: calculate the evidence of any two sensors about u _kcompatible coefficient be:

$R_{s,j}\left(u_k\right)=\frac{m_s\left(u_k\right)\×m_j\left(u_k\right)}{\frac{m_s{\left(u_k\right)}^2+m_j{\left(u_k\right)}^2}{2}}(k=\mathrm{1,2,3};s,j=\mathrm{1,2,3,4})---\left(4\right)$

S, j represent any two class sensors, example: as k=1, calculate evidence between two two sensorses about u ₁the compatible coefficient of (having fire);

To calculate between all the sensors the compatible coefficient of evidence between two according to formula (5), obtain consistent matrix:

$\left(\begin{array}{cccc}{R}_{\mathrm{1,1}}& R_{\mathrm{1,2}}& R_{\mathrm{1,3}}& R_{\mathrm{1,4}}\\ R_{\mathrm{2,1}}& R_{\mathrm{2,2}}& R_{\mathrm{2,3}}& R_{\mathrm{2,4}}\\ R_{\mathrm{3,1}}& R_{\mathrm{3,2}}& R_{\mathrm{3,3}}& R_{\mathrm{3,4}}\\ R_{\mathrm{4,1}}& R_{\mathrm{4,2}}& R_{\mathrm{4,3}}& R_{\mathrm{4,4}}\end{array}\right)---\left(5\right)$

4.3: the absolute compatible degree calculating every bar evidence is:

$D_s\left(u_k\right)=\underset{j=1,s\≠j}{\overset{4}{\mathrm{\Σ}}}{R}_{s,j}\left(u_k\right)(s=\mathrm{1,2,3,4};k=\mathrm{1,2,3})---\left(6\right)$

Owing to expecting that between evidence, compatible coefficient is 1, desirable compatible degree is n-1 (n is compatible evidence number, n=4 in the present invention), and this evidence is about u _kconfidence level definition:

$\mathrm{Re}{l}_s\left(u_k\right)=\frac{D_s\left(u_k\right)}{n-1}=w_s(k=\mathrm{1,2,3};s=\mathrm{1,2,3,4})---\left(7\right)$

Confidence level is simultaneously as this evidence u _kweight;

4.4: evidential probability situation new after obtaining weighting by repetition step 4.2-4.3:

m ₁(u ₁)＝w ₁p _1,1；m ₁(u ₂)＝w′ ₁p _1,2；m ₁(u ₃)＝w″ ₁p _1,3

m ₂(u ₁)＝w ₂p _2,1；m ₂(u ₂)＝w′ ₂p _2,2；m ₂(u ₃)＝w″ ₂p _2,3

m ₃(u ₁)＝w ₃p _3,1；m ₃(u ₂)＝w′ ₃p _3,2；m ₃(u ₃)＝w″ ₃p _3,3

m ₄(u ₁)＝w ₄p _4,1；m ₄(u ₂)＝w′ ₄p _4,2；m ₄(u ₃)＝w″ ₄p _4,3

4.5: ask for inconsistent degree, by the conflict spectrum between two between evidence obtained above, i.e. evidential probability situation, calculating K value:

Wherein, m ₁and m ₂represent the basic trust allocation probability after step 4.3-4.4 weighting, A, B represent evidence, represent any two class evidences in the ultraviolet ray intensity of carbonomonoxide concentration, smokescope, temperature, flame in the present invention; K reflects the conflict spectrum between each evidence, i.e. the inconsistent factor, as K=0, represents that two kinds of evidences are completely compatible, as K=1, represents that two kinds of evidences conflict completely; As 0<K<1, represent two kinds of evidence partially compatible;

4.6: ask the probable value after two kinds of sensor fusion by D-S evidence theory:

In formula, coefficient 1/K is regularization factors, both its role is to first for Jiao standardization, m ₁and m ₂obtained by step 4.3-4.4;

4.7: repeat step 4.2 ~ 4.6, the result merge first time and the sampling fusion value of the third sensor merge according to D-S evidence theory method, obtain merging after probability, then with the 4th class sensor fusion, obtain the final probability distribution situation that fire occurs: m (X)={ P _u1, P _u2, P _u3.Technique effect of the present invention is:

In sample phase, can effectively filtering misdata and noise, significantly improve accuracy and the reliability of fire alarm; At data preprocessing phase, by threshold decision, remove the redundant data of sensor of the same type, effectively reduce the consumption of energy; In the data fusion stage, by the method for D-S evidence theory, merge the sampled data from four class sensors, thus avoid the wrong report brought by single class sensor and the generation of failing to report phenomenon, and the redundant information deleted between dissimilar sensor, the energy consumption of information transmission should be able to be reduced mutually.

Accompanying drawing explanation

Fig. 1 is schematic diagram of the present invention.

Fig. 2 is process flow diagram of the present invention.

Fig. 3 is filter window and sliding type schematic diagram thereof in the present invention.

Fig. 4 is the process flow diagram that in the present invention, multi-sensor data carries out D-S evidence theory.

Embodiment

As depicted in figs. 1 and 2, the method of the invention process Fusion, mainly for sensor fire alarm current in the fire alarms such as forest, park, warehouse wrong report and rate of failing to report this difficult problem high, provides a kind of multi-sensor data collection, fusion, decision analysis a kind of method finally obtained compared with accurate information.Now describe specific embodiment of the invention step in detail in conjunction with concrete probable value as follows:

Step 1: obtain temperature sensor, smokescope sensor, carbonomonoxide concentration sensor, the sampled data of ultraviolet flame sensor within the sampling period in bunch territory respectively; Pre-service is carried out to sampled data, judges whether data should be sent to aggregators;

1.1: in bunch territory, arrange z temperature sensor, a z smokescope sensor, a z carbonomonoxide concentration sensor, a z ultraviolet flame sensor node respectively, the node total number of the sensor arranged adds up to 4z;

1.2:4z sampling node is at sampling period T _ninterior equal interval sampling environmental data n time;

Step 2: the X matrix setting up various kinds of sensors node sample data in the sampling period;

X matrix expression is as follows:

$x_A=\left[\begin{array}{cccc}{x}_{11}& x_{12}& \mathrm{\&Lambda;}& x_{1n}\\ \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}\\ x_{i1}& x_{i2}& \mathrm{\&Lambda;}& x_{\mathrm{in}}\\ \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}\\ x_{z1}& x_{z2}& \mathrm{\&Lambda;}& x_{\mathrm{zn}}\end{array}\right]A\left(\mathrm{1,2,3,4}\right)---\left(1\right)$

Wherein (A=1,2,3,4) represent the kind of sensor, represent temperature sensor, smokescope sensor, carbon monoxide gas concentration sensor, ultraviolet flame sensor respectively; The row of matrix represents that i-th (i=1,2, Λ, z) individual node is at sampling period T respectively _nn sampled data of interior continuous acquisition, is designated as x _i1, x _i2..., x _in;

Step 3: medium filtering process is carried out to every data line of X matrix, finds jump signal on this basis, obtain the fire information arrived by nodal test.To matrix X _athe step of the i-th row data processing is as follows:

3.1: set up the one dimension median filter that length of window is s, wherein, s is odd number, and S can be divided exactly by n.With the 1st column position of the i-th row for starting point, intermediate value is got to the data in window, is designated as Q _air (), wherein r represents the number of times getting intermediate value.

3.2: with step-length s to the right moving window once, make r=r+1, intermediate value Q is got to the data in window _ai(r).(2) twice, front and back intermediate value is judged according to formula, that is:

α _Ai＝[Q _Ai(r+1)-Q _Ai(r)]-δ _A>0 (A＝1,2,3,4；i＝1,2,Λ,z) ⑵

Wherein δ _a(A=1,2,3,4) represent the pre-judgment threshold of 4 class sensor, if meet formula (2), i.e. and α _abe greater than zero, illustrate that data occur that non-stationary changes, and this point is designated as not stationary point X from this sampling _air, and jump to step 3.3 continue next step process; Otherwise, abandon data Q _air () also re-executes step 3.2; If same type of sensor has the sampled data of part of nodes can not meet formula (2) in the sampling period, then this cycle abandons the sampled data of this part of nodes;

3.3: the windowing cancelling the sampled data to the r time median operation.To the sampled data remained, with sampled point X _airfor true origin sets up the coordinate system of sampled data about sampling number.

3.4: under this coordinate system, obtain the function F of sampled data about sampling number by curve _a(x _w), and ask single order to lead to each sampled data relative on this curve, by ask max [F ' _a(x _w)] obtain each sampling node gradient maxima.The physical meaning of differentiate is to solve the rate of change of function to variable, namely in the present invention sampled data about the situation of change of sampling number, when derivative value is maximum, can be regarded as the environmental data change that this sampling obtain fast the most remarkable, also the catastrophe of fire in actual environment can be reacted, by the sampled data Y corresponding to gradient maxima _a(x _i) send into step 3.5;

3.5: ask for same type of sensor node at T _nthe maximal value of output valve by obtaining after medium filtering and differentiate process in cycle length:

m ₁＝max{Y ₁(x ₁)，Y ₁(x ₂)，Λ，Y ₁(x _z)}

m ₂＝max{Y ₂(x ₁)，Y ₂(x ₂)，Λ，Y ₂(x _z)}

m ₃＝max{Y ₃(x ₁)，Y ₃(x ₂)，Λ，Y ₃(x _z)} ⑶

m ₄＝max{Y ₄(x ₁)，Y ₄(x ₂)，Λ，Y ₄(x _z)}

Wherein max{} represents and gets maxima operation, and maximal value step 3.5 exported is sent into next link and merged;

Step 4: according to the method for the D-S evidence theory improved, average is merged, calculating sampling fusion value; 4.1: definition event type Θ={ u ₁, u ₂, u ₃there is fire }={, without fire, uncertain }; Table 1 represents the event occurrence rate under different sampling fusion value; Obtained the sampling fusion value of four class sensors by step 3, the elementary probability situation obtaining corresponding evidence according to table 1 is:

m ₁(u ₁)＝p _1,1；m ₁(u ₂)＝p _1,2；m ₁(u ₃)＝p _1,3

m ₂(u ₁)＝p _2,1；m ₂(u ₂)＝p _2,2；m ₂(u ₃)＝p _2,3

m ₃(u ₁)＝p _3,1；m ₃(u ₂)＝p _3,2；m ₃(u ₃)＝p _3,3

m ₄(u ₁)＝p _4,1；m ₄(u ₂)＝p _4,2；m ₄(u ₃)＝p _4,3

Wherein, m _f(f=1,2,3,4) represent four class sensor fusion values, i.e. m ₁: carbonomonoxide concentration, m ₂: smokescope, m ₃: temperature, m ₄: the ultraviolet ray intensity of flame;

Reference table 1 obtains probability distribution corresponding to four class fusion values, and sensor fusion value and existing fire model knowledge map mutually (the knowledge only use for example of this fire model), and the interval belonging to fusion value obtains event occurrence rate distribution.

The relation of table 1 event occurrence rate and sampling fusion value

After obtaining the probability distribution of sensor, successively it is merged between two.Such as, carbonomonoxide concentration value is measured as 17, known with reference to table 1, and the probability that now all kinds of event occurs is: m ₁{ having fire, without fire, uncertain }={ 0.6,0.2,0.2}.Suppose that sensor obtains m by the Data Fusion of Sensor of single type ₁, m ₂, m ₃, m ₄be followed successively by carbonomonoxide concentration: 19, smokescope: 6.5, temperature: 56, the ultraviolet ray intensity of flame: 3; Obtain the distribution of each evidential probability as follows:

m ₁：m ₁(u ₁)＝0.6,m ₁(u ₂)＝0.2,m ₁(u ₃)＝0.2

m ₂：m ₂(u ₁)＝0.75,m ₂(u ₂)＝0.1,m ₂(u ₃)＝0.15

m ₃：m ₃(u ₁)＝0.7,m ₃(u ₂)＝0.2,m ₃(u ₃)＝0.1

m ₄：m ₄(u ₁)＝0.8,m ₄(u ₂)＝0.1,m ₄(u ₃)＝0.1

Article 4.2: two, evidence basic probability assignment function is respectively m _s(u _k), m _j(u _k), calculate these two evidences about u _kcompatible coefficient be:

$R_{s,j}\left(u_k\right)=\frac{m_s\left(u_k\right)\&times;m_j\left(u_k\right)}{\frac{m_s{\left(u_k\right)}^2+m_j{\left(u_k\right)}^2}{2}}---\left(4\right)$

To proposition u ₁process, according to formula (4), obtain

R _1,2(u ₁)＝0.9756

R _1,3(u ₁)＝1.077

R _1,4(u ₁)＝0.96

R _2,3(u ₁)＝0.9976

R _2,4(u ₁)＝0.9979

R _3,4(u ₁)＝0.9912

4.3: known n bar evidence, obtains the compatible coefficient of evidence between two according to formula (4), obtain proposition u _kevidence consistent matrix:

$\left(\begin{array}{cccc}{R}_{\mathrm{1,1}}& R_{\mathrm{1,2}}& R_{\mathrm{1,3}}& R_{\mathrm{1,4}}\\ R_{\mathrm{2,1}}& R_{\mathrm{2,2}}& R_{\mathrm{2,3}}& R_{\mathrm{2,4}}\\ R_{\mathrm{3,1}}& R_{\mathrm{3,2}}& R_{\mathrm{3,3}}& R_{\mathrm{3,4}}\\ R_{\mathrm{4,1}}& R_{\mathrm{4,2}}& R_{\mathrm{4,3}}& R_{\mathrm{4,4}}\end{array}\right)---\left(5\right)$

Proposition u is obtained by step 4.3 ₁evidence consistent matrix be:

$\left[\begin{array}{cccc}1& 0.9756& 1.077& 0.96\\ 0.9756& 1& 0.9976& 0.9979\\ 1.077& 0.9976& 1& 0.9912\\ 0.96& 0.9979& 0.9912& 1\end{array}\right]$

4.4: the absolute compatible degree computing formula of evidence is:

$D_s\left(u_k\right)=\underset{j=1,s\&NotEqual;j}{\overset{4}{\mathrm{\&Sigma;}}}{R}_{s,j}\left(u_k\right)(s=\mathrm{1,2,3,4};k=\mathrm{1,2,3})---\left(6\right)$

The absolute compatible degree that can obtain every bar evidence is:

D ₁(u ₁)＝R _1,2+R _1,3+R _1,4＝3.0126

D ₂(u ₁)＝R _2,1+R _2,3+R _2,4＝2.9711

D ₃(u ₁)＝R _3，1+R _3,2+R _3,4＝3.0658

D ₄(u ₁)＝R _4，1+R _4，2+R _4，3＝2.9491

Owing to expecting that between evidence, compatible coefficient is 1, desirable compatible degree is n-1, n=4 in the present invention, and this evidence is about u _kconfidence level definition:

$\mathrm{Re}{l}_s\left(u_k\right)=\frac{D_s\left(u_k\right)}{n-1}=w_s(s=\mathrm{1,2,3,4};k=\mathrm{1,2,3})---\left(7\right)$

Confidence level is simultaneously as this evidence u _kweight; The confidence level of every bar evidence is as follows:

$w_1=\frac{D_1\left(u_1\right)}{n-1}=1.0042$

$w_2=\frac{D_2\left(u_1\right)}{n-1}=0.9904$

$w_3=\frac{D_3\left(u_1\right)}{n-1}=1.022$

$w_4=\frac{D_4\left(u_1\right)}{n-1}=0.983$

Using confidence level as the corresponding proposition u of this evidence ₁weight, basic probability assignment function is processed, obtains the distribution of new evidential probability as follows:

m ₁(u ₁)＝1.0042×0.6＝0.603

m ₂(u ₁)＝0.9904×0.75＝0.7428

m ₃(u ₁)＝1.022×0.7＝0.7154

m ₄(u ₁)＝0.983×0.8＝0.7864

4.5: repeat step 4.2 ~ 4.4 couple proposition u ₂, u ₃, u ₄do respectively and as above process, obtain

m ₁(u ₂)＝0.2×0.866＝0.1732

m ₂(u ₂)＝0.1×0.866＝0.0866

m ₃(u ₂)＝0.2×0.866＝0.1732

m ₄(u ₂)＝0.1×0.866＝0.0866

m ₁(u ₃)＝0.2×2.56/3＝0.171

m ₂(u ₃)＝0.15×2.806/3＝0.1403

m ₃(u ₃)＝0.1×2.723/3＝0.09

m ₄(u ₃)＝0.1×2.723/3＝0.09

After above-mentioned process, source evidence becomes new evidence function:

m ₁：m ₁(u ₁)＝0.6023,m ₁(u ₂)＝0.1732,m ₁(u ₃)＝0.171

m ₂：m ₂(u ₁)＝0.7428,m ₂(u ₂)＝0.0866,m ₂(u ₃)＝0.1403

m ₃：m ₃(u ₁)＝0.7154,m ₃(u ₂)＝0.1732,m ₃(u ₃)＝0.09

m ₄：m ₄(u ₁)＝0.7864,m ₄(u ₂)＝0.0866,m ₄(u ₃)＝0.09

4.6: carry out data fusion to dissimilar sensing data, fusion method is D-S evidence theory method.Ultimate principle is as follows: ask for inconsistent degree, the conflict spectrum namely between each evidence provided by forest fire database;

4.6.1:K value expression:

Wherein, m ₁and m ₂represent corresponding basic trust partition function, K reflects the conflict spectrum between each evidence, i.e. the inconsistent factor, as K=0, represents that two kinds of evidences are completely the same, as K=1, represents that two kinds of evidences conflict completely; As 0<K<1, represent two kinds of evidence partially compatible;

4.6.2: ask the probable value after two kinds of sensor fusion by D-S evidence theory:

In formula, coefficient 1/K becomes regularization factors, and effect is first for Jiao X standardization.Wherein meaning be to obtain a new elementary probability by after two independently elementary probability combination;

4.6.3: first ask for m ₁and m ₂d-S merge probability, according to formula (8) calculate normaliztion constant K:

K＝1-(0.6023×0.0866+0.7428×0.1732)＝0.82

4.6.4: the probable value after utilizing D-S combining evidences rule to calculate two kinds of sensor fusion respectively:

First ask for regularization factors:

1/k＝1.22

The probability of breaking out of fire: m (u ₁)=0.804

The probability of not breaking out of fire: m (u ₂)=0.1668

Unknown situation: m (u ₃)=0.0292

Therefore, the probability after carbonomonoxide concentration and smokescope two kinds of sensor fusion is obtained for { having fire, without fire, uncertain }={ 0.804,0.1668,0.0292};

4.6.5: repeat step 4.6.1 ~ 4.6.4, after result after this being merged and temperature sensor merge, merge with ultraviolet flame acquisition sensor again, the probability distribution situation finally obtaining fire generation is { 0.8904,0.0966,0.013}, the probabilistic forecasting of the known breaking out of fire of result thus reaches 0.8904, and the prediction probability of uncertain whether breaking out of fire is decreased to 0.013.

Known by last probability, the uncertainty of system drops to negligible degree, namely the determinacy of system is significantly improved, originally single-sensor can not judge whether breaking out of fire, the correct probability judged can be improved after fusion, embody the data anastomosing algorithm of the present invention's proposition to the advantage improving system reliability.

Claims (1)

1., based on a multisensor fire alarm method for data fusion, comprise the following steps:

Step 1: obtain temperature sensor, smokescope sensor, carbonomonoxide concentration sensor, the sampled data of ultraviolet flame sensor within the sampling period in bunch territory respectively; Pre-service is carried out to sampled data, judges whether data should be sent to aggregators;

1.1: in bunch territory, arrange z temperature sensor, a z smokescope sensor, a z carbonomonoxide concentration sensor, a z ultraviolet flame sensor node respectively, the node total number of the sensor arranged adds up to 4z;

1.2:4z sampling node is at sampling period T _ninterior equal interval sampling environmental data n time;

Step 2: the X matrix setting up various kinds of sensors node sample data in the sampling period;

X matrix expression is as follows:

$X_A=\left[\begin{array}{cccc}{x}_{11}& x_{12}& \mathrm{\&Lambda;}& x_{1n}\\ \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}\\ x_{i1}& x_{i2}& \mathrm{\&Lambda;}& x_{\mathrm{in}}\\ \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}& \mathrm{\&Lambda;}\\ x_{z1}& x_{z2}& \mathrm{\&Lambda;}& x_{\mathrm{zn}}\end{array}\right]A=\left(\mathrm{1,2,3,4}\right)---\left(1\right)$

Wherein (A=1,2,3,4) represent the kind of sensor, represent temperature sensor, smokescope sensor, carbon monoxide gas concentration sensor, ultraviolet flame sensor respectively; The row of matrix represents that i-th (i=1,2, Λ, z) individual node is at sampling period T respectively _nn sampled data of interior continuous acquisition, is designated as x _i1, x _i2..., x _in;

Step 3: carry out medium filtering process to every data line of X matrix, finds the catastrophe point in sampled data on this basis, obtains the fire information arrived by nodal test.To matrix X _arow X _aithe step of data processing is as follows:

3.1: set up the one dimension median filter that length of window is s (s is odd number, and S can be divided exactly by n).With row X _aifar Left be starting point, intermediate value is got to the data in window, and is designated as Q _ai(r), wherein (A=1,2,3,4), (i=1,2, Λ z), r represents the number of times getting intermediate value, now r=0;

3.2: with step-length s to the right moving window once, make r=r+1, intermediate value Q is got to the data in window _air (), judges twice, front and back intermediate value according to formula (2), that is:

α _Ai＝[Q _Ai(r+1)-Q _Ai(r)]-δ _A>0 (A＝1,2,3,4；i＝1,2,Λ,z) (2)

Wherein δ _a(A=1,2,3,4) represent the pre-judgment threshold of 4 class sensor, if meet formula (2), i.e. α _abe greater than zero, illustrate that data occur that non-stationary changes, and this point is designated as astable some X from this sampling _air, and jump to step 3.3 continue next step process; Otherwise, abandon data Q _air () also re-executes step 3.2; If same type of sensor has the sampled data of part of nodes can not meet formula (2) in the sampling period, then this cycle abandons the sampled data of this part of nodes;

3.3: the windowing cancelling the sampled data to the r time median operation.To the sampled data remained, with sampled point X _airfor true origin sets up the coordinate system of sampled data about sampling number;

3.4: under this coordinate system, obtain the function F of sampled data about sampling number by curve _a(x _w), and first order derivative is asked to each sampled data relative on this curve.By ask max [F ' _a(x _w)] obtain each sampling node gradient maxima, by the sampled data Y of its correspondence _a(x _i) send into step 3.5, because this value can reflect the catastrophe of fire in actual environment;

3.5: ask for same type of sensor node at T _nby obtaining the maximal value of output valve after medium filtering and differentiate process in cycle length:

m ₁＝max{Y ₁(x ₁)，Y ₁(x ₂)，Λ，Y ₁(x _z)}

m ₂＝max{Y ₂(x ₁)，Y ₂(x ₂)，Λ，Y ₂(x _z)}

m ₃＝max{Y ₃(x ₁)，Y ₃(x ₂)，Λ，Y ₃(x _z)} (3)

m ₄＝max{Y ₄(x ₁)，Y ₄(x ₂)，Λ，Y ₄(x _z)}

Wherein max{} represents and gets maxima operation, and maximal value step 3.5 exported is sent into next link and merged;

Step 4: average is merged according to the D-S evidence theory method improved;

4.1: definition event type Θ={ u ₁, u ₂, u ₃there is fire }={, without fire, uncertain }; Definition list 1 is the event occurrence rate under different sampling fusion value; For the four class sensor sample fusion values that step 3.5 exports, the elementary probability situation obtaining corresponding evidence according to table 1 is:

m ₁(u ₁)＝p _1,1；m ₁(u ₂)＝p _1,2；m ₁(u ₃)＝p _1,3

m ₂(u ₁)＝p _2,1；m ₂(u ₂)＝p _2,2；m ₂(u ₃)＝p _2,3

m ₃(u ₁)＝p _3,1；m ₃(u ₂)＝p _3,2；m ₃(u ₃)＝p _3,3

m ₄(u ₁)＝p _4,1；m ₄(u ₂)＝p _4,2；m ₄(u ₃)＝p _4,3

Wherein, m ₁, m ₂, m ₃, m ₄represent the evidence of four class sensors, i.e. m ₁: carbonomonoxide concentration, m ₂: smokescope, m ₃: temperature, m ₄: flame ultraviolet ray intensity;

4.2: calculate the evidence of any two sensors about u _kcompatible coefficient be:

$\begin{array}{cc}{R}_{s,j}\left(u_k\right)=\frac{m_s\left(u_k\right)\&times;m_j\left(u_k\right)}{\frac{m_s{\left(u_k\right)}^2+m_j{\left(u_k\right)}^2}{2}}& (k=\mathrm{1,2,3};s,j=\mathrm{1,2,3,4})---\left(4\right)\end{array}$

S, j represent any two class sensors, example: as k=1, calculate evidence between two two sensorses about u ₁the compatible coefficient of (having fire);

To calculate between all the sensors the compatible coefficient of evidence between two according to formula (5), obtain consistent matrix:

$\left(\begin{array}{cccc}{R}_{\mathrm{1,1}}& R_{\mathrm{1,2}}& R_{\mathrm{1,3}}& R_{\mathrm{1,4}}\\ R_{\mathrm{2,1}}& R_{\mathrm{2,2}}& R_{\mathrm{2,3}}& R_{\mathrm{2,4}}\\ R_{\mathrm{3,1}}& R_{\mathrm{3,2}}& R_{\mathrm{3,3}}& R_{\mathrm{3,4}}\\ R_{\mathrm{4,1}}& R_{\mathrm{4,2}}& R_{\mathrm{4,3}}& R_{\mathrm{4,4}}\end{array}\right)---\left(5\right)$

4.3: the absolute compatible degree calculating every bar evidence is:

$D_s\left(u_k\right)=\underset{j=1,s\&NotEqual;j}{\overset{4}{\mathrm{\&Sigma;}}}{R}_{s,j}\left(u_k\right)(s=\mathrm{1,2,3,4};k=\mathrm{1,2,3})---\left(6\right)$

Owing to expecting that between evidence, compatible coefficient is 1, desirable compatible degree is n-1 (n is compatible evidence number, n=4 in the present invention), and this evidence is about u _kconfidence level definition:

$\begin{array}{cc}{\mathrm{Rel}}_s\left(u_k\right)=\frac{D_s\left(u_k\right)}{n-1}=w_s& (k=\mathrm{1,2,3};s=\mathrm{1,2,3,4})\end{array}---\left(7\right)$

Confidence level is simultaneously as this evidence u _kweight;

4.4: evidential probability situation new after obtaining weighting by repetition step 4.2-4.3:

m ₁(u ₁)＝w ₁p _1,1；m ₁(u ₂)＝w′ ₁p _1,2；m ₁(u ₃)＝w″ ₁p _1,3

m ₂(u ₁)＝w ₂p _2,1；m ₂(u ₂)＝w′ ₂p _2,2；m ₂(u ₃)＝w″ ₂p _2,3

m ₃(u ₁)＝w ₃p _3,1；m ₃(u ₂)＝w′ ₃p _3,2；m ₃(u ₃)＝w″ ₃p _3,3

m ₄(u ₁)＝w ₄p _4,1；m ₄(u ₂)＝w′ ₄p _4,2；m ₄(u ₃)＝w″ ₄p _4,3

4.5: ask for inconsistent degree, by the conflict spectrum between two between evidence obtained above, i.e. evidential probability situation, calculating K value:

Wherein, m ₁and m ₂represent the basic trust allocation probability after step 4.3-4.4 weighting, A, B represent evidence, represent any two class evidences in the ultraviolet ray intensity of carbonomonoxide concentration, smokescope, temperature, flame in the present invention; K reflects the conflict spectrum between each evidence, i.e. the inconsistent factor, as K=0, represents that two kinds of evidences are completely compatible, as K=1, represents that two kinds of evidences conflict completely; As 0<K<1, represent two kinds of evidence partially compatible;

4.6: ask the probable value after two kinds of sensor fusion by D-S evidence theory:

In formula, coefficient 1/K is regularization factors, both its role is to first for Jiao standardization, m ₁and m ₂obtained by step 4.3-4.4;

4.7: repeat step 4.2 ~ 4.6, the result merge first time and the sampling fusion value of the third sensor merge according to D-S evidence theory method, obtain merging after probability, then with the 4th class sensor fusion, obtain the final probability distribution situation that fire occurs: m (X)={ P _u1, P _u2, P _u3.