CN104091078A - Product multi-information fusion identification failure remedying method based on D-S evidence theory - Google Patents

Abstract

The invention discloses a product multi-information fusion identification failure remedying method based on the D-S evidence theory. The technical problem that tracking disconnection is caused by an existing method in the production process of individual products is solved. According to the technical scheme, the method comprises the steps that multi-source information correlation and data structure existing in the product physical and manufacturing process is analyzed, a product scattering and incomplete multi-source information model needed in the scattering process is built, the feature data of failure identification parts and product historical database similarity measure are calculated through a class-center-based Euclidean distance and improved variable coefficient weighting method, fusion identification is carried out on a feature layer through the D-S evidence theory, the identifications of the failure products are effectively recovered and remedied, and then the technical problem that tracking disconnection is caused in the production process of the individual products due to DM code failure is solved. Experimental results show that the identification accuracy of preventing tracking failure is larger than 95%, efficiency is high, recovering and remedying are reliable, and practicability is good.

Description

The many information fusion marks of product inefficacy means to save the situation based on D-S evidence theory

Technical field

The present invention relates to a kind of product identification inefficacy means to save the situation, particularly relate to the many information fusion marks of a kind of product based on D-S evidence theory inefficacy means to save the situation.

Background technology

Document " a kind of image synthesis preprocess method for workpiece 3D bar-code identification. Xi'an University of Technology's journal, 2002, 18 (4): 332-336. " a kind of image synthesis preprocess method that recovers identification for workpiece 3D bar code is disclosed, laser-induced thermal etching is in the bar code of surface of the work, is furnished with the industrial camera picked-up 3D bar code image of secondary light source, through histogram equalization, image projection transformation, adaptive threshold law spacing wave is extracted, after the empty information true and false of bar correction etc. is processed, finally obtain workpiece numbering by GB decoding, part has met the identification recovery of product under the concrete condition of some operation and has followed the trail of requirement.Described in document, method is mainly to know method for distinguishing after the image pre-service by bar code again, nonetheless, the techniques such as wearing and tearing, pollution, thermal treatment and surface treatment in product lifecycle production run still can make DM code and cleartext information recognition all fail, and the picture breakdown of product identification cannot extract reliable characteristic or can not extract feature completely from image when more serious, so, existing technical measures can not be stopped product identification completely and not lose efficacy in its all production link, cause product individuality to follow the trail of and open circuit in process of production.

Summary of the invention

Cause product individuality to follow the trail of in process of production the deficiency opening circuit in order to overcome existing method, the invention provides the many information fusion marks of a kind of product based on D-S evidence theory inefficacy means to save the situation.Associated and data structure in the method analytic product physics and manufacture process between multi-source information, set up in Discrete Manufacturing Process the scattered and imperfect multi-source information model of product, by the Euclidean distance based on class center and improved coefficient of variation weighted method, lost efficacy the mark characteristic of part and the historical data base similarity of product are estimated to calculating, finally adopt D-S evidence theory to merge identification to its characteristic layer, effectively recover to remedy the mark of inefficacy product, and then prevent DM code to lose efficacy and cause product individuality to follow the trail of and open circuit in process of production.

The technical solution adopted for the present invention to solve the technical problems is: the many information fusion marks of a kind of product based on D-S evidence theory inefficacy means to save the situation, is characterized in adopting following steps:

Step 1, the information source that analysis is relevant to product identification, comprise product physical message and manufacture process information, and many information fusion are incorporated into product represent to lose efficacy and remedy.If a multi-source information model, wherein object set is

P＝{x ₁,x ₂,x ₃,.....x _i} (1)

The main body of this multi-source information, element x _i, i={1,2,3 ... ..n} represents different information sources; Property set is

C＝{c ₁,c ₂,c ₃,.....c _j} (2)

It is the description of product external information attribute; The codomain of attribute is discrete values, is designated as U _m(m≤j), wherein m represents the attribute of product different aforementioned sources, and j represents all properties of product different aforementioned sources, and the set of relations of object set P and property set C is

F＝{f _i:m≤j} (3)

Wherein, f _m: P → U _m(m≤j), f _irepresent product multi-source information object set, f _mrepresenting the property set of actual different aforementioned sources, expressed contacting between object set U and property set C, is the basis of information source.Establish again the status attribute collection of E for being determined by property set C,

E＝{e ₁,e ₂,e ₃,.....e _j} (4)

Formula (4) represents the state that each attribute of product object comprises, its codomain U _m' (m≤j) is that qualitative value specified states is divided into mark and lost efficacy and do not lose efficacy, and property set C is called conditional attribute collection.Product multi-source information model represents with following formula:

I＝{P,C,F,E} (5)

For the outside multiple information sources of product arbitrarily, above-mentioned model representation becomes a two-dimensional data table.

Step 2, when mark occur lost efficacy time, gather scattered and incomplete product physics and manufacturing information, merge the requirement of identification to meet product, similarity between judgement sample often adopts neighbour's criterion, the degree that adopts the mark that lost efficacy to mate with center different classes of in Sample Storehouse is high, repeatedly measures the correlated characteristic of inefficacy product and carries out calculating with the similarity measure of historical record data.

1. the many characteristic models of product based on record.

If (X ₍₁₎, X ₍₂₎...., X _(n)) be that the multi-source information n dimension of product is overall, therefrom obtain sample data

(x ₁₁,x ₁₂,...,x _1n) ^T，(x ₂₁,x ₂₂,...,x _2n) ^T，….(x _N1,x _N2,...,x _Nn) ^T

In process of producing product, i observation data is designated as

X _i＝(x _i1,x _i2,...,x _in) ^T i＝1,2,...,N (6)

Formula (6) represents multiple features of single measurement product.Introduce data observation matrix

$X={\left[\begin{array}{cccc}{x}_{11}& x_{12}& \·\·\·& x_{1n}\\ x_{21}& x_{22}& \·\·\·& x_{2n}\\ \·\·\·& \·\·\·& \·\·\·& \·\·\·\\ x_{N1}& x_{N2}& \·\·\·& x_{\mathrm{Nn}}\end{array}\right]}_{N\×n}=[X_{\left(1\right)},X_{\left(2\right)}...X_{\left(n\right)}]---\left(7\right)$

Formula (7) is N × n matrix, and N is capable is N sample X ₁, X ₂...., X _n, composition is from the overall (X of n dimension of the many features of product ₁, X ₂...., X _n) sample.The n row of observing matrix X are respectively n variable X ₍₁₎, X ₍₂₎...., X _(n)the value of getting in N test.Be designated as

X _j＝(x _1j,x _2j,...,x _Nj) ^T j＝1,2,...,N (8)

2. the Euclidean distance method based on class center is estimated.

Be provided with M classification: w ₁, w ₂... w _m, every class has N _iindividual sample, is expressed as for any sample X=(x to be identified ₁, x ₂..., x _n), calculate distance wherein it is i Lei Lei center.If relatively X meets to all kinds of distances:

$d(X,\stackrel{\&OverBar;}{X^{\left(w_i\right)}})<d(X,\stackrel{\&OverBar;}{X^{\left(w_j\right)}}),j=\mathrm{1,2},...,M,i\&NotEqual;j$ X ∈ w _i(9)

3. improved coefficient of variation weighted method.

According to the weight of each feature of statistical computation, the historical data of certain feature changes comparatively stable in production run, and it is larger that the distance in batch between part more approaches the shared weight of its state, establishes X _ij(i=1,2..., n; J=1,2..., m) be the historical data of i feature; Wherein X _j(j=1,2..., m) is the historical data of j feature:

Average is $\stackrel{\&OverBar;}{x_j}=\frac{1}{N}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_{\mathrm{ij}}---\left(10\right)$

Variance is: $s_i=\sqrt{\left(\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(x_i-\stackrel{\&OverBar;}{x_j})}^2\right)}---\left(11\right)$

The corresponding coefficient of variation is: $v_j=\frac{s_j}{\stackrel{\&OverBar;}{x_j}}---\left(12\right)$

For product feature X _j(j=1,2..., m), according to the historical data of evaluation object, the less corresponding weight of this each attribute change amplitude degree is: weight is i the shared weight of property value of many features in process of producing product,

Step 3, the weight of part, length, intensity, skin hardness and coating thickness feature are designated as: wherein, d _i' be actual measurement feature, feature and the historical data of repeatedly measuring the mark that lost efficacy obtain one group of distance, and the eigenwert of product is designated as: { d _ij| i=1 ..., 5, j=1 ..., n}, wherein, n is the historical number of measuring of feature.Five reliability partition functions of its correspondence, are designated as: m ₁, m ₂..., m ₅.The task that basic reliability is distributed is exactly to be m ₁, m ₂..., m ₅the probable value that corresponding power set element is given.Its range averaging value is designated as standard scale feature is all dull and independently, the relation that size and the power of supporting evidence are inversely proportional to.If data corresponding to actual measurement feature think that close to historical data supporting evidence is strong, otherwise larger with gap, support more weak.Degree of confidence for the burnt first A of i feature is distributed (Bel _i(A) be):

${\mathrm{Bel}}_i\left(A\right)=m_i\left(A\right)=\begin{array}{cc}\frac{d_{\min }}{{\stackrel{\&OverBar;}{d}}_i}& \frac{d_{\min }}{{\stackrel{\&OverBar;}{d}}_i}\&le;1---\left(13\right)\end{array}$

The nearest similarity of each feature and historical data is higher, is interpreted as likelihood degree maximum, otherwise larger with Euclidean distance, likelihood support is more weak, wherein what standard technology specified is the distance reasonable change value of each feature.Likelihood degree for the burnt first A of i feature distributes (Pl _i(A)), constructing basic reliability for each part feature distributes:

From the interval relation of evidence, the interval reliability value of refusal evidence is: m _i(B)=1-Pl _i(A); The interval reliability value of uncertain evidence is: m _i(Θ)=1-m _i(A)-m _i(B).

Adopt fusion formula to carry out Fusion Features: to establish Bel ₁and Bel ₂for two belief functions of same identification framework Θ, m ₁and m ₂be respectively its corresponding elementary probability assignment, burnt unit is respectively A ₁..., A _kand B ₁..., B _r, combinatorial formula is m=m ₁⊕ m ₂determined by following formula:

$m\left(C\right)=\left\{\begin{array}{cc}{(1-K_1)}^{-1}\underset{A_i\&cap;B_j=C}{\mathrm{\&Sigma;}}{m}_1\left(A_i\right)m_2\left(B_j\right),& \&ForAll;C\&Element;\mathrm{\&Theta;},C\&NotEqual;\mathrm{\&phi;}\\ 0,& C=\mathrm{\&phi;}\end{array}\right.---\left(15\right)$

Wherein, K ₁represent the size of conflicting between evidence, if K ₁=1 shows m ₁and m ₂contradiction, can not combine elementary probability assignment completely.For many Evidence Combination Methods, combine by team.

The invention has the beneficial effects as follows: the associated and data structure in the method analytic product physics and manufacture process between multi-source information, set up in Discrete Manufacturing Process the scattered and imperfect multi-source information model of product, by the Euclidean distance based on class center and improved coefficient of variation weighted method, lost efficacy the mark characteristic of part and the historical data base similarity of product are estimated to calculating, finally adopt D-S evidence theory to merge identification to its characteristic layer, effectively recover to remedy the mark of inefficacy product, and then prevent DM code lost efficacy and cause product individuality to follow the trail of and open circuit in process of production.Experiment showed, that the inventive method prevents from following the trail of the recognition correct rate losing efficacy more than 95%, efficiency is high, recovers to remedy reliably, has good practicality.

Below in conjunction with the drawings and specific embodiments, the present invention is elaborated.

Brief description of the drawings

Fig. 1 the present invention is based on the process flow diagram of the many information fusion marks of the product inefficacy means to save the situation of D-S evidence theory.

Embodiment

With reference to Fig. 1.The many information fusion marks of the product inefficacy means to save the situation concrete steps that the present invention is based on D-S evidence theory are as follows:

Step 1, set up the scattered and imperfect multi-source information model of product.

Analyze the information source relevant to product identification, comprise product physical message and manufacture process information, and many information fusion are incorporated into product represent to lose efficacy and remedy, for the ease of the mathematical description to product multi-source information, if a multi-source information model, wherein object set is

P＝{x ₁,x ₂,x ₃,.....x _i} (1)

The main body of this multi-source information, element x _i, i={1,2,3 ... ..n} represents different information sources; Property set is

C＝{c ₁,c ₂,c ₃,.....c _j} (2)

It is the description of product external information attribute; The codomain of attribute is discrete values, and is designated as U _m(m≤j), the set of relations of object set P and property set C is

F＝{f _i:m≤j} (3)

Wherein, f _m: P → U _m(m≤j), expressed contacting between object set U and property set C is the basis of information source.Establish again the status attribute collection of E for being determined by property set C,

E＝{e ₁,e ₂,e ₃,.....e _j} (4)

Formula (4) represents the state that each attribute of product object comprises, its codomain U _m' (m≤j) is that qualitative value specified states is divided into mark and lost efficacy and do not lose efficacy, and property set C is called conditional attribute collection.Product multi-source information model represents with following formula:

I＝{P,C,F,E} (5)

For the outside multiple information sources of product arbitrarily, above-mentioned model representation becomes a two-dimensional data table.Before note does not occur to lose efficacy, part is numbered A, B, and C, after mark loses efficacy, part was designated as 1,2,3.The scattered Incomplete information historical record of product of the mark that lost efficacy, as entered, comprises that the characteristic parameter of weight, part length, intensity, skin hardness, coating thickness is as shown in table 1:

The table 1 product historical multi-source information that do not lose efficacy

Step 2, the many characteristic similarities of product are estimated calculating.

In the time that mark occurs to lose efficacy, by scattered and incomplete product physics and manufacturing information collection, merge the requirement of identification to meet product, similarity between judgement sample often adopts neighbour's criterion, the degree that adopts the mark that lost efficacy to mate with center different classes of in Sample Storehouse is high, repeatedly measures the correlated characteristic of inefficacy product and carries out calculating with the similarity measure of historical record data.

1. the many characteristic models of product based on record.

If (X ₍₁₎, X ₍₂₎...., X _(n)) be that the multi-source information n dimension of product is overall, therefrom obtain sample data

(x ₁₁,x ₁₂,...,x _1n) ^T，(x ₂₁,x ₂₂,...,x _2n) ^T，….(x _N1,x _N2,...,x _Nn) ^T

In process of producing product, i observation data is designated as

X _i＝(x _i1,x _i2,...,x _in) ^T i＝1,2,...,N (6)

Represent multiple features of single measurement product.Introduce data observation matrix

$X={\left[\begin{array}{cccc}{x}_{11}& x_{12}& \&CenterDot;\&CenterDot;\&CenterDot;& x_{1n}\\ x_{21}& x_{22}& \&CenterDot;\&CenterDot;\&CenterDot;& x_{2n}\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ x_{N1}& x_{N2}& \&CenterDot;\&CenterDot;\&CenterDot;& x_{\mathrm{Nn}}\end{array}\right]}_{N\&times;n}=[X_{\left(1\right)},X_{\left(2\right)}...X_{\left(n\right)}]---\left(7\right)$

Be N × n matrix, N is capable is N sample X ₁, X ₂...., X _n, composition is from the overall (X of n dimension of the many features of product ₁, X ₂...., X _n) sample.The n row of observing matrix X are respectively n variable X ₍₁₎, X ₍₂₎...., X _(n)the value of getting in N test.Be designated as

X _j＝(x _1j,x _2j,...,x _Nj) ^T j＝1,2,...,N (8)

2. the Euclidean distance method based on class center is estimated.

Be provided with M classification: w ₁, w ₂... w _m, every class has N _iindividual sample, is expressed as for any sample X=(x to be identified ₁, x ₂..., x _n), calculate distance wherein it is i Lei Lei center.If relatively X meets to all kinds of distances:

$d(X,\stackrel{\&OverBar;}{X^{\left(w_i\right)}})<d(X,\stackrel{\&OverBar;}{X^{\left(w_j\right)}}),j=\mathrm{1,2},...,M,i\&NotEqual;j$ X ∈ w _i(9)

3. improved coefficient of variation weighted method.

According to the weight of each feature of statistical computation, the historical data of certain feature changes comparatively stable in production run, and it is larger that the distance in batch between part more approaches the shared weight of its state, establishes X _ij(i=1,2..., n; J=1,2..., m) be the historical data of i feature; Wherein X _j(j=1,2..., m) is the historical data of j feature:

Average is $\stackrel{\&OverBar;}{x_j}=\frac{1}{N}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_{\mathrm{ij}}---\left(10\right)$

Variance is: $s_i=\sqrt{\left(\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(x_i-\stackrel{\&OverBar;}{x_j})}^2\right)}---\left(11\right)$

The corresponding coefficient of variation is: $v_j=\frac{s_j}{\stackrel{\&OverBar;}{x_j}}---\left(12\right)$

For product feature X _j(j=1,2..., m), according to the historical data of evaluation object, the less corresponding weight of this each attribute change amplitude degree is: weight is i the shared weight of property value of many features in process of producing product,

According to the weight coefficient of the coefficient of variation weighted method of history making Information Statistics analysis and modification calculate, determine that weight in multi-source information feature, length, intensity, skin hardness, the shared weight of coating thickness are respectively 0.09,0.11,0.21,0.36,0.23.

If d _i' be actual measurement feature, repeatedly to measure the mark that lost efficacy and obtain one group of distance with historical data, the eigenwert of product is designated as: { d _ij| i=1, ..., 5, j=1, ..., n}, the present embodiment choose that 3 different inefficacies mark parts are repeatedly measured and with historical part A, B, the C Euclidean distance of the coefficient of variation weighted method of computed improved respectively, inefficacy part 15 pattern measurements corresponding with it 10 secondary data become power Euclidean distances, and mean value and the minimum distance value of asking for its character pair are as follows:

$\left\{\begin{array}{cc}\stackrel{\&OverBar;}{d_1}=7.2& d_{1\min }=4\\ \stackrel{\&OverBar;}{d_2}=0.22& d_{2\min }=0.11\\ \stackrel{\&OverBar;}{d_3}=0.98& d_{3\min }=0.21\\ \stackrel{\&OverBar;}{d_4}=3& d_{4\min }=1.44\\ \stackrel{\&OverBar;}{d_5}=0.84& d_{5\min }=0.23\end{array}\right.,$ $\left\{\begin{array}{cc}\stackrel{\&OverBar;}{d_1}=12.87& d_{1\min }=5.76\\ \stackrel{\&OverBar;}{d_2}=7.81& d_{2\min }=5.39\\ \stackrel{\&OverBar;}{d_3}=4.9& d_{3\min }=1.89\\ \stackrel{\&OverBar;}{d_4}=19.44& d_{4\min }=17.64\\ \stackrel{\&OverBar;}{d_5}=4.29& d_{5\min }=0.92\end{array}\right.,$ $\left\{\begin{array}{cc}\stackrel{\&OverBar;}{d_1}=10.23& d_{1\min }=5.76\\ \stackrel{\&OverBar;}{d_2}=8.43& d_{2\min }=5.39\\ \stackrel{\&OverBar;}{d_3}=3.57& d_{3\min }=0.84\\ \stackrel{\&OverBar;}{d_4}=14.52& d_{4\min }=12.96\\ \stackrel{\&OverBar;}{d_5}=4.29& d_{5\min }=0.92\end{array}\right.,$

Step 3, many Fusion Features DM code identification based on D-S theory.

The feature of the part of the present embodiment comprises weight, part length, intensity, skin hardness, five features of coating thickness, and five features are designated as: wherein d _i' be actual measurement feature, feature and the historical data of repeatedly measuring the mark that lost efficacy obtain one group of distance, and the eigenwert of product is designated as: { d _ij| i=1 ..., 5, j=1 ..., n}, wherein n is the historical number of measuring of feature.Five reliability partition functions of its correspondence, are designated as: m ₁, m ₂..., m ₅.The task that basic reliability is distributed is exactly to be m ₁, m ₂..., m ₅corresponding power set element is given appropriate, conforms with the probable value of explanation.Its range averaging value is designated as standard scale feature is all dull and independently, the relation that size and the power of supporting evidence are inversely proportional to.If data corresponding to actual measurement feature can think that close to historical data supporting evidence is strong, otherwise larger with gap, support more weak.Be designated as the front part that do not lose efficacy and be numbered A, B, C, after mark loses efficacy, part was designated as 1,2,3.Repeatedly measure the mark that lost efficacy and obtain one group of distance with historical data, calculate the power that becomes power Euclidean distance supporting evidence by similarity measure.So the degree of confidence for the burnt first A of i feature is distributed (Bel _i(A) be):

${\mathrm{Bel}}_i\left(A\right)=m_i\left(A\right)=\begin{array}{cc}\frac{d_{\min }}{{\stackrel{\&OverBar;}{d}}_i}& \frac{d_{\min }}{{\stackrel{\&OverBar;}{d}}_i}\&le;1---\left(13\right)\end{array}$

The nearest similarity of each feature and historical data is higher, is interpreted as likelihood degree maximum, otherwise larger with Euclidean distance, likelihood support is more weak, wherein standard technology regulation be the distance reasonable change value of each feature, such as the undulating quantity of machining tolerance, weight etc.Arrange what standard technology specified is the distance reasonable change value of each feature, (is designated as respectively ), distribute (Pl for the likelihood degree of the burnt first A of i feature _i(A)), for the basic reliability of each latent structure of the example of each part is distributed:

From the interval relation of evidence, the interval reliability value of refusal evidence is: m _i(B)=1-Pl _i(A); The interval reliability value of uncertain evidence is: m _i(Θ)=1-m _i(A)-m _i(B).

Adopt fusion formula to carry out Fusion Features: to establish Bel ₁and Bel ₂for two belief functions of same identification framework Θ, m ₁and m ₂be respectively its corresponding elementary probability assignment, burnt unit is respectively A ₁..., A _kand B ₁..., B _r, combinatorial formula is m=m ₁⊕ m ₂determined by following formula:

$m\left(C\right)=\left\{\begin{array}{cc}{(1-K_1)}^{-1}\underset{A_i\&cap;B_j=C}{\mathrm{\&Sigma;}}{m}_1\left(A_i\right)m_2\left(B_j\right),& \&ForAll;C\&Element;\mathrm{\&Theta;},C\&NotEqual;\mathrm{\&phi;}\\ 0,& C=\mathrm{\&phi;}\end{array}\right.---\left(15\right)$

Wherein K ₁represent the size of conflicting between evidence, if K ₁=1 shows m ₁and m ₂contradiction, can not combine elementary probability assignment completely.For many Evidence Combination Methods, utilize said method to combine by team.

The interval reliability value of refusal evidence is: m _i(B)=1-Pl _i(A) thus, be, constructing the following table is the basic reliability of this example with part A, B, the each feature of C of part 1:

The part 1 of table 2 inefficacy mark distributes with part A feature degree of confidence

Calculate the identification losing efficacy between mark part 1 and part A, according to Dempster fusion rule, first merge weight characteristics and length characteristic, the result of establishing after merging is m ⁽¹⁾(A), m ⁽¹⁾(B), m ⁽¹⁾(C):

K ⁽¹⁾＝m ₁(A)m ₂(B)+m ₁(B)m ₂(A)＝0.4404

$\begin{array}{c}{m}^{\left(1\right)}\left(A\right)=\frac{m_1\left(A\right)m_2\left(A\right)+m_1\left(A\right)m_2\left(C\right)+m_1\left(C\right)m_2\left(A\right)}{1-K^{\left(1\right)}}\\ =\frac{0.44\&times;0.5+0.44\&times;0.04+0.02\&times;0.5}{1-0.4404}=0.411\end{array}$

$\begin{array}{c}{m}^{\left(1\right)}\left(B\right)=\frac{m_1\left(B\right)m_2\left(B\right)+m_1\left(B\right)m_2\left(C\right)+m_1\left(C\right)m_2\left(B\right)}{1-K^{\left(1\right)}}\\ =\frac{0.52\&times;0.41+0.52\&times;0.09+0.04\&times;0.41}{1-0.4404}=0.4939\end{array}$

m ⁽¹⁾(C)＝0.095

The part 1 and part A Fusion Features result of table 3 inefficacy mark

By the result m after merging ⁽¹⁾(A), m ⁽¹⁾(B), m ⁽¹⁾and m (C) ₃(A), m ₃(B), m ₃(C) carry out Dempster fusion, this merges whole features by team class, and part 1 obtains last fusion results with part A and is: m ⁽⁴⁾(A)=0.931, m ⁽⁴⁾(B)=0.068, m ⁽⁴⁾(C)=0.001.This calculates the identification losing efficacy between mark part 1 and part B, m class ⁽⁴⁾(A)=0.338, m ⁽⁴⁾(B)=0.662, m ⁽⁴⁾(C)=0.This calculates the identification losing efficacy between mark part 1 and part C, m class ⁽⁴⁾(A)=0.488, m ⁽⁴⁾(B)=0.512, m ⁽⁴⁾(C)=0.The probability that after Fusion Features, the support of target draws than original single features is all high, the matching degree most significant digit 0.931 of the mark that lost efficacy part 1 and part A, the matching degree most significant digit 0.338 of the mark that lost efficacy part 1 and part B, the matching degree most significant digit 0.488 of the mark that lost efficacy part 1 and part C, after too much Fusion Features, the part 1 of the mark that lost efficacy is the highest with part A identification, and uncertainty reduces, can recover to remedy the mark of part, continue to merge correlated characteristic and improve the reliability of target identification.

Claims (1)

1. the many information fusion marks of the product based on a D-S evidence theory inefficacy means to save the situation, is characterized in that comprising the following steps:

Step 1, the information source that analysis is relevant to product identification, comprise product physical message and manufacture process information, and many information fusion are incorporated into product represent to lose efficacy and remedy; If a multi-source information model, wherein object set is

P＝{x ₁,x ₂,x ₃,.....x _i} (1)

The main body of this multi-source information, element x _i, i={1,2,3 ... ..n} represents different information sources; Property set is

C＝{c ₁,c ₂,c ₃,.....c _j} (2)

It is the description of product external information attribute; The codomain of attribute is discrete values, is designated as U _m(m≤j), wherein m represents the attribute of product different aforementioned sources, and j represents all properties of product different aforementioned sources, and the set of relations of object set P and property set C is

F＝{f _i:m≤j} (3)

Wherein, f _m: P → U _m(m≤j), f _irepresent product multi-source information object set, f _mrepresenting the property set of actual different aforementioned sources, expressed contacting between object set U and property set C, is the basis of information source; Establish again the status attribute collection of E for being determined by property set C,

E＝{e ₁,e ₂,e ₃,.....e _j} (4)

Formula (4) represents the state that each attribute of product object comprises, its codomain U _m' (m≤j) is that qualitative value specified states is divided into mark and lost efficacy and do not lose efficacy, and property set C is called conditional attribute collection; Product multi-source information model represents with following formula:

I＝{P,C,F,E} (5)

For the outside multiple information sources of product arbitrarily, above-mentioned model representation becomes a two-dimensional data table;

Step 2, when mark occur lost efficacy time, gather scattered and incomplete product physics and manufacturing information, merge the requirement of identification to meet product, similarity between judgement sample often adopts neighbour's criterion, the degree that adopts the mark that lost efficacy to mate with center different classes of in Sample Storehouse is high, repeatedly measures the correlated characteristic of inefficacy product and carries out calculating with the similarity measure of historical record data;

1. the many characteristic models of product based on record;

If (X ₍₁₎, X ₍₂₎...., X _(n)) be that the multi-source information n dimension of product is overall, therefrom obtain sample data

(x ₁₁,x ₁₂,...,x _1n) ^T，(x ₂₁,x ₂₂,...,x _2n) ^T，….(x _N1,x _N2,...,x _Nn) ^T

In process of producing product, i observation data is designated as

X _i＝(x _i1,x _i2,...,x _in) ^T i＝1,2,...,N (6)

Formula (6) represents multiple features of single measurement product; Introduce data observation matrix

$X={\left[\begin{array}{cccc}{x}_{11}& x_{12}& \&CenterDot;\&CenterDot;\&CenterDot;& x_{1n}\\ x_{21}& x_{22}& \&CenterDot;\&CenterDot;\&CenterDot;& x_{2n}\\ \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;& \&CenterDot;\&CenterDot;\&CenterDot;\\ x_{N1}& x_{N2}& \&CenterDot;\&CenterDot;\&CenterDot;& x_{\mathrm{Nn}}\end{array}\right]}_{N\&times;n}=[X_{\left(1\right)},X_{\left(2\right)}...X_{\left(n\right)}]---\left(7\right)$

Formula (7) is N × n matrix, and N is capable is N sample X ₁, X ₂...., X _n, composition is from the overall (X of n dimension of the many features of product ₁, X ₂...., X _n) sample; The n row of observing matrix X are respectively n variable X ₍₁₎, X ₍₂₎...., X _(n)the value of getting in N test; Be designated as

X _j＝(x _1j,x _2j,...,x _Nj) ^T j＝1,2,...,N (8)

2. the Euclidean distance method based on class center is estimated;

Be provided with M classification: w ₁, w ₂... w _m, every class has N _iindividual sample, is expressed as for any sample X=(x to be identified ₁, x ₂..., x _n), calculate distance wherein it is i Lei Lei center; If relatively X meets to all kinds of distances:

$d(X,\stackrel{\&OverBar;}{X^{\left(w_i\right)}})<d(X,\stackrel{\&OverBar;}{X^{\left(w_j\right)}}),j=\mathrm{1,2},...,M,i\&NotEqual;j$ X ∈ w _i(9)

3. improved coefficient of variation weighted method;

According to the weight of each feature of statistical computation, the historical data of certain feature changes comparatively stable in production run, and it is larger that the distance in batch between part more approaches the shared weight of its state, establishes X _ij(i=1,2..., n; J=1,2..., m) be the historical data of i feature; Wherein X _j(j=1,2..., m) is the historical data of j feature:

Average is $\stackrel{\&OverBar;}{x_j}=\frac{1}{N}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_{\mathrm{ij}}---\left(10\right)$

Variance is: $s_i=\sqrt{\left(\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{(x_i-\stackrel{\&OverBar;}{x_j})}^2\right)}---\left(11\right)$

The corresponding coefficient of variation is: $v_j=\frac{s_j}{\stackrel{\&OverBar;}{x_j}}---\left(12\right)$

For product feature X _j(j=1,2..., m), according to the historical data of evaluation object, the less corresponding weight of this each attribute change amplitude degree is: weight is i the shared weight of property value of many features in process of producing product,

Step 3, the weight of part, length, intensity, skin hardness and coating thickness feature are designated as: wherein, d _i' be actual measurement feature, feature and the historical data of repeatedly measuring the mark that lost efficacy obtain one group of distance, and the eigenwert of product is designated as: { d _ij| i=1 ..., 5, j=1 ..., n}, wherein, n is the historical number of measuring of feature; Five reliability partition functions of its correspondence, are designated as: m ₁, m ₂..., m ₅; The task that basic reliability is distributed is exactly to be m ₁, m ₂..., m ₅the probable value that corresponding power set element is given; Its range averaging value is designated as standard scale feature is all dull and independently, the relation that size and the power of supporting evidence are inversely proportional to; If data corresponding to actual measurement feature think that close to historical data supporting evidence is strong, otherwise larger with gap, support more weak; Degree of confidence for the burnt first A of i feature is distributed (Bel _i(A) be):

${\mathrm{Bel}}_i\left(A\right)=m_i\left(A\right)=\begin{array}{cc}\frac{d_{\min }}{{\stackrel{\&OverBar;}{d}}_i}& \frac{d_{\min }}{{\stackrel{\&OverBar;}{d}}_i}\&le;1---\left(13\right)\end{array}$

The nearest similarity of each feature and historical data is higher, is interpreted as likelihood degree maximum, otherwise larger with Euclidean distance, likelihood support is more weak, wherein what standard technology specified is the distance reasonable change value of each feature; Likelihood degree for the burnt first A of i feature distributes (Pl _i(A)), constructing basic reliability for each part feature distributes:

From the interval relation of evidence, the interval reliability value of refusal evidence is: m _i(B)=1-Pl _i(A); The interval reliability value of uncertain evidence is: m _i(Θ)=1-m _i(A)-m _i(B);

Adopt fusion formula to carry out Fusion Features: to establish Bel ₁and Bel ₂for two belief functions of same identification framework Θ, m ₁and m ₂be respectively its corresponding elementary probability assignment, burnt unit is respectively A ₁..., A _kand B ₁..., B _r, combinatorial formula is m=m ₁⊕ m ₂determined by following formula:

$m\left(C\right)=\left\{\begin{array}{cc}{(1-K_1)}^{-1}\underset{A_i\&cap;B_j=C}{\mathrm{\&Sigma;}}{m}_1\left(A_i\right)m_2\left(B_j\right),& \&ForAll;C\&Element;\mathrm{\&Theta;},C\&NotEqual;\mathrm{\&phi;}\\ 0,& C=\mathrm{\&phi;}\end{array}\right.---\left(15\right)$

Wherein, K ₁represent the size of conflicting between evidence, if K ₁=1 shows m ₁and m ₂contradiction, can not combine elementary probability assignment completely; For many Evidence Combination Methods, combine by team.