CN105654175A - Part supplier multi-target preferable selection method orienting bearing manufacturing enterprises - Google Patents

Abstract

The invention provides a part supplier multi-target preferable selection method orienting bearing manufacturing enterprises. The degree of capacity of each expert is acquired from historical samples through a neural network training method by considering objective difference of the experience levels and the knowledge levels of different experts; the degree of capacity of the experts is introduced and then a multi-order fuzzy rough set integrated analytic hierarchy process with combination of an analytic hierarchy process in which fuzzy numbers replace accurate numbers and rough sets is provided so that combined processing of multiple experts for multiple part suppliers on the score result of a single indicator is realized; a multi-weight information integration model is established, and integration weights of the evaluation indicators are solved; and the part suppliers are ranked according to the indicator value of each evaluation indicator and the integration weights of the indicators of multiple part suppliers, and the optimal part supplier is obtained.

Description

A kind of parts supplier multiple goal preferred method towards bearing mnanufacture enterprise

One, art

The present invention relates to the Supplier Selection field in enterprise supply chain management, it is specially a kind of parts supplier multiple goal preferred method towards bearing mnanufacture enterprise.

Two, background technology

Bearing is formed primarily of parts such as roller, lasso, sealing-ring, maintenance frame, dust guards. Along with being growing more intense of market competition, strength is more and more concentrated on self core business by bearing mnanufacture enterprise, and provides a large amount of part by outside supplier, such as spin, sealing-ring etc. The quality of finished product is had direct impact by these outsourcing parts, how to carry out preferably having become the key that the market competitiveness improves in bearing mnanufacture enterprise to parts supplier.

Parts supplier for bearing mnanufacture enterprise is selected, current evaluation method is calculated as master with the quantification of specific targets and tax power, but the attention rate of parts supplier evaluation index is had notable difference by different bearing mnanufacture enterprises, this kind of method is only applicable to the bearing mnanufacture enterprise of certain particular type, and replicability is poor; Index compose temporary use single weight or multiple weight on average carry out parameter weight, ignore the syntagmatic between multiple weight. In addition, method of expertise also has application to a certain degree, and this kind of method, taking the experience ability of expert as core, has certain saving grace, but directly application often makes preferred effect poor; All there is difference in the experience level of expert, know-how etc., the selection result of different expert comes in and goes out very big; Numerous evaluation index also is difficult to take into account. Bearing as the huge basic equipment of production value, its manufacturing concern selecting to consider green effect during parts supplier, i.e. the feature such as the environmental influence of parts supplier and optimizing network resource utilization.

Three, summary of the invention

Based on the problems referred to above, the present invention considers the objective difference of different expertise level, know-hows etc., is obtained the energy dynamics of each expert from historical sample by the method for neural network training; After introducing experts ability degree, it is provided that a kind of analytical hierarchy process that fuzzy number substitutes accurately number assembles, with thick, the comprehensive analytical hierarchy process of many rank Fuzzy Rough Sets closed, it is achieved several expert is to the combined treatment of many parts suppliers appraisal result in single index; Set up many weight informations integrated model, solve the integrated weight of evaluation index; According to the desired value of many parts suppliers in each evaluation index and index integration weight, parts supplier is sorted, thus preferably provide a kind of scientific and reasonable solution for the parts supplier of bearing mnanufacture enterprise.

The technical scheme of the present invention is:

Described a kind of parts supplier multiple goal preferred method towards bearing mnanufacture enterprise, it is characterised in that: comprise the steps:

Step 1: preferred for the parts supplier of bearing mnanufacture enterprise general objective is decomposed into five sub-goals: quality sub-goal, deliverability sub-goal, economic situation sub-goal, partner services sub-goal and environmental influence sub-goal, it is corresponding in turn to five evaluation indexes, uses Index successively₁, Index₂..., Index_mRepresent, m=5;

Step 2: the energy dynamics being obtained each expert by the method for neural network training from historical sample:

S (s > 1) individual expert is to r object O₁, O₂..., O_rSort, the total individual sample of N '. Experts ability degree vector is $\mathrm{\ρ}={\[{\mathrm{\ρ}}_1,{\mathrm{\ρ}}_2,...,{\mathrm{\ρ}}_s\]}^T\left(\underset{i=1}{\overset{s}{\Σ}}{\mathrm{\ρ}}_i=1,0\≤{\mathrm{\ρ}}_i\≤1\right).$

Sample N (N=1,2 ..., N ') in, expert i (i=1,2 ..., ranking results s) is:WithRepresentIn be in O_jObject number afterwards is (containing O_jSelf), also represent O_j?In score. In sample N, useTo O₁, O₂..., O_rFinally sorting, s expert is to O₁, O₂..., O_rFinal ranking results represent and be:

O₁, O₂..., O_rTrue ranking results be:

Set up the neural network structure calculating experts ability degree:

At input layer, neuron number is s �� r, and input matrix isEqual

At hidden layer, neuron number is r (equal with object number). For neurone j (i.e. object j, j=1,2 ..., r), it is input asOutput isConnection weight matrix between input layer and hidden layer node can be divided into s sub-matrix-block, wherein submatrix block i (i=1,2 ..., s) be ${\mathrm{\Γ}}_i^{\left(N\right)}=\left[\begin{array}{cccc}\underset{i}{{\mathrm{\γ}}_{11}^{\left(N\right)}}& 0& \mathrm{...}& 0\\ 0& \underset{i}{{\mathrm{\γ}}_{22}^{\left(N\right)}}& \mathrm{...}& 0\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ 0& 0& \mathrm{...}& \underset{i}{{\mathrm{\γ}}_{rr}^{\left(N\right)}}\end{array}\right];$

At output layer, only 1 neurone, hidden layer node and the connection weight being output between node are vectorial is (1,1 ..., 1)^T��

The deduction process of the neural network calculating experts ability degree is:

For hidden layer node j (i.e. object j, j=1,2 ..., r), it is input asOutput is ${\mathrm{dout}}_j^{\left(N\right)}=f({\mathrm{yin}}_j^{\left(N\right)}+{\mathrm{\θ}}_j),$ Action function is $f\left(y\right)=\frac{1}{1+e^{-y}},$

For output layer node, it is input asOutput is ${\mathrm{bout}}^{\left(N\right)}=({\mathrm{bout}}_1^{\left(N\right)},{\mathrm{bout}}_2^{\left(N\right)},...,{\mathrm{bout}}_r^{\left(N\right)}),{\mathrm{bout}}_j^{\left(N\right)}$ (j=1,2 ..., r) it is object O_j? ${\mathrm{LI}}^{\left(N\right)}:O_1^{\left(N\right)},O_2^{\left(N\right)},...,O_r^{\left(N\right)}$ In score, ${\mathrm{LI}}^{\left(N\right)}:O_1^{\left(N\right)},O_2^{\left(N\right)},...,O_r^{\left(N\right)}$ By ${\mathrm{dout}}_1^{\left(N\right)},{\mathrm{dout}}_2^{\left(N\right)},...,{\mathrm{dout}}_r^{\left(N\right)}$ Sort and obtain.

For error function, the error of sample N isIt is object O_j?In score,It is object O_j?In score.

The experts ability degree obtained (namely connection weight) between input layer and hidden layer node is normalized: for submatrix block after completing by the neural network training calculating experts ability degreeCalculate the mean value of its diagonal lines elementCalculate successivelyAfter be normalized, so the energy dynamics of expert i isFinally calculate experts ability degree vector for ��=[��₁, ��₂..., ��₃]��

Step 3: introduce experts ability degree, by the comprehensive analytical hierarchy process of many rank Fuzzy Rough Sets, it is achieved several expert is to the combined treatment of many parts suppliers appraisal result in single index:

Having n family's parts supplier, s expert participates in scoring, and evaluation indice is Index={Index₁, Index₂..., Index_m, m=5.

Sub-step 3-1: at index Index_tOn �� Index, the rating matrix of n family's parts supplier, on the basis grasping history supply situation and investigation, is followed successively by by s expertWherein expert i (i=1,2 ..., s) to the rating matrix of n family's parts supplier be: $\widetilde{H^{i,t}}=\left[\begin{array}{cccc}(1,1,1,1)& \widetilde{H_{1,2}^{i,t}}& \mathrm{...}& \widetilde{H_{1,n}^{i,t}}\\ \widetilde{H_{2,1}^{i,t}}& (1,1,1,1)& \mathrm{...}& \widetilde{H_{2,n}^{i,t}}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ \widetilde{H_{n,1}^{i,t}}& \widetilde{H_{n,2}^{i,t}}& \mathrm{...}& (1,1,1,1)\end{array}\right],$ Represent expert i provide at index Index_tUpper parts supplier q (q=1,2 ..., n) relative to parts supplier p (p=1,2 ..., n) trapezoidal fuzzy number scoring, p �� q. S rating matrix is carried out consistency check respectively, consistency check method is obtained by method well known in the art, reference: Luo Zhimeng, Zhou Jianzhong, Yang person of outstanding talent, Deng. based on the Virtual Research Centers partner selection [J] of fuzzy AHP. Central China University of Science and Technology's journal: natural science edition, 2008,36 (12): 100-103. If s rating matrix is all by consistency check, perform sub-step 3-2; Otherwise go back to sub-step 3-1, defective rating matrix is marked again by the expert of correspondence.

Sub-step 3-2: decomposeFor A^{I, t}, B^{I, t}, C^{I, t}, D^{I, t}.Wherein

$A^{i,i}=\left[\begin{array}{cccc}1& a_{1,2}^{i,t}& \mathrm{...}& a_{1,n}^{i,t}\\ a_{2,1}^{i,t}& 1& \mathrm{...}& a_{2,n}^{i,t}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ a_{n,1}^{i,t}& a_{n,2}^{i,t}& \mathrm{...}& 1\end{array}\right],$

Other is analogized with this.

Sub-step 3-3: according to A^{1, t}, A^{2, t}..., A^{S, t}Build coarse group decision matrix: $A^t=\left[\begin{array}{cccc}1& A_{1,2}^t& \mathrm{...}& A_{1,n}^t\\ A_{2,1}^t& 1& \mathrm{...}& A_{2,n}^t\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ A_{n,1}^t& A_{n,2}^t& \mathrm{...}& 1\end{array}\right],$ Wherein, $\begin{array}{c}{A}_{p,q}^t=\{a_{p,q}^{1,t},a_{p,q}^{2,t},\mathrm{...},a_{p,q}^{s,t}\},\\ p\≠q,\\ p=1,2,\mathrm{...},n,\\ q=1,2,\mathrm{...},n\end{array}.$

So,The coarse number corresponding by s expert analysis mode data comprised is: $RN\left(a_{p,q}^{i,t}\right)=\[a_{p,q}^{i-,t},a_{p,q}^{i+,t}\],i=1,2,...,s.$ So, $RN\left(A_{p,q}^t\right)=\{\[a_{p,q}^{1-,t},a_{p,q}^{1+,t}\],\[a_{p,q}^{2-,t},a_{p,q}^{2+,t}\],...,\[a_{p,q}^{q-,t},a_{p,q}^{q+,t}\]\}.$ Introducing experts ability degree, can try to achieve average roughness interval is $ARN\left(A_{p,q}^t\right)=\[a_{p,q}^{-,t},a_{p,q}^{+,t}\]=\[\frac{\underset{i=1}{\overset{s}{\Σ}}{\mathrm{\ρ}}_i\·a_{p,q}^{i-,t}}{s},\frac{\underset{i=1}{\overset{s}{\Σ}}{\mathrm{\ρ}}_i\·a_{p,q}^{i+,t}}{s}\].$ Try to achieve A successively^tIn the average roughness of other elements interval.

Sub-step 3-4: construct coarse Judgement Matrix ${\mathrm{EA}}^t=\left[\begin{array}{cccc}\[1,1\]& \[a_{1,2}^{-,t},a_{1,2}^{+,t}\]& \mathrm{...}& \[a_{1,n}^{-,t},a_{1,n}^{+,t}\]\\ \[a_{2,1}^{-,t},a_{2,1}^{+,t}\]& \[1,1\]& \mathrm{...}& \[a_{2,n}^{-,t},a_{2,n}^{+,t}\]\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ \[a_{n,1}^{-,t},a_{n,1}^{+,t}\]& \[a_{n,2}^{-,t},a_{n,2}^{+,t}\]& \mathrm{...}& \[1,1\]\end{array}\right].$ By EA^tIt is decomposed into coarse lower boundary matrix EAI^tWith coarse upper boundary matrix EAS^t: ${\mathrm{EAI}}^t=\left[\begin{array}{cccc}1& a_{1,2}^{-,t}& \mathrm{...}& a_{1,n}^{-,t}\\ a_{2,1}^{-,t}& 1& \mathrm{...}& a_{2,n}^{-,t}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ a_{n,1}^{-,t}& a_{n,2}^{-,t}& \mathrm{...}& 1\end{array}\right],$ ${\mathrm{EAS}}^t=\left[\begin{array}{cccc}1& a_{1,2}^{+,t}& \mathrm{...}& a_{1,n}^{+,t}\\ a_{2,1}^{+,t}& 1& \mathrm{...}& a_{2,n}^{+,t}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ a_{n,1}^{+,t}& a_{n,2}^{+,t}& \mathrm{...}& 1\end{array}\right],$ The proper vector corresponding to maximum eigenwert is asked to be VAI respectively^t=[vai₁, vai₂..., vai_n]^T, VAS^t=[vas₁, vas₂..., vas_n]^T, after average it is ${\mathrm{ga}}_p^t=\frac{\left|{\mathrm{vai}}_p\right|+\left|{\mathrm{vas}}_p\right|}{2},$ P=1,2 ..., n, thus have ${\mathrm{GA}}^t=\{{\mathrm{ga}}_1^t,{\mathrm{ga}}_2^t,...,{\mathrm{ga}}_n^t\}.$

Sub-step 3-5: with reason, duplicon step 3-3,3-4, can obtain ${\mathrm{GB}}^t=\{{\mathrm{gb}}_1^t,{\mathrm{gb}}_2^t,...,{\mathrm{gb}}_n^t\},{\mathrm{GC}}^t=\{{\mathrm{gc}}_1^t,{\mathrm{gc}}_2^t,...,{\mathrm{gc}}_n^t\},{\mathrm{GD}}^t=\{{\mathrm{gd}}_1^t,{\mathrm{gd}}_2^t,...,{\mathrm{gd}}_n^t\},$ Thus n family's parts supplier is at index Index_t(t=1,2 ..., the fuzzy value on m) is followed successively by $({\mathrm{ga}}_1^t,{\mathrm{gb}}_1^t,{\mathrm{gc}}_1^t,{\mathrm{gd}}_1^t),({\mathrm{ga}}_2^t,{\mathrm{gb}}_2^t,{\mathrm{gc}}_2^t,{\mathrm{gd}}_2^t),...,({\mathrm{ga}}_n^t,{\mathrm{gb}}_n^t,{\mathrm{gc}}_n^t,{\mathrm{gd}}_n^t).$

Sub-step 3-6: with reason, duplicon step 3-1 to 3-5, can try to achieve the fuzzy value of n family's parts supplier in other index, thus the index value matrix of n family's parts supplier in m index isFor parts supplier p (p=1,2 ..., n) index t (t=1,2 ..., the desired value on m), represents by trapezoidal fuzzy several form.

Step 4: set up many weight informations integrated model, solves the integrated weight of evaluation index:

Sub-step 4-1: the parts supplier index value matrix will tried to achieve in step 3 by center of gravity methodIt is converted into center of gravity form V=(v_{P, t})_n��m, $v_{p,1}=Gra\left(\widetilde{v_{p,t}}\right)=\frac{\[{\left({\mathrm{gd}}_p^t\right)}^2+{\mathrm{gd}}_p^t\·{\mathrm{gc}}_p^t+{\left({\mathrm{gc}}_p^t\right)}^2\]-\[{\left({\mathrm{ga}}_p^t\right)}^2+{\mathrm{ga}}_p^t\·{\mathrm{gb}}_p^t+{\left({\mathrm{gb}}_p^t\right)}^2\]}{3({\mathrm{gd}}_p^t+{\mathrm{gc}}_p^t-{\mathrm{ga}}_p^t-{\mathrm{gb}}_p^t)}.$

Sub-step 4-2: the weight vectors composing m the index that power method is calculated by l kind is followed successively by ��₁, ��₂..., ��_l, wherein ��_k=[��_{K, 1}, ��_{K, 2}..., ��_{K, m}]^T,Integrated weight is ��=��₁����₁+��₂����₂+...+��_l����_l, wherein ��₁, ��₂.., ��_lFor integrated coefficient,Ideally, multiple weight ��₁, ��₂..., ��_lThe information provided for the preferred general objective of parts supplier should be impartial, therefore sets up many weight informations integrated model: $min\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{({\mathrm{\δ}}_k\·{\mathrm{\ψ}}_{k,t}\·v_{p,t}-{\mathrm{\δ}}_g\·{\mathrm{\ψ}}_{g,t}\·v_{p,t})}^2,$ S.t.k=1,2 ..., l, g=1,2 ..., l, k �� g.

Sub-step 4-3: assume integrated weight ��=��₁����₁+��₂����₂+...+��₁����_lIn except ��_kAnd ��_gThe integrated coefficient of other weights outer is zero, i.e. ��_k+��_g=1, by ��_g=1-��_kSubstitute in many weight informations integrated model and can solve ${\mathrm{\δ}}_k=\frac{\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{v_{p,t}}^2\·{\mathrm{\ψ}}_{g,t}\·({\mathrm{\ψ}}_{k,t}+{\mathrm{\ψ}}_{g,t})}{\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{v_{p,t}}^2\·{({\mathrm{\ψ}}_{k,t}+{\mathrm{\ψ}}_{g,t})}^2},{\mathrm{\δ}}_g=\frac{\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{v_{p,t}}^2\·{\mathrm{\ψ}}_{k,t}\·({\mathrm{\ψ}}_{k,t}+{\mathrm{\ψ}}_{g,t})}{\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{v_{p,t}}^2\·{({\mathrm{\ψ}}_{k,t}+{\mathrm{\ψ}}_{g,t})}^2},$ Make ��_k=1, then ${\mathrm{\δ}}_g=\frac{{\mathrm{\δ}}_g}{{\mathrm{\δ}}_k}.$

Sub-step 4-4: duplicon step 4-3, integrated coefficient �� can be obtained₁, ��₂..., ��_lJust reciprocal comparator matrix be�� is carried out consistency check, consistency check is obtained by method well known in the art, reference: Zhang Hanye, Guo Ruiling, bang river. based on the research [J] that the lithium cells shape used for electric vehicle of analytical hierarchy process is selected. machinery science and technology, 2012,31 (004): 523-527. If �� is by consistency check, solve the Maximum characteristic root �� of ��_maxCorresponding proper vector ��=[��₁, ��₂..., a_l]^TAnd make normalized and can obtain each integrated coefficient ��₁, ��₂..., ��_lValue, substitute into ��=��₁����₁+��₂����₂+...+��_l����_lAfter calculate to obtain integrated weight ��; Otherwise, according to method well known in the art, reference: the research [J] of the consistence of the satisfaction of judgment matrix in the .AHP of Wu Wen river. the practice of mathematics and understanding, 2010 (19): 181-189, �� is adjusted.

Step 5: by n family's parts supplier at each evaluation index Index₁, Index₂..., Index_mOn center of gravity form index value matrix V=(v_{P, t})_n��mWith index integration weight ��=[��₁, �� 2 ..., ��_m]^T, obtain Weighted Guidelines value matrix Z=(z_{P, t})_n��m, wherein z_{P, t}=��_t��v_{P, t}.According to Weighted Guidelines value matrix Z, according to method well known in the art, reference: Zhu Zhu, Zhang Lin, Ye Xiaowen, etc. based on the Land_use change Comprehensive Benefit Evaluation [J] of TOPSIS method. economic geography, 2012,32 (10): 139-144, n family's parts supplier is sorted, then arranges the first parts supplier and be optimal supplier.

Useful effect:

(1) all there is difference in the experience level of expert, know-how etc., treats its energy dynamics with a certain discrimination more scientific and reasonable;

(2) analytical hierarchy process uses accurately number to describe expert to the relative score of different index, cannot reflect the fuzzy property of true thinking properly. The fuzzy number with membership function more can reflect the true judge thinking of people. The information that thick energy collecting quantitative analysis is inaccurate, inconsistent, imperfect with process, this theory can more profoundly spy upon the true perception of expert. The analytical hierarchy process that fuzzy number substitutes accurately number, with slightly assembling conjunction, presses close to truth more when the several expert analysis mode result of combined treatment;

(3) evaluation index adopts single tax information temporary can be caused to lose, and the multiple simple average composing power have ignored its otherness, and the integrated process of many weights then can address these problems;

(4) method is easy, is easy to programming realization.

Four, accompanying drawing explanation

The preferred general objective exploded view of parts supplier of Tu1Shi bearing mnanufacture enterprise

Fig. 2 is the neural network structure figure calculating experts ability degree

Fig. 3 is the neural network training process figure calculating experts ability degree

Five, embodiment

Below in conjunction with specific embodiment, the present invention is described:

In the present embodiment, the outsourcing part of certain bearing mnanufacture enterprise is certain type roller, the supplier of this part have first, second, the third 3, these 3 roller suppliers need to be carried out multiple goal preferred, selected 1 is optimal supplier.

Step 1: according to the decomposition of the bearing mnanufacture enterprise preferred general objective of parts supplier, set up assessment indicator system, i.e. { Index₁, Index₂..., Index₅, it is corresponding in turn to quality index (sub-goal), deliverability index (sub-goal), economic situation index (sub-goal), partner services index (sub-goal) and environmental impact indicators (sub-goal);

Step 2: the energy dynamics being obtained each expert by the method for neural network training from historical sample:

There are 3 experts, have 6 samples, as shown in the table:

In sample 1,4 objects are O₁, O₂, O₃, O₄, it truly sorts as O₂, O₃, O₄, O₁; In sample 2,4 objects are O₅, O₆, O₇, O₈, it truly sorts as O₅, O₆, O₈, O₇; In sample 3,4 objects are O₉, O₁₀, O₁₁, O₁₂, it truly sorts as O₁₀, O₁₂, O₁₁, O₉. In sample 4,4 objects are O₁₃, O₁₄, O₁₅, O₁₆, it truly sorts as O₁₃, O₁₆, O₁₅, O₁₄. In sample 5,4 objects are O₁₇, O₁₈, O₁₉, O₂₀, it truly sorts as O₂₀, O₁₇, O₁₈, O₁₉. In sample 6,4 objects are O₂₁, O₂₂, O₂₃, O₂₄, it truly sorts as O₂₃, O₂₂, O₂₄, O₂₁��

After neural network training, calculate experts ability degree vector for ��=[0.42850.33460.2369]^T��

Step 3: introduce experts ability degree, by the comprehensive analytical hierarchy process of many rank Fuzzy Rough Sets, it is achieved 3 experts are to the combined treatment of 3 roller supplier appraisal result in single index:

Sub-step 3-1:3 expert to first, second, the third 3 roller suppliers at index Index₂On desired value rating matrix respectively

$\widetilde{H^{1,2}}=\left[\begin{array}{ccc}(1,1,1,1)& (7/8,13/9,9/5)& (4,5,6,7)\\ (5/9,9/13,1,8/7)& (1,1,1,1)& (2,5/2,3,7/2)\\ (1/7,1/6,1/5,1/4)& (2/7,1/3,2/5,1/2)& (1,1,1,1)\end{array}\right]$

For $\widetilde{H^{2,2}}=\left[\begin{array}{ccc}(1,1,1,1)& (5/8,11/9,8/5)& (4,5,31/5,7)\\ (5/8,9/11,1,8/5)& (1,1,1,1)& (2,5/2,3,10/3)\\ (1/7,5/31,1/5,1/4)& (3/10,1/3,2/5,1/2)& (1,1,1,1)\end{array}\right],$ All by consistency check, perform

$\widetilde{H^{3,2}}=\left[\begin{array}{ccc}(1,1,1,1)& (3/8,1,13/9,9/5)& (4,41/8,6,57/8)\\ (5/9,9/13,1,8/7)& (1,1,1,1)& (2,5/2,3,7/2)\\ (8/57,1/6,8/41,1/4)& (2/7,1/3,2/5,1/2)& (1,1,1,1)\end{array}\right]$

Sub-step 3-2;

Sub-step 3-2: willDecompose respectively, wherein $A^{1,2}=\left[\begin{array}{ccc}1& 7/8& 4\\ 5/9& 1& 2\\ 1/7& 2/7& 1\end{array}\right],$ Other is analogized with this.

Sub-step 3-3: according to A^1,2, A^2,2, A^3,2, build coarse group decision matrix

$A^2=\left[\begin{array}{ccc}1& \{7/8,5/8,3/8\}& \{4,4,4\}\\ \{5/9,5/8,5/9\}& 1& \{2,2,2\}\\ \{1/7,1/7,8/57\}& \{2/7,3/10,2/7,\}& 1\end{array}\right].$ Wherein, $A_{1,2}^2=\{7/8,5/8,3/8\}$ In division " 5/8 ",

On it, approximate set is that { 7/8,5/8}, lower aprons collection is that { 5/8,3/8}, its coarse number is L(5/8)=(5/8+3/8)/2=0.5000, therefore RN (5/8)=[0.5000,0.7500], with managing RN (7/8)=[0.6250,0.8750], RN (3/8)=[0.3750,0.7500]. So having $RN\left(A_{1,2}^2\right)=\{\[0.6250,0.8750\],\[0.5000,0.7500\],\[0.3750,0.7500\]\},$ Introduce experts ability degree ��=[0.42850.33460.2369]^T, can obtainAverage roughness interval be Try to achieve A successively²In the average roughness of other elements interval.

Sub-step 3-4: construct coarse Judgement Matrix ${\mathrm{EA}}^2=\left[\begin{array}{ccc}\[1,1\]& \[0.1747,0.2580\]& \[1.3212,1.3212\]\\ \[0.1876,0.1987\]& \[1,1\]& \[0.7123,0.7123\]\\ \[0.0526,0.0501\]& \[0.0989,0.1020\]& \[1,1\]\end{array}\right].$ By EA²It is decomposed into coarse lower boundary matrix EAI²With coarse upper boundary matrix EAS², ask the proper vector corresponding to maximum eigenwert to be VAI respectively²=[0.7828,0.5891,0.2004]^T, VAS²=[0.8053,0.5615,0.1906]^T, finally try to achieve GA²={ 0.7941,0.5753,0.1955}.

Sub-step 3-5: with reason, duplicon step 3-3,3-4, can obtain GB²={ 0.8352,0.6087,0.2490}, GC²={ 0.8796,0.7038,0.2875}, GD²=0.9233,0.7865,0.3246}, thus 3 roller suppliers are at index Index₂On fuzzy value be followed successively by (0.7941,0.8352,0.8796,0.9233), (0.5753,0.6087,0.7038,0.7865), (0.1955,0.2490,0.2875,0.3246).

Sub-step 3-6: with reason, duplicon step 3-1 to 3-5, can try to achieve the fuzzy value of 3 roller suppliers in other index, thus index value matrix in 5 indexs of 3 roller suppliers isFor roller supplier p (p=1,2,3) index t (t=1,2 ..., 5) on desired value, represent by trapezoidal fuzzy several form.

Step 4: set up many weight informations integrated model, solves the integrated weight of evaluation index:

Sub-step 4-1: the roller supplier index value matrix will tried to achieve in step 3 by center of gravity methodIt is converted into center of gravity form V=(v_{P, t})_3��5��

Sub-step 4-2: according to tax power method known in field: entropy assessment, variation coefficient method, CRITIC method, reference [1] Ni nine group, Li Ping, Wei Chaofu, Deng. region land exploitation-renovation Potential Evaluation [J] of power is composed based on AHP and entropy assessment. Transactions of the Chinese Society of Agricultural Engineering, 2009 (5): 202-209.[2] Sun Kai, Ju Xiaofeng, Li Yuhua. the enterprise incubator based on variation coefficient method runs performance appraisal [J]. Harbin University of Science and Technology's journal, 2007, 12 (3): 165-167.[3] Xu Ping .CRITIC method quality of medical work evaluate in application [J]. value engineering, 2011, 30 (1): 200-201.

The weight vectors composing 5 indexs that power method is calculated by these 3 kinds is followed successively by ��₁=[0.36070.30120.11070.10460.1228]^T, ��₂=[0.30970.10070.10470.18380.3011]^T, ��₃=[0.33990.11480.31140.12090.1130]^T. Integrated weight is ��=��₁����₁+��₂����₂+��₃����₃. Set up many weight informations integrated model:S.t.k=1,2,3, g=1,2,3, k �� g.

Sub-step 4-3: make ��₃=0, then ��₁+��₂=1. By ��₂=1-��₁Substitute into and many weight informations integrated model can solve ��₁=0.3188, ��₂=0.6812, make ��₁=1, then

Sub-step 4-4: duplicon step 4-3, integrated coefficient �� can be obtained₁, ��₂, ��₃Just reciprocal comparator matrix be $\mathrm{\Θ}=\left[\begin{array}{ccc}1& 2.1368& 3.8677\\ 0.4680& 1& 3.0120\\ 0.2586& 0.3320& 1\end{array}\right].$ �� passes through consistency check, its Maximum characteristic root ��_max=3.0290, corresponding proper vector ��=[0.85900.47640.1874]^T, make normalized and can obtain ��₁=0.5641, ��₂=0.3128, ��₃=0.1231, thus integrated weight ��=[0.34220.21550.13350.13140.1774]^T��

Step 5: by 3 roller suppliers at each evaluation index Index₁, Index₂..., Index₅On center of gravity form index value matrix V=(v_{P, t})_3��5With index integration weight ��, obtain Weighted Guidelines value matrix $Z={\left(z_{p,t}\right)}_{3\×5}=\left[\begin{array}{ccccc}0.7798& 0.2378& 0.5876& 0.9081& 0.4239\\ 0.4897& 0.3490& 0.1230& 0.5782& 0.1876\\ 0.6065& 0.1008& 0.2139& 0.6329& 0.2305\end{array}\right],$ Wherein z_{P, t}=��_t��v_{P, t}.According to Weighted Guidelines value matrix Z, according to method well known in the art, reference: Zhu Zhu, Zhang Lin, Ye Xiaowen, etc. based on the Land_use change Comprehensive Benefit Evaluation [J] of TOPSIS method. economic geography, 2012,32 (10): the ranking results of roller supplier of 139-144,3 family is first > third > second, then first is the optimum roller supplier of this bearing mnanufacture enterprise.

Claims (1)

1. the parts supplier multiple goal preferred method towards bearing mnanufacture enterprise, it is characterised in that: comprise the steps:

Step 1: preferred for the parts supplier of bearing mnanufacture enterprise general objective is decomposed into five sub-goals: quality sub-goal, deliverability sub-goal, economic situation sub-goal, partner services sub-goal and environmental influence sub-goal, it is corresponding in turn to five evaluation indexes, uses Index successively₁, Index₂..., Index_mRepresent, m=5;

Step 2: the energy dynamics being obtained each expert by the method for neural network training from historical sample:

S (s > 1) individual expert is to r object O₁, O₂..., O_rSort, the total individual sample of N '. Experts ability degree vector is $\mathrm{\ρ}={\[{\mathrm{\ρ}}_1,{\mathrm{\ρ}}_2,...,{\mathrm{\ρ}}_s\]}^T(\underset{i=1}{\overset{s}{\Σ}}{\mathrm{\ρ}}_i=1,0\≤{\mathrm{\ρ}}_i\≤1).$

Sample N (N=1,2 ..., N ') in, expert i (i=1,2 ..., ranking results s) is:WithRepresentIn be in O_jObject number afterwards is (containing O_jSelf), also represent O_j?In score. In sample N, useTo O₁, O₂..., O_rFinally sorting, s expert is to O₁, O₂..., O_rFinal ranking results represent and be:

O₁, O₂..., O_rTrue ranking results be:

Set up the neural network structure calculating experts ability degree:

At input layer, neuron number is s �� r, and input matrix is Equal

At hidden layer, neuron number is r (equal with object number). For neurone j (i.e. object j, j=1,2 ..., r), it is input asOutput isConnection weight matrix between input layer and hidden layer node can be divided into s sub-matrix-block, wherein submatrix block i (i=1,2 ..., s) be $\underset{i}{{\mathrm{\Γ}}^{\left(N\right)}}=\left[\begin{array}{cccc}\underset{i}{{\mathrm{\γ}}_{11}^{\left(N\right)}}& 0& ...& 0\\ 0& \underset{i}{{\mathrm{\γ}}_{22}^{\left(N\right)}}& ...& 0\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ \\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ 0& 0& ...& \underset{i}{{\mathrm{\γ}}_{rr}^{\left(N\right)}}\end{array}\right];$

At output layer, only 1 neurone, hidden layer node and the connection weight being output between node are vectorial is (1,1 ..., 1)^T. The deduction process of the neural network calculating experts ability degree is:

For hidden layer node j (i.e. object j, j=1,2 ..., r), it is input asOutput is ${\mathrm{dout}}_j^{\left(N\right)}=f({\mathrm{yin}}_j^{\left(N\right)}+{\mathrm{\θ}}_j),$ Action function is $f\left(y\right)=\frac{1}{1+e^{-y}},$

For output layer node, it is input asOutput is ${\mathrm{bout}}^{\left(N\right)}=({\mathrm{bout}}_1^{\left(N\right)},{\mathrm{bout}}_2^{\left(N\right)},...,{\mathrm{bout}}_r^{\left(N\right)}),{\mathrm{bout}}_j^{\left(N\right)}$ (j=1,2 ..., r) it is object O_j?In score,BySort and obtain.

For error function, the error of sample N is It is object O_j?In score,It is object O_j?In score.

The experts ability degree obtained (namely connection weight) between input layer and hidden layer node is normalized: for submatrix block after completing by the neural network training calculating experts ability degreeCalculate the mean value of its diagonal lines elementCalculate successivelyAfter be normalized, so the energy dynamics of expert i isFinally calculate experts ability degree vector for ��=[��₁, ��₂..., ��_s]��

Step 3: introduce experts ability degree, by the comprehensive analytical hierarchy process of many rank Fuzzy Rough Sets, it is achieved several expert is to the combined treatment of many parts suppliers appraisal result in single index:

Having n family's parts supplier, s expert participates in scoring, and evaluation indice is Index={Index₁, Index₂..., Index_m, m=5.

Sub-step 3-1: at index Index_tOn �� Index, the rating matrix of n family's parts supplier, on the basis grasping history supply situation and investigation, is followed successively by by s expertWherein expert i (i=1,2 ..., s) to the rating matrix of n family's parts supplier be: $\widetilde{H^{i,t}}=\left[\begin{array}{cccc}\left(1,1,1,1\right)& \widetilde{h_{1,2}^{i,t}}& ...& \widetilde{h_{1,n}^{i,t}}\\ \widetilde{h_{2,1}^{i,t}}& \left(1,1,1,1\right)& ...& \widetilde{h_{2,n}^{i,t}}\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ \\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ \widetilde{h_{n,1}^{i,t}}& \widetilde{h_{n,2}^{i,t}}& ...& \left(1,1,1,1\right)\end{array}\right],$ $\widetilde{h_{p,q}^{i,t}}=(a_{p,q}^{i,t},b_{p,q}^{i,t},c_{p,q}^{i,t},d_{p,q}^{i,t})$ Represent expert i provide at index Index_tUpper parts supplier q (q=1,2 ..., n) relative to parts supplier p (p=1,2 ..., n) trapezoidal fuzzy number scoring, p �� q.S rating matrix is carried out consistency check respectively, consistency check method is obtained by method well known in the art, reference: Luo Zhimeng, Zhou Jianzhong, Yang person of outstanding talent, Deng. based on the Virtual Research Centers partner selection [J] of fuzzy AHP. Central China University of Science and Technology's journal: natural science edition, 2008,36 (12): 100-103. If s rating matrix is all by consistency check, perform sub-step 3-2; Otherwise go back to sub-step 3-1, defective rating matrix is marked again by the expert of correspondence.

Sub-step 3-2: decomposeForWherein

$A^{i,t}=\left[\begin{array}{cccc}1& a_{1,2}^{i,t}& ...& a_{1,n}^{i,t}\\ a_{2,1}^{i,t}& 1& ...& a_{2,n}^{i,t}\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ \\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ a_{n,1}^{i,t}& a_{n,2}^{i,t}& ...& 1\end{array}\right],$

Other is analogized with this.

Sub-step 3-3: according to A^{1, t}, A^{2, t}..., A^{S, t}Build coarse group decision matrix: $A^t=\left[\begin{array}{cccc}1& A_{1,2}^t& ...& A_{1,n}^t\\ A_{2,1}^t& 1& ...& A_{2,n}^t\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ \\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ A_{n,1}^t& A_{n,2}^t& ...& 1\end{array}\right],$ Wherein,

$A_{p,q}^t=\{a_{p,q}^{1,t},a_{p,q}^{2,t},...,a_{p,q}^{s,t}\},$

P �� q,

P=1,2 ..., n,

Q=1,2 ..., n

So,The coarse number corresponding by s expert analysis mode data comprised is: $RN\left(a_{p,q}^{i,t}\right)=\[a_{p,q}^{i-,t},a_{p,q}^{i+,t}\],i=1,2,...,s.$ So, $RN\left(A_{p,q}^t\right)=\{\[a_{p,q}^{1-,t},a_{p,q}^{1+,t}\],\[a_{p,q}^{2-,t},a_{p,q}^{2+,t}\],...,\[a_{p,q}^{q-,t},a_{p,q}^{q+,t}\]\}.$ Introducing experts ability degree, can try to achieve average roughness interval is $ARN\left(A_{p,q}^t\right)=\[a_{p,q}^{-,t},a_{p,q}^{+,t}\]=\[\frac{\underset{i=1}{\overset{s}{\Σ}}{\mathrm{\ρ}}_i\·a_{p,q}^{i-,t}}{s},\frac{\underset{i=1}{\overset{s}{\Σ}}{\mathrm{\ρ}}_i\·a_{p,q}^{i+,t}}{s}\].$ Try to achieve A successively^tIn the average roughness of other elements interval.

Sub-step 3-4: construct coarse Judgement Matrix ${\mathrm{EA}}^t=\left[\begin{array}{cccc}\[1,1\]& \[a_{1,2}^{-,t},a_{1,2}^{+,t}\]& ...& \[a_{1,n}^{-,t},a_{1,n}^{+,t}\]\\ \[a_{2,1}^{-,t},a_{2,1}^{+,t}\]& \[1,1\]& ...& \[a_{2,n}^{-,t},a_{2,n}^{+,t}\]\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ \\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ \[a_{n,1}^{-,t},a_{n,1}^{+,t}\]& \[a_{n,2}^{-,t},a_{n,2}^{+,t}\]& ...& \[1,1\]\end{array}\right].$ By EA^tIt is decomposed into coarse lower boundary matrix EAI^tWith coarse upper boundary matrix EAS^t: ${\mathrm{EAI}}^t=\left[\begin{array}{cccc}1& a_{1,2}^{-,t}& ...& a_{1,n}^{-,t}\\ a_{2,1}^{-,t}& 1& ...& a_{2,n}^{-,t}\\ \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}.\\ .\\ .\end{array}& \begin{array}{c}\\ \\ \end{array}& \begin{array}{c}.\\ .\\ .\end{array}\\ a_{n,1}^{-,t}& a_{n,2}^{-,t}& ...& 1\end{array}\right],$ ${\mathrm{EAS}}^t=\left[\begin{array}{cccc}1& a_{1,2}^{+,t}& \mathrm{...}& a_{1,n}^{+,t}\\ a_{2,1}^{+,t}& 1& \mathrm{...}& a_{2,n}^{+,t}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ a_{n,1}^{+,t}& a_{n,2}^{+,t}& \mathrm{...}& 1\end{array}\right],$ The proper vector corresponding to maximum eigenwert is asked to be respectively ${\mathrm{VAI}}^t={\[{\mathrm{vai}}_1,{\mathrm{vai}}_2,...,{\mathrm{vai}}_n\]}^T,$ ${\mathrm{VAS}}^t={\[{\mathrm{vas}}_1,{\mathrm{vas}}_2,...,{\mathrm{vas}}_n\]}^T,$ After average it is ${\mathrm{ga}}_p^t=\frac{\left|{\mathrm{vai}}_p\right|+\left|{\mathrm{vas}}_p\right|}{2},p=1,2,...,n,$ Thus have ${\mathrm{GA}}^t=\{{\mathrm{ga}}_1^t,{\mathrm{ga}}_2^t,...,{\mathrm{ga}}_n^t\}.$

Sub-step 3-5: with reason, duplicon step 3-3,3-4, can obtain ${\mathrm{GB}}^t=\{{\mathrm{gb}}_1^t,{\mathrm{gb}}_2^t,...,{\mathrm{gb}}_n^t\},$ ${\mathrm{GC}}^t=\{{\mathrm{gc}}_1^t,{\mathrm{gc}}_2^t,...,{\mathrm{gc}}_n^t\},$ ${\mathrm{GD}}^t=\{{\mathrm{gd}}_1^t,{\mathrm{gd}}_2^t,...,{\mathrm{gd}}_n^t\},$ Thus n family's parts supplier is at index Index_t(t=1,2 ..., the fuzzy value on m) is followed successively by

Sub-step 3-6: with reason, duplicon step 3-1 to 3-5, can try to achieve the fuzzy value of n family's parts supplier in other index, thus the index value matrix of n family's parts supplier in m index is For parts supplier p (p=1,2 ..., n) index t (t=1,2 ..., the desired value on m), represents by trapezoidal fuzzy several form.

Step 4: set up many weight informations integrated model, solves the integrated weight of evaluation index:

Sub-step 4-1: the parts supplier index value matrix will tried to achieve in step 3 by center of gravity methodIt is converted into center of gravity form V=(v_{P, t})_n��m, $v_{p,t}=Gra\left(\widetilde{v_{p,t}}\right)=\frac{\[{\left({\mathrm{gd}}_p^t\right)}^2+{\mathrm{gd}}_p^t\·{\mathrm{gc}}_p^t+{\left({\mathrm{gc}}_p^t\right)}^2\]-\[{\left({\mathrm{ga}}_p^t\right)}^2+{\mathrm{ga}}_p^t\·{\mathrm{gb}}_p^t+{\left({\mathrm{gb}}_p^t\right)}^2\]}{3({\mathrm{gd}}_p^t+{\mathrm{gc}}_p^t-{\mathrm{ga}}_p^t-{\mathrm{gb}}_p^t)}.$

Sub-step 4-2: the weight vectors composing m the index that power method is calculated by l kind is followed successively by ��₁, ��₂..., ��_l, wherein ��_k=[��_{K, 1}, ��_{K, 2}..., ��_{K, m}]^T, k=1,2 ..., l,Integrated weight is ��=��₁����₁+��₂����₂+...+��_l����_l, wherein ��₁, ��₂..., ��_lFor integrated coefficient, ��₁, ��₂..., ��_l>=0,Ideally, multiple weight ��₁, ��₂..., ��_lThe information provided for the preferred general objective of parts supplier should be impartial, therefore sets up many weight informations integrated model: $min\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{({\mathrm{\δ}}_k\·{\mathrm{\ψ}}_{k,t}\·v_{p,t}-{\mathrm{\δ}}_g\·{\mathrm{\ψ}}_{g,t}\·v_{p,t})}^2,$ S.t.k=1,2 ..., l, g=1,2 ..., l, k �� g.

Sub-step 4-3: assume integrated weight ��=��₁����₁+��₂����₂+...+��_l����_lIn except ��_kAnd ��_gThe integrated coefficient of other weights outer is zero, i.e. ��_k+��_g=1, by ��_g=1-��_kSubstitute in many weight informations integrated model and can solve ${\mathrm{\δ}}_k=\frac{\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{v_{p,t}}^2\·{\mathrm{\ψ}}_{g,t}\·({\mathrm{\ψ}}_{k,t}+{\mathrm{\ψ}}_{g,t})}{\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{v_{p,t}}^2\·{({\mathrm{\ψ}}_{k,t}+{\mathrm{\ψ}}_{g,t})}^2},$ ${\mathrm{\δ}}_g=\frac{\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{v_{p,t}}^2\·{\mathrm{\ψ}}_{k,t}\·({\mathrm{\ψ}}_{k,t}+{\mathrm{\ψ}}_{g,t})}{\underset{p=1}{\overset{n}{\Σ}}\underset{t=1}{\overset{m}{\Σ}}{v_{p,t}}^2\·{({\mathrm{\ψ}}_{k,t}+{\mathrm{\ψ}}_{g,t})}^2},$ Make ��_k=1, then ${\mathrm{\δ}}_g=\frac{{\mathrm{\δ}}_g}{{\mathrm{\δ}}_k}.$

Sub-step 4-4: duplicon step 4-3, integrated coefficient �� can be obtained₁, ��₂..., ��_lJust reciprocal comparator matrix be�� is carried out consistency check, consistency check is obtained by method well known in the art, reference: Zhang Hanye, Guo Ruiling, bang river. based on the research [J] that the lithium cells shape used for electric vehicle of analytical hierarchy process is selected. machinery science and technology, 2012,31 (004): 523-527. If �� is by consistency check, solve the Maximum characteristic root �� of ��_maxCorresponding proper vector ��=[��₁, ��₂..., ��_l]^TAnd make normalized and can obtain each integrated coefficient ��₁, ��₂..., ��_lValue, substitute into ��=��₁����₁+��₂����₂+...+��_l����_lAfter calculate to obtain integrated weight ��;Otherwise, according to method well known in the art, reference: the research [J] of the consistence of the satisfaction of judgment matrix in the .AHP of Wu Wen river. the practice of mathematics and understanding, 2010 (19): 181-189, �� is adjusted.

Step 5: by n family's parts supplier at each evaluation index Index₁, Index₂..., Index_mOn center of gravity form index value matrix V=(v_{P, t})_n��mWith index integration weight ��=[��₁, ��₂..., ��_m]^T, obtain Weighted Guidelines value matrix Z=(z_{P, t})_n��m, wherein z_{P, t}=��_t��v_{P, t}. According to Weighted Guidelines value matrix Z, according to method well known in the art, reference: Zhu Zhu, Zhang Lin, Ye Xiaowen, etc. based on the Land_use change Comprehensive Benefit Evaluation [J] of TOPSIS method. economic geography, 2012,32 (10): 139-144, n family's parts supplier is sorted, then arranges the first parts supplier and be optimal supplier.