实例推理中遗传训练算法用于机械失效模式识别的研究_英文_

实例推理中遗传训练算法用于机械失效模式识别的研究_英文_ Research of Genetic Tra in ing Algorithm f or Identif ying Mechan ical Fa il ure Modes within the Fra me work of Ca se2Ba sed Rea son ing XU Yuan2ming , ZHAN G Yang , CH EN Li2na ( S chool of A e ron a u t ical S cience a n d En gi nee ri n g Tech nology , B ei j i n g U ni ve rsi t y of )100 08 3 , Chi n a A e ron a u t ics a n d A st ron a u t ics , B ei j i n g ( ) ( ) Abstract : The co mbinatio n of case2based reaso ning CB Rand genetic algorit hm GAis co nsidered in t he p roblem of failure mode identificatio n in aero nautical co mpo nent failure analysis. Several imple2 mentatio n issues such as matching at t ributes selectio n , similarity measure calculatio n , weight s learning and t raining evaluatio n policies are caref ully st udied. The testing applicatio ns illust rate t hat an accuracy of 74167 % can be achieved wit h 75 balanced2dist ributed failure cases covering 3 failure modes , and t hat t he resulting learning weight vector can be well applied to t he ot her 2 failure modes , achieving 7313 % of recognitio n accuracy. It is also p roved t hat it s pop ularizing capabilit y is good to t he recognitio n of even more mixed failure modes. Key words : failure mode identificatio n ; case2based reaso ning ; genetic algorit hm ; learning t rain 实例 (推理中遗传训练算法用于机械失效模式识别的研究 . 徐元铭 ,张洋 ,陈丽娜. 中国航空学报 英 ) () 文版, 2005 , 18 2: 122 - 129 . 摘 要 :采用实例推理和遗传算法相结合的 ,研究了航空机械零部件失效模式识别的问题 。 对用于识别的失效属性的选择 、检索相似度计算 、训练用遗传算法的适应度函数设计以及训练策 略的影响进行了较为详细的描述 。应用测试表明 ,对包含分布均衡的 3 种模式的情况取得了高于 ( ) 74167 %的识别率 ,所获得的最佳权值向量对另外 2 种模式具有很好的识别精度 大于 7313 %,对 混合多模式情况也具有较好的推广能力 。验证了该方法对航空零部件失效模式的识别是可行的 。 关键词 :失效模式识别 ; 基于实例的推理 ; 遗传算法 ; 学习训练 () 文章编号 : 100029361 20050220122208 中图分类号 : TP391 . 5 ; TP311 . 11文献标识码 :A main , f ailure mo de identificatio n is usually co mplex and time co nsuming. It involves , in many cases , a 1 Int ro ductio n 1 gro up of expert s fo r making synt hetic decisio ns. Failure mo des no r mally refer to t ho se fo r ms of The applicatio n of logic based Artificial Intelligence () A Itechniques gives a p ro mising met ho d fo r aid2 exhibiting a f ailure of a mechanical co mpo nent in eit her macro scopic o r micro scopic sense , o r t he ing human’s f ailure analysis task . Several notice2 able research wo r k have been do ne in t his research classificatio n of t he co mpo nent ’s f ailure mecha2 2 nisms acco rding to t he p hysical , chemical o r ot her area , but still very limited : Mayer used an ex2 p rocesses w hich have led to a f ailure . Identifying pert system app roach to identif y t he basic boiler 3 f ailure mo de of a f ailed co mpo nent is t he mo st im2 t ube f ailure mechanisms ; Ko mai , et al investi2po rtant step in t he entire task of f ailure analysis , gated image p rocess and pat ter n recognitio n fo r i2 since it can give t he effective task2o riented guide2 dentif ying six different f ract ure surf ace mo rp holo2 4 ,5 lines fo r subsequent analysis decisio ns o n deter min2 integrated database wit h ex2 gies ; Liao , et al ing f ailure causes and reco mmending p recautio n ac2 pert systems , as well as case based reaso ning fo r tio ns. f ailure mechanism recognitio n in pet ro2chemical in2 6 ,7 dust ry applicatio n ; and Xu , et al , gave an rule In aero nautical equip ment f ailure analysis do2 Received date : 2003211217 ; Revisio n received date : 2005201225 ( )Fo undatio n item : 973 Fo undatio n Program of China G1999065010 based uncertaint y expert system fo r aero nautical e2 quip ment f ailure analysis , and investigated case ( ) based reaso ning CB Rin t his do main . This paper describes f urt her t he st udy result s ( ) of Genetic Algo rit hm GA based t raining wit h CB R p ro blem2solving paradigms fo r f ailure mo de i2 dentificatio n of aero nautical equip ment . 2 Failure Mo de Identificatio n U sing CB R CB R is a met ho dology w hich stems f ro m hu2 man’s reaso ning behavio r by recalling o r resem2 bling past similar sit uatio ns. It s basic idea is t reat2 ing first t he p ro blem to be solved as “ TargetFig. 1 The flow chart of CB R system Case”, and a gro up of already solved o r old p ro b2 at t ribute o r an at t ribute vecto r is deter mined by lems as“Base Cases”; and p roceeding wit h t he as2 weight lear ning p rocess t hro ugh GA based t raining sessment of a similarit y bet ween t he target case and ( ) see Sectio n 5upo n a histo rical f ailure case base . t he base cases by designing p roper quantitative Finally , t he system o utp ut s t he mo st similar f ailure “weight s of judgment ”schemes. Based o n t he de2 mo de w hich can be validated by user to decide (gree of similarit y , t he solutio n pat ter n o r solutio n w het her it is to be sto red into t he case base o r not . ) it self of an old p ro blem can be p rocessed o r adap t2 (ed to infer t he new solutio n pat ter n o r even solu2 ) tio nfo r t he target p ro blem. The advantages of ap2 3 Failure At t ribute Selectio n and Gro uping plying CB R in f ailure analysis can be summarized as It is co mmo n sense in aero nautical equip ment ( ) 1 It does not require t he explicit do main f ailure analysis t hat t he aspect s of general visual o r knowledge info r matio n , o nly a collectio n of f ailure surf ace co nditio ns of a f ailed co mpo nent , it s f rac2 cases needed to be sto red , t hus avoiding t he bot tle2 t ure/ crack f ace feat ures , cro ss2sectio n/ subsurf ace neck of knowledge elicitatio n . feat ures , and elect ro n f ractograp hic feat ures , etc . () 2The mat ure and advanced database tech2 have to be examined befo re t he detailed analyses of nology can be used to manage t hese f ailure cases. ( material co mpo sitio n and p hysical p roperties e . () 3The identificatio n capabilit y can be incre2 ) g . , hardness , brit tle , ductility , etc . Failure mentally imp roved by lear ning t hro ugh new cases. mo de identificatio n p hase requires t he co nsideratio n Fig11 show s t he flow chart of CB R system fo r of t ho se feat ures w hich are universal and clo sely re2 f ailure mo de identificatio n . The w hole system can lated to f ailure occurrence and evolutio n of t he be characterized by t wo stages : case ret rieval and f ailed co mpo nent ; w hereas t he basic info r matio n weight lear ning. Case ret rieval start s by inp ut ting abo ut t he co mpo nent ’s no r mal states , wo r king sufficient info r matio n abo ut target case to be co n2 co nditio ns , and exter nal loads , alo ng wit h t he sidered , and uses t he at t ribute selectio n criteria analyses of material co mpo sitio n and mechanical ( ) 2 see Sectio n 3 belowto ext ract t he mo st signifi (p roperties and so o n , are not o bligato ry p ractices cant relevant f ailure at t ributes associated wit h t he demo nst rate t hat t hey are in f act mo re usef ul fo r identificatio n p rocess. The match bet ween t he se2 ) f ailure cause deter minatio n. Therefo re , t he classi2 lected at t ributes of t he target case and t ho se in case ficatio n of f ailure at t ributes selected fo r f ailure base is co nducted by a similarit y measure using a ( ( ) mo de identificatio n af ter co nsulting human ex2 weighed K Nearest Neighbo rs KN N technique ( ) ) see Sectio n 4. The ro bust op timal weight of an pert s and f ailure analysis handboo kscan be show n in Fig12 . similarit y measures uses Boolean logic : 1 - exis2 tence of t he at t ribute ; - 1 - no existence of t he at2 t ribute ; 0 - unknow n of t he at t ribute . L ear ning by GA Based Training Weight 5 The p urpo se of using Genetic Algo rit hm in CB R is to acquire t he ro bust op timal weight s of f ailure at t ributes upo n an o riginal f ailure case base . This is an impo rtant part of f ailure mo de identifica2 tio n since it has a great influence o n identificatio n perfo r mance. Because t he total numbers of at2 Fig. 2 The hierarchy of failure at t ributes t ribute element s and at t ribute vecto rs are 343 and ( ) 14 see Fig12respectively , it wo uld be unrealistic Three basic at t ribute gro up s , namely , general to assign a weight fo r each at t ribute element o r ( ) surf ace co nditio ns 57 , f ract ure/ crack feat ures each at t ribute vecto r due to t he co mp utatio n bur2 ( ) () 58and f ract ure mo rp hologies 228, are defined den . Therefo re , it is decided to set a same weight in t his research . The number at tached here rep re2 fo r at t ribute element s o r vecto rs in t he same gro up . sent s t he total number of element s t hat each gro up In t his way , The total number of weight s used fo r can take . Each gro up can be f urt her divided into t raining is 3 , i . e . weight vecto r W = [ w , w , 1 2 o ne o r mo re at t ribute vecto rs as show n in Fig. 2 . It w ] . The weight assignment fo r case ret rieval is sho uld be noted t hat different ways of gro uping 3 perfo r med based o n t he searching and lear ning ca2 f ailure at t ributes are po ssible . It really depends o n pabilites of GA . GA is t he op timizatio n algo rit hm t heir effect s o n t he GA based weight t raining co m2 based o n t he nat ural evolutio n co ncep t co ming f ro m p utatio n co st and identificatio n accuracy. Darwin’s t heo ry of evolutio n . The nat ural selectio n eighed KNN Ret rieval Met ho d A W4 increases t he surviving capabilities of a pop ulatio n over t he generatio ns. The genetic info r matio n of The general fo r m of similarity measure f unc2 each individual is sto red in a chro mo so mal st ring tio n is as follow s : n and t he goo dness of individual is measured by ( ) )( dist x , y W ×1 - i i i ? defining a fit ness f unctio n based t he st ring. Only i = 1 ( ) S IM X , Y = n t he individuals wit h bet ter characteristics survive W i ?i = 1 during t he evolutio nary p rocess so t hat t he fit ness ) ( w here S IM X , Y is t he similarit y bet ween case f unctio n is maximized. The detailed p rocedures of X and Y , W is t he weight of at t ribute element i , i GA is given in Ref . 8 . ( ) n is t he number of at t ributes ; dist x , y is t he i i 5 . 1 The f itness f unct ion no r malized distance of t he i t h at t ribute bet ween In CB R do main , t he classificatio n accuracy t wo cases , and takes t he fo r m as follow s : rate of t raining case set fo r a particular weight vec2 ( ) ( ) ( ) dist x , y = | x - y | / | max i - min i | i i i i to r is adop ted as t he fit ness f unctio n of GA lear ning 9 ,10 w here x , y are i t h at t ribute values of case X and p rocess. However , to avoid p re2mat ure co n2 i i ( ) ( ) Y respectively , max i and min i denote t he vergence and keep high rate of accuracy , a penalt y upper and lower values of t he i t h at t ribute respec2 f unctio n is caref ully defined based o n massive t rial tively. If x o r y is unknow n , w hich means just analysis. The mat hematical fo r m of t he fit ness and i i o ne of t he cases o r bot h has missing at t ribute val2 penalt y f unctio ns designed are exp ressed as 5 m ) ( ues , it usually set s dist x , y = 015 . i i ( ) F= [ S IM T , S + P]wit h l i k li ?The assignment of t he at t ribute values fo r i = 1 p ( ) t hat co uld po ssess a high pop ularized capabilit y. = 0 . 5 S IM T , S if t he f ailure mo des of lii k () 2L eave One Out policy T and S are same i k This policy takes o nly o ne case o ut of a select2 o r ed test case set fo r testing , and matches it wit h t he ( ) = - 1 S IM T , S if t he f ailure mo des ofp i k li mo st similar case in t he rest of t he case set , and T and S are different i k judges t he success of match by o utco mes of t he t wo w here Fis t he fit ness f unctio n of l t h weight vec2 l cases. And af ter t his , ret ur n t he case into t he test to r , m is t he number of test cases , Pis t he penal2 li set and take next o ne . The p rocess repeat s until t he t y f unctio n , T is t he i t h test case , and S is a ref2 i k p re2specified test cycles are satisfied o r all cases in erence case w hich is mo st similar to T , i . e . , i t he test set are tested. The o bvio us feat ure of t his ( ) ( ) S IM T , S = { maxS IM T , S | j = 1 , 2 , . . . , n} i k i j policy is t he every case in a test set serves as eit her n is t he number w here S is t he j t h reference case , j a test case o r a reference case . Therefo re , t he of reference cases. t raining classificatio n capabilit y is equally dist ribut2 5 . 2 Tra in ing pol ic ies ed. However , w hen t he test set is large , it will The t raining of weight s fo r f ailure at t ributes suffer t he p ro blem of a co mp utatio n burden . De2 has to solve t he p ro blem of selectio n of a test case spite t his , L eave One Out policy can still be use2 set and a reference case set . Two kinds of t raining f ul fo r validatio n of effectiveness of op timal weight s policies w hich are no r mally used in CB R are as fol2 ( ) acquired by Policy 1 w hen it wo r ks o n a fixed low s : volume of a t raining set available . () 1Test Reference Set policy In t his st udy , a mixed t raining policy is f ur2 This policy requires t he divisio n of t he w hole t her p ropo sed , w hich suggest s t hat w hen t he test f ailure case set w hich participates in t raining into a ( ratio is relatively small e . g . bet ween 0125 , 2 test set and a reference set . A test ratio can be de) () 015Policy 1is f avo red ; w hereas w hen t he test fined here , w hich is t he p ropo rtio n of t he number ( ) ratio lies bet ween 0155,110 , Policy 2is advo2 of cases in a test set to t he number of cases in t he cated to wo r k o n t he test set . The effect of t his w hole t raining set . Given a specified test ratio , an policy will be demo nst rated in Sectio n 6 below . op timal weight vecto r is searched by GA operatio n No mat ter w hich t raining policy is used to ac2 acco rding to t he co mp utatio n of a fit ness f unctio n quire t he op timal weight vecto rs , t he ro bust nesses described in Sectio n 511 , by recursively taking of t hese weight vecto rs have to be validated based each f ailure case in t he test set and matching it o n t he co mpariso n of t heir t raining classificatio n ac2 wit h t he mo st similar case in t he reference set . curacies wit h validated classificatio n accuracies as ( During t his p rocess , t he o utco mes i . e . f ailure described in t he next subsectio n . Therefo re , cer2 ) mo desof t he t wo matched cases are co mpared and tain validatio n criteria are p ropo sed in t his st udy. t he success of identificatio n can be judged. Finally , 5 . 3 Val idat ion of tra in ing eff ects t he percentage of t he successf ul matched cases over As stated befo re , finding t he mo st ro bust op ti2 t he w hole test case set is co unted and signified as mal weight vecto r based o n t he limited test / refer2 t he t raining classificatio n accuracy. ence set , w hich has a goo d pop ularized capabilit y , The advantage of applying t his t raining policy is t he aim of t he weight lear ning p rocess. This can is t hat t he op timal weight vecto r can be searched by be do ne by t he validatio n of a given op timal weight means of a limited o r relatively small set of testing ( ) ( ) vecto r w hich is acquired by Policy 1o r 2o r a cases if p roperly designed. However , different test mixed policy t hro ugh assessment of t he weight vec2 ratio s co uld give different searched op timal weight to r o n t he w hole t raining set in a case base . In t his vecto rs. The validatio n p rocess has to be perfo r med case , t he validated classificatio n accuracy must be to choo se t he mo st ro bust op timal weight vecto r reco rded by applying t his given op timal weight vec2 ( ) 2granular f ract ure 28 . The number ro sio n interto r o n t he w hole t raining set by L eave One Out policy. The validatio n criteria co nsidered can be in bracket s denotes t he amo unt of t raining cases fo r described as follow s : each f ailure mo de . The reaso n fo r such selectio n is t hat t hese mo des are in t he same level and bal2 () 1Criterio n fo r effectiveness of t raining If co mpariso n bet ween a t raining classificatio n anced2dist ributed. Clearly t he t raining experiment ( ) accuracy denoted by A fo r an op timal weight based o n o nly 75 cases is in f act a small sample ( vecto r and it s validated classificatio n accuracy de2 p ro blem. ) noted by B gives result s such t hat : 6 . 1 Eff ects of diff erent tra in ing pol ic ies () iA is much higher t han B , t hen t he pop u2 Fig13 show s t he t raining and validatio n result s larized capabilit y of t he resulted op timal vecto r is ( of applying Test Reference Set policy i . e . p roved to be poo r and ineffective ; ( ) ) Policy 1 at different test ratio s w hich change () A is much lower t han B , t hen t he unreli2 iif ro m 0105 to 0195 by an increment of 0105 . The able o r unp redictable result s wo uld be anticipated , solid line rep resent s changes of t raining classifica2 t herefo re t he resulted weight vecto r is still regarded tio n accuracies wit h test ratio s by applying such as ineffective ; policy , w hereas t he dot ted line rep resent s changes () iiiA is near o r equal to B , t hen t he op timal of validated classificatio n accuracies by applying t he weight vecto r is co nsidered to be co nsistent . And () op timal weight vecto r gained t hro ugh Policy 1to f urt her mo re , if A o r B is greater t han a user speci2 t he w hole 75 t raining case subset . ( ) fied percentage value e . g . , 70 %, t he op timal weight vecto r is p roved to be stable and effective . () 2Criterio n fo r efficiency of t raining Fo r an effective op timal weight vecto r , t he co mpariso n of it s validated classificatio n accuracy ( ( ) ) i . e . B value in iii above wit h t he t raining classificatio n accuracy of directly applying t he L eave One Out policy o n t he w hole t raining set available is co nducted. And if t he values of bot h ac2 curacies are near o r same , t hen t he op timal weight ( vecto r w hich is in f act acquired by applying Policy ( Pop ulatio n size :60 max generatio ns : 300 ; cro ss rate : 019 ; muta2 () ( ) ) 1o r 2o r mixed o n a limited test case set is ( )tio n rate :0105 ; search interval : 5 , 55 ; divisio n accurac y : 1 p roved to be efficient . ()Fig. 3 Classificatio n accuracy curves by Policy 1 Clearly , o nly if t he effectiveness of t raining is It can be seen t hat at interval of 0105,0135 , satisfied , t he evaluatio n of efficiency of t raining t he t raining classificatio n accuracies are much high2 will make sense . er t han t he validated classificatio n accuracies. It 6 Experiment s and Discussio ns explains t hat t he sufficient lear ning samples in a reference set can guarantee t he t raining accuracy , In t his st udy , 358 f ailure analysis cases of but as a result of too few testing cases in a test set , aero nautical equip ment have been collected f ro m it may not po ssess goo d pop ularized capabilities , as jo ur nals and f ailure analysis repo rt s. A f ailure anal2 displayed by low values of validated classificatio n ysis case base is established by using Access accuracy. At interval of 016 ,0195 , t he t raining database technology. Here selecting a t raining case classificatio n accuracies decreased , w hich demo n2 subset covering t hree f ailure mo des fo r experi2 st rates lear ning samples in a reference set is rat her ( ) ments , i . e . low cycle f atigue f ract ure 25 , ( ) limited , so as to unable to give reliable lear ning high cycle f atigue f ract ure 22, and st ress co r2 classificatio n result s. The co nsistency o nly exist s bet ween test ratio s at interval of 0135 , 015 , w here t he t raining classificatio n accuracies coincide wit h t he validated classificatio n accuracies and keep stable and as high as 74167 %. Fig. 4 showes result s by applying t he L eave (() ) One Out policy i . e . Policy 2at different test ratio s. Fig15 Classificatio n accuracies by t he mixed policy criteria described in Subsectio n 513 , t he mixed pol2 icy gives bet ter t raining effect s t han t he ot her t wo t raining policies and po ssesses t he highly efficient op timal weight vecto rs. Table 1 list s t he best e2 quivalent weight vecto rs at different test ratio s in applying t he mixed t raining policy. Ta ble 1 Best weight vectors at diff erent test ratios ()Fig14 Classificatio n accuracy curves by Policy 2 Test ratio s Best equivalent weight vecto rs It can be seen t hat w hen test ratio s lies be2 0 . 35 , 0 . 4 , 0 . 45 , 0 . 5 , ( )2 ,1 ,1 0 . 7 , 0 . 75 , 0 . 8 , 0 . 85 t ween 011,015 , t he t raining classificatio n accura2 ( )0 . 55 , 0 . 6 , 0 . 65 , 0 . 9 , 0 . 95 3 ,1 ,1 cies and validated classificatio n accuracies o scillate ( ) 1 . 0 5 ,1 ,1 much and show much differences , w hich p roves The ro bust op timal weight vecto r sho uld be t hat t he L eave One Out policy used fo r t raining () 2 , 1 , 1 w hich occurs mo st f requently amo ng all at small test ratio s is ineffective , w hereas t he re2 t he test ratio s. sult s f ro m 015,110 give co nsistent and co nvergent 6 . 2 Test on ident if icat ion capa bil ity using other values of t wo classificatio n accuracies , and validates fa il ure modes t hat t he classificatio n accuracy keep as high as 74167 % as show n in Fig13 . This demo nst rates 2 To check t he pop ularized capabilit y of t he opt hat w hen t he test case set is relatively large , Poli2 timal weight vecto r gained t hro ugh above lear ning ( () p rocess , t he applicatio n of t he weight vecto r 2 , cy 2can safeguard t he t raining effectiveness. ) 1 , 1 to identificatio n of ot her t wo f ailure mo des Based o n t he analysis described above , a by CB R p rocess is co nducted. In t his case , 15 brit2 mixed t raining policy is p ropo sed , w hich suggest s tle cleavage f ract ure f ailure mo de cases and 14 t hat w hen test ratio s lies in 0125,015 , t he Test t her mal f atigue f ract ure f ailure mo de cases are Reference Set policy sho uld be applied , and w hen t he test ratio s lies in 0155,110 t he L eave One co nsidered . Table 2 show s t he identificatio n result s Out policy sho uld be applieds. Fig15 gives result s and t heir co mpariso n wit h t he t raining effect s by of applying t he mixed t raining policy. applying L eave One Out policy alo ne to t his 28 Clearly , bot h curves of t raining classificatio n f ailure cases. accuracy and validated classificatio n accuracy show 2 It can be seen f ro m Table 2 t hat t he percentmuch imp roved co nsistency , and t he validated clas2 ( ) age values of identificatio n by 2 , 1 , 1vecto r and sificatio n accuracy of 74167 % maintains co nstant by L eave One Out policy are quite near and since t he test ratio of 0135 . Acco rding to validatio n high , w hich p roves t hat t he op timal weight vecto r v s the best weight vectors Ta ble 3 Test ratios () 2 ,1 , 1o btained t hro ugh above 75 t raining cases can be pop ularized to ot her f ailure mo de identifica2 Test ratio s Best equivalent weight vecto rs 0 . 35 , 0 . 4 , 0 . 45 , tio n cases. ( )2 ,1 ,1 0 . 5 , 0 . 95 , 1 . 0 Ta ble 2 The recogn ition eff ect on other patterns ( ) Ot hers 1 ,1 ,1 by weight vector ( 2 , 1 , 1) No . of Identificatio n L eave One Failure mo des ( ) cases by 2 ,1 ,1Out effect s 7 Co nclusio ns ( ) ( )11 73 . 3 %11 73 . 3 % Brit tle cleavage f ract ure 15 ( ) ( ) Ther mal fatigue f ract ure 14 12 85 . 7 %13 92 . 8 % The GA based weight t raining p rocess has been co nducted in a CB R system fo r f ailure mo de 2dis2 6 . 3 Tra in ing eff ects of mixed unbalanced identificatio n of aero nautical equip ment . Failure at2 tributed fa il ure mode ca ses t ributes used fo r matching in case ret rieval p hase Fig. 6 show s t he t raining result s of mixing 75 are gro uped into t hree catego ries , and a weighed K Failure cases described in Subsectio n 611 wit h 28 Nearest Neighbo r app roach is adop ted fo r similarit y Failure cases described in Subsectio n 612 The total measures bet ween an old o r sto red case and t he new of 5 f ailure mo des are co nsidered w hich are clearly o r target case . The fit ness f unctio n of GA fo r evo2 unbalanced2dist ributed. The t raining policy adop t2 lutio nary searching of op timal weight vecto rs is ed is t he mixed policy described in Subsectio n 611 . caref ully st udied. The perfo r mance of t he system is tested based o n t he co nsideratio n of t hree kinds of t raining policies , in w hich t he mixed t raining poli2 cy is newly p ropo sed by t he co mbined use of Test Reference Set policy and L eave One Out policy at different test ratio s. The evaluatio n o r validatio n of t raining effect s are st udied and novel validatio n criteria are set up . The experimental result s show t hat ( ) 1 The sufficient and balanced2dist ributed f ailure mo de t raining case set can give bet ter t rain2 Fig. 6 Classificatio n accuracies for 5 modes ing result s t han unbalanced o nes. The t raining curves show t hat t he stable vali2 ( ) 2The highly efficient op timal weight vecto r dated classificatio n accuracy reached 63146 % w hen can be achieved by using t he mixed t raining policy. t he test ratio s are at intervals of 0135 , 015 and ( ) 3 The pop ularized capabilit y of an op timal 0195,110 . Anot her stable value is 61146 % w hichweight vecto r by applying t he mixed t raining policy is slightly less at interval of 0155 , 019 . This is effective and ro bust fo r identificatio n of mo re demo nst rates t hat t he mixed t raining policy can f ailure mo des , and is valid fo r p ractical use . still achieved relatively bet ter result s of identifica2 tio n fo r unbalanced2dist ributed f ailure case set . Ref erences Nevert heless , it is argued t hat t he best t raining ef2 1 钟群鹏 ,张峥 ,田永江. 机械装备失效分析诊断技术J . 北 fect wo uld be taken in t he mo st balanced2dist ribut2 ( ) 京航空航天大学学报 ,2002 ,28 5:497 - 502 . ed f ailure mo de cases. Table 3 list s t he op timal Zho ng Q P , Zhang Z , Tian Y J . Failure analysis diagno sis of weight vecto rs fo r t he 5 f ailure mo de identificatio n mechanical equip ment J . Jo urnal of Bei jing U niversit y of ( ) ( Aero nautics and Ast ro nautics , 2002 , 28 5 : 497 - 502 . in p rocesses wit h different test ratio s. )Chinese Again , it is show n t hat t he ro bust op timal Maye D. Boiler t ube failure mechanism recognitio n —an ex2 2 () weight vecto r is 2 ,1 ,1as acquired in Subsectio n ( ) pert systemJ . CIM Bulletin , 1990 , 83 939:92 - 95 . Ko mai K , Mino shima K I. Recognitio n of different f ract ure 3 6 . 2 . surface mo rp hologies using co mp uter image p rocessing tech2 1989 . Shin K S , Han I G. Case based reaso ning suppo rted by ge2 ( ) 9 nique J . J SM E Int J Series A , 1993 , 36 2:220 - 227 . 4 netic algo rit hms fo r co rpo rate bo nd ratingJ . Ex pert System Liao T W , Zhan Z H , Mo unt C R. An intergrated database ( ) and expert system fo r failure mechanism identificatio n : Part I wit h Applicatio ns , 1999 16:85 - 95 . ( ) and Part I I J . En gineering Failure Analysis , 1999 6 : 王玉 , 邢渊 , 阮雪榆. 基于事例的推理循环中人工神经网 10 387 - 421 . 络和遗传算法的 4 种应用模型J . 上海交通大学学报 , ( ) Liao T W , Zhan Z H , Mo unt C R. A case based reaso ning 5 2003 , 37 2:202 - 204 . system fo r identif ying failure mechanismsJ . Artificial Intel 2 Wang Y , Zho u Y , Ruan X Y. Fo ur applicatio n mo dels fo r ( ) ligence ,2000 13:199 - 213 . artificial neural net wo r k and genetic algo rit hm in case based 徐元铭 , 骆晶妍 , 邢小楠. 航空机械装备失效分析专家系 reaso ning cycle J . Jo urnal of Shan ghai J iaoto ng U niversit y , 6 ( ) 统应用研究J . 机械设计与研究 ,2003 ,19 1:53 - 56 .( ) ()2003 , 37 2:202 - 204 . in Chinese Xu Y M , L uo J Y , Xing X N . The applicatio n of expert sys2 tem fo r failure analysis of aero nautical equip ment J . Ma 2 Biogra phies : ( ) ( chine Design and Research , 2003 , 19 1 : 53 - 56 . in Chi2 ( )He received XU Yuan2ming 1965 - )nese his Ph. D degree in Aero nautical Engi2 7 徐元铭 ,张洋. 实例推理技术在航空动部件失效分析中的neering. He is now a p rofessor and doc2 ) ( 应用J . 材料工程 ,2003 增刊:289 - 291 .torial st udent supervisor of Beijing U ni2 Xu Y M , Zhang Y. The applicatio n st udy of case based rea2 versit y of Aero nautics and Ast ro nautics. so ning fo r failure analysis of aero nautical equip ment J . Ma 2 His research interest s include : aircraf t ( ) () terial Engineering , 2003 Sup: 289 - 291 . in Chinese st ruct ural op timizatio n , intelligent CAD , Goldberg D E. Genetic algo rit hms in search , optimizatio n 8 intelligent based fault diagnosis and failure analysis. and machine learning M . Lo ndo n : Addisio n 2Wesley ;