畢業(yè)論文外文翻譯-并條機自調(diào)勻整利用人工神經(jīng)網(wǎng)絡(luò)確定在自調(diào)勻整作用點(外文原文+中文翻譯)

ResearchseofArtificialNeuralNetworkstheActionPointatAuto-levelingDrawAssadFarooq1andChokriofTextileClothingTechnology,TechnischeUniversit?tDresden,Abstract

ArticleArtificialneuralnetworkswithabilityoffromhavebeenappliedinthetextileThepointisofimportantparametersofthedrawingqualityofreportsamethodoftheactionusingartificialnetworks.actionvariableswereselectedasfortrainingtheartificialneuralnetworkswiththeaimtooptimizeauto-levelingbylimitingthelevelingactionpointsearchrange.TheMarquardtalgorithmisincorporatedintotheback-propagationtoacceleratetheBayesianregularizationistheofnetworks.arequitepromising.Keyartificialauto-lev-eling,frame,levelingactionpoint。Theoftheanroleinwhilesliverisofthecriticalwhenqualityyarn.ismajorcriteriaforassessmentofoperationofInprinciple,twotoirregularities.istothedraftingandrecognizetheforirregularities,thatmeansbethem.Thevaluableistouseauto-levelersintheinadequatetocorrectthevariationsinsliver.Thecontrolofirregularitiesloweronsliveruniformity,ambientframeparameters.Atthedrawframevariationsinarecontinuallybyamechanicaldevice(aroll)convertedintoelectricalThetoanelectronicwithvariable,timedelayedThetimeallowsdraftbetweenthemid-rollthedeliveryrollofdrawtoadjustthatmomentwhendefectivesliverpiece,whichhadmeasuredbyapairofscanningfindsitselfatapointofAtthispoint,dependingupontheamountofinsliverpiece.ThedistancethatscanningpairofdraftcalledthepointofregulationthelevelingactionpointinFigure1.Thisleadstocalculatedthecorrespondingdefective[2,3].Inframes,especiallyintheofoffibermaterial,orbatchesmachineparametersmustbeTheLAPismostauto-levelingparameterwhichisinfluencedbyparametersspeed,material,draftgauge,draftgauge,feedingtension,draft,andsettingtheetc.UseofArtificialNeuralfortheActionPointA.C.1ofanPreviously,sampleshadbewithsettings,tothelaboratory,onthetesteroptimumLAPwasdrawautomaticsearchfunctionforoptimumofLAP.Duringthisfunction,isscannedbyLAPcess,thequalityparametersareconstantlyalgorithmautomaticallycalculatestheoptimumLAPbythepointwiththeminimumsliverCV%.Atpresentasearchof120mmscanned,21examinedusing100ofineach2100ofisnecessarysearchatime-consumingmethodaccompaniedbythehencedirectlyaffectingcostparameters.Inthiswork,wefindthepossibilityofLAP,artificialneurallimitautomaticsearchandtheabove-mentioneddisadvantages.ArtificialNeuralNetworksThemotivationofusingartificialneuralnetworksliesintheirflexibilityofinformationnothave.Theneuralnetworksystemcansolveaproblem“byexperienceandlearning”theinputoutputpatternsprovidedbytheuser.Inthefieldofartificialneuralnetworks(mostlyhavebeenextensivelystudiedthelasttwodecades[4–6].thefieldofpreviousonpredictingpropertiesandspinningprocessusingthefiberpropertiesacombinationoffiberpropertiesandmachinesettingsasofneuralnetworks[7–12].Back-propagationissupervisedmostfrequentlyforneuraltraining.back-propagationalgorithmbasedtheWidrow-Hoffdeltaruleinwhichweightoutthroughthemeanofoutputtoinput[13].ThesetpatternstheuntiltheTheusessteepestdescentamethodtodetermineofOverfittingTheofneuralnetworktrainingproduceanetworkproducesontrainingwhichalsorespondsproperlynovelinputs.WhennetworkwelloninputsasonsetthesaidtobeThecapacitytheislargelybythe(numberofhiddenneurons)thisvitalrolethetraining.Anetworkwhichnottoalltheinformationintheisbeunderfitted,whileathatiscomplextofitthe“noise”inthedataleadstooverfitting.“Noise”meansvariattheunpredictablefrominputsofspecificAllstandardneuralarchitecturesasfullytooverfitting.itisdifficultacquirethenoisefromspinningdependenceofonthevariationsandenvironmentalconditions,etc.stoppingcommonlyusedtotacklethisproblem.Thisinvolvesofdataintothreei.e.atrainingavalidationsetandwiththatlargepartofthedata(validationset)canneverbetheofTheotherofoverfittingwhichthemethodofimprovingconstrainingofnetworkweights.[14]discussedBayesianforback-propagationnetworks,whichconsistentlyproducedwithgoodTheobjectiveofprocessismini-mizetheoferrors:n(t)Diii

2

（1）Wheretthetargetsandaresponsestheiitoreduceofsquarederrorsregularizationadditionalterm,theD

（2）Inequation

E

isthesumofsquaresofthenetworkweights,andαandwfunctionparameters.relativeofobjectivefunctionparametersdictatestheemphasisfortraining.Ifα<<β,thenthetrainingwilldrivetheerrorssmaller.α>>β,trainingwillemphasizeweightsizereductionattheexpenseofnetworksmootherresponse[15].Theofisonviewwhatittofromdata,inwhichusedtorepresenttheuncertaintythebeinglearned.anythepriorwhatmightbecanbeindistributionthatdefineAfterthedata,therevisedarebyaposteriordistributionoverweights.Networkthatplausiblebutwhichthewillbebeingmuchwhiletheforvaluesofweightsdofitthedatawillhaveincreased[16].InframeworktheweightsofareconsideredAfterthetaken,thefunctionfortheweightsbeupdatedaccordingtoBayes’rule:(w/D,

M)

(D/,)(w/(D/M)

M)

(3)InequationDdataset,Mneuralnetworkwistheofweights.P(/,M)isthepriorprobability,whichrepresentsourofbeforeanydatais(D/,

M)whichprobabilityofdataoccurring,giventheweightsw.Pw

)isathatthetotal1[15].Inthisstudy,employedMATLABNeuralNet-worksToolbox“trainbr”whichisanincorporationoftheLevenbergMarqaurdtregularizationtheorem(orBayesianintotrainthenetworktoreduceofapproximationofHessianmatrixproducegoodcapabilities.Thisaofparametersbiases)effectivelyusedbytheofshouldthesame,irrespectiveoftotalofparametersinThiseliminatesrequireddeterminingthenetworkExperimentalTheexperimentaldatawasobtainedfromRieter,ofdrawframeRSB-D40[17].Forthematerialexperimentalwasbasedofuseinspinningindustry.Forexample,isthefrequentlyusedmaterial,soitwereperformedoncottonwithallsettings,wasnotcasewithotherAlso,owingtothefactthatallcouldnotbeprocessedwithsamepressuredraftdifferentweredesigned.Thewiththeirplansgivenin1.Theprocedurewasappliedtomaterialsandstandardforsettings(sliverlineardensity,LAP,wasadopted.AcomparisonofautomaticwasperformedandresultswerethesearchfunctionfromRSBD-40.ThereforetheLAPsearcheswereaccomplishedbyQualityMonitorabstractoftheinFigure2.UseofArtificialNeuralfortheActionPointA.C.2Abstractnetworkmodel.HeretobeconsideredisthatisnoinmachinetheLAPinfluencingparameter,i.e.feedingSofeedingspeedconsideredberelatedspeedofdoublings(4).Thedeliveryspeedvariedbetween3001100thedoublings7,toachieveofCount×speed/（Doublings×Feed）（）TrainingandTestingSetsFortrainingtheneuralnetwork,thedataintofirstphaseincludedexperimentaldataforcompilationthedatasubsequentThepriortheinfluencingi.e.speed,deliveryspeed,gaugesofdraft,andofthesliverwastoselecttheSofirstphasecontainedtheinwhichstandardweretakenasafoundationandthenLAPinfluencingineachInsecondphase,experimentswereselectedinwhichinfluencingparameterandnetworkwasallowedtotheThiswasontheofinfluencingparameters,withtheaimtoLAPlength.thirdphaseinvolvedtheexperimentsconductedonpilotscaleThesepilotwerecarriedbymachinefordifferentsettings.theseresultsweretoassesstheoftheneuraldataNormalizingtargetvariablesmaketrainingimprovingnumericalconditiontheproblem.Alsoitcanreducechancesofinforneuralnetworkofthelargespansofnetwork-inputdata,targetsforbetterperformance.Atfirsttheinputswerebetweeninterval[whichnotanypromisingresults.databetweentheinterval1]networkswithsuccess.NeuralNetworkTrainingWefiveneuralnetworkstopredictLAPbyofdataThesetsdividedintotrainingasinwasperformedtrainingweretotheofneuralinformFigure3thetrainingperformanceofthenetwork5.3TrainingperformanceofNNResultsandDiscussionalreadyinTable2,differentofdatawereusedtotrainfivenetworksatheirtestonunseendata.Asistoeliminatetheguessfortheappropriateofhiddentheofhiddenrangingfrom1422,wereintwolayerstrainnetworks.theselectionoftopologythatshouldbecomplextolearnalltheintheaspossibilityofwastackledwithduringtrainingthetestsetsbeseenfromFigures44ofNN1.UseofArtificialNeuralfortheActionPointC.5ofNN2.6ofNN3.7ofNN4.8ofNN5.wascalculatedtopredictedactualvalues,bothduringtrainingandtesting,andthearepresentedinFigure9.overallresultscanbeexplainedthepresenceorof–outputinteractionsindataandtheincreaseinthewiththeoftrainingdatasets.Figure9clarifiestheincreaseintestingerror,evenwithanincreaseintheoftraining,sets,whentrainingandperformedonfromdifferentphases(NN2&2).theerroraofphasewastrainremainderusedtesting,inNN3and5comparisonwithNN2and4,Thepresenceofdifferentinputoutputinteractionsindifferentphasesthistrend.behaviorofNN1withtoabove-mentionedfactattributedtosetsinphase1.9betweentrainingtestsets.InordertoassesstheoffitofNN5,10-foldcross-validationpreformed,offorandfortesting,timestestingnetworktimeon10%ofunseenAnaverageR

2

reported.wasadoptedfor80%andtestthecalculatedof

2

wasintheperformancethe80%datasetsforconfirmaveryfitofNN5ConclusionTheartificialnetworkmodelwasthenetworkstrainedatofTextileClothingTechnology,TechnischeUniversitatDresden.TheuseofBayesianregularizationtestingforpracticalapplicationspromisingresults.testingofNN5asin3,aofabout2mmisobservable,whichfallswellwithinmmnegligiblerangedeterminationofLAP.Thisthatcanbeinthefutureforquickcomputationofthewiththeadvantagesofofmaterialandtime.TheaccuracyincomputationcanleadtoCV%quality.TestofNN紡織研究期刊并條機自調(diào)勻整利用人工神經(jīng)網(wǎng)絡(luò)確定在自調(diào)勻整作用點摘要用他們的人工神經(jīng)網(wǎng)絡(luò)從數(shù)據(jù)中學(xué)習(xí)的能力已經(jīng)成功的應(yīng)用在紡織行業(yè)，在并條機自調(diào)勻整參數(shù)中勻整作用點是一個重要的參數(shù)且強烈的影響到紗線的質(zhì)量本文主要是講述的是用人工神經(jīng)網(wǎng)絡(luò)來檢測勻整作用點各種各樣的勻整作用點會影響導(dǎo)致不同的變量，我們把這些變量輸入到人工神經(jīng)網(wǎng)絡(luò)進(jìn)行測試，目的是為了通過限制勻整作用點改變范圍來優(yōu)化并條機的自調(diào)勻整，Levenberg-Marquardt算法是納入反向傳播加速系和貝葉斯正則化方法,進(jìn)了網(wǎng)絡(luò)的泛化。得到的結(jié)果是非常的精確的。關(guān)鍵詞

人工神經(jīng)網(wǎng)絡(luò)自調(diào)勻整并條機勻整作用點紗線的均勻度在紡織行業(yè)中發(fā)揮著越來越重要的角,

而棉條均勻度是產(chǎn)生高質(zhì)量紗線的重要因素之一。棉條均勻度也是并條機運轉(zhuǎn)的主要評價標(biāo)準(zhǔn)。在原則上,有兩種方法可以減少棉條不均勻。一種途徑是研究牽伸機構(gòu)和認(rèn)識棉條不均的原因,以這意味著可能會發(fā)現(xiàn)可以降低不均勻度的方法。其他更有有效的方法是使用自調(diào)勻整,因為在大多數(shù)情況下并條機是不足以正確的改變紗線的均勻度。長片段不勻控制在一個較低水平，取決于梳棉棉條均勻性、環(huán)境條件、和幀參數(shù)。在自調(diào)勻整并條機（RSB-40喂入的棉條厚度的不斷變化變化可以通過一種機械檢測設(shè)備來監(jiān)測其變化并且將監(jiān)測到的信號轉(zhuǎn)換為電子信號量數(shù)據(jù)被傳輸?shù)揭粋€可變電子記憶器里時間延遲響應(yīng)當(dāng)棉條片段有缺陷的時候可以在一定的時間延遲允許并條機牽伸的中間羅拉和傳送羅拉之間來進(jìn)行棉條的不均勻度精確調(diào)整當(dāng)已被一對檢測羅拉測量后發(fā)現(xiàn)在本身的一個牽伸點在這一點上伺服電機運行取決于在棉條部分被檢測的數(shù)量的變化一對分離的檢測羅拉和牽伸點之間的距離，稱為零點調(diào)節(jié)或勻整作用點LAP），如下1所示，這在計算相應(yīng)有缺陷的材料應(yīng)該進(jìn)行相應(yīng)的修正2、3）,在并條機的自調(diào)勻整裝置上,特別是在纖維材料改變的情況下或者是各種不同型號機器選擇和過程控制參數(shù)必須優(yōu)化整作用點是自調(diào)勻整中最重要的參數(shù)許多的因素影響，比喻喂條速度，棉條材料，前牽伸測量計，喂入張緊力，前牽伸和棉條導(dǎo)條羅拉設(shè)置等。insliver:均勻度變化的棉條

Draft主要牽伸區(qū)Rollers:檢測羅拉BreakDraft前牽伸區(qū)PressuereBar:壓力棒actionpoint勻整作用點此前，棉條樣品必須在實驗室在不同的條件下進(jìn)行，在測試儀上進(jìn)行檢測棉條均勻度，直到找到最好的勻整作用點（手動進(jìn)行檢測RSB-40并條機的自調(diào)勻整具有自動搜索確定最好的勻整作用點這一功能，棉條在進(jìn)行此功能期間，棉條被自動檢測是通過調(diào)整臨時的勻整作用點和被記錄的有用的數(shù)據(jù)此過程中，不斷的監(jiān)測的質(zhì)量參數(shù)和通過選擇棉條的最低點的棉條的V%，通過一種算法自動計算出最佳的自調(diào)勻整作用點

，

目前的一個搜索檢測范圍是120米，例如在每檢查100m的棉條是21點，因此2100m的棉條必須進(jìn)行搜索功能，這是一種比較費時的方法材料和生產(chǎn)的損失會直接影響成本在這項工作中我們試圖找出一種能夠預(yù)測出勻整的作用點利用人工神經(jīng)網(wǎng)絡(luò)限制自動搜索范圍和減少上述缺點。人工神經(jīng)網(wǎng)絡(luò)利用人工神經(jīng)網(wǎng)絡(luò)的目的在于他們的靈活性和信息處理能力，傳統(tǒng)的計算方法不具備。神經(jīng)網(wǎng)絡(luò)系統(tǒng)可以由用戶提供的輸入輸出模式通過“經(jīng)驗和學(xué)習(xí)”這種方式來解決問題，在紡織品領(lǐng)域，人工神經(jīng)網(wǎng)絡(luò)主要使用反向傳播)在過去20年已經(jīng)在廣泛的研究紡織領(lǐng)域之前的研究領(lǐng)域一直專注于預(yù)測紗線性能和紡紗纖維工藝使用性能或結(jié)合纖維的性能和機器型號作為的人工神經(jīng)網(wǎng)絡(luò)的輸入。反向傳播是一種監(jiān)督學(xué)習(xí)技術(shù)，頻繁用于人工神經(jīng)網(wǎng)絡(luò)的訓(xùn)練。反向傳播算法是基于Widrow-的霍夫三角洲學(xué)習(xí)法則，其中權(quán)重是通過樣輸出到樣品輸入之間的響應(yīng)的平方差進(jìn)行調(diào)整。這系列單一模式反復(fù)提供給網(wǎng)絡(luò),直到誤差值最小化差反向傳播算法采用最快下降法這實質(zhì)上是一個一階的方法確定一個合適的梯度運動方向。過度擬合神經(jīng)網(wǎng)絡(luò)的訓(xùn)練的目標(biāo)是產(chǎn)生一個網(wǎng)絡(luò)這個開發(fā)的網(wǎng)絡(luò)上產(chǎn)生較小的錯誤,并且這種情況依然可能會反映出的是一個異常的輸入。當(dāng)一種網(wǎng)絡(luò)模式異常的輸入作為研發(fā)系統(tǒng)的輸入,這種網(wǎng)絡(luò)被認(rèn)為是廣義的網(wǎng)絡(luò)。很大程度上是受到網(wǎng)絡(luò)的泛化能力的網(wǎng)絡(luò)架構(gòu)(一些隱藏的神經(jīng)元這在開發(fā)過程中起著至關(guān)重要的作用,而網(wǎng)絡(luò)是太復(fù)雜的，不足以來收集的所有信息數(shù)據(jù)，這被認(rèn)為是不充分?jǐn)M合,而當(dāng)一個網(wǎng)絡(luò)是過于復(fù)雜,那么很難適應(yīng)與紊亂的數(shù)據(jù)將會導(dǎo)致過度擬合噪音”的意思是從一個特定的網(wǎng)絡(luò)中的輸入值中變化的目標(biāo)值是預(yù)測不到的所有的標(biāo)準(zhǔn)例如完全連接的多層感應(yīng)器神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)容易出現(xiàn)過擬合此外從紡織行業(yè)它是很難獲得沒有紊亂的數(shù)據(jù)由于最終產(chǎn)品依賴于固有材料變化和環(huán)境條件等問題解決這個問題最普遍的方法是阻止這些問題的出現(xiàn)這包括訓(xùn)練的數(shù)據(jù)劃分成的三個組,即一個訓(xùn)練集驗證集和測試集,這其中有個缺點是一個大量的數(shù)據(jù)部分(驗證系統(tǒng))不可能是訓(xùn)練的部分。正規(guī)化另一個過擬合解決方案是正規(guī)化這是提高泛化通過約束網(wǎng)絡(luò)的權(quán)值的大小的方法。麥凱討論了一個實用的貝葉斯框架網(wǎng)——反向傳播網(wǎng)絡(luò)連續(xù)生產(chǎn)的網(wǎng)絡(luò)具有良好的泛化。最初的開發(fā)的過程中目標(biāo)是是求最小平方誤差的總和n(t)Diii

2

（1）其中t是目標(biāo)函數(shù)和a是神經(jīng)網(wǎng)絡(luò)反應(yīng)各自的目標(biāo)通常情況下開發(fā)的目標(biāo)是減少平方差和F=Ed然而正規(guī)化增加了一個額外的長期目標(biāo)函數(shù)這是由以下函數(shù)得到的：D

（2）在方程（2）中，是網(wǎng)絡(luò)權(quán)重的平方的總和是目標(biāo)函數(shù)的參數(shù)，目標(biāo)函數(shù)參數(shù)的相對大小決定了訓(xùn)練的重點，如然后訓(xùn)練的算法就會產(chǎn)生較小的誤差練的神經(jīng)網(wǎng)絡(luò)就會在網(wǎng)絡(luò)錯誤代價強調(diào)權(quán)重的減小，從而產(chǎn)生一個流暢的網(wǎng)絡(luò)響應(yīng)。貝葉斯學(xué)校的統(tǒng)計數(shù)據(jù)是基于一個完全不同的角度來獲取數(shù)據(jù)其中的概率是用來表示關(guān)系的不確定性有學(xué)問看到任何數(shù)據(jù)之前先前的看法可能是真正的關(guān)系可以用概率分布來表示通過網(wǎng)絡(luò)權(quán)值來定義這種關(guān)系方案構(gòu)思后的數(shù)據(jù)訂后的觀點通過之后的絡(luò)權(quán)值分布來捕獲以前網(wǎng)絡(luò)權(quán)重似乎有道理，但是與數(shù)據(jù)不相符現(xiàn)在被看做是不太可能而對不適合數(shù)據(jù)的權(quán)重值的概率將增加。在貝葉斯框架的權(quán)重網(wǎng)絡(luò)被認(rèn)為是隨機變量。數(shù)據(jù)被采取后后驗概率密度函數(shù)的權(quán)重可以根據(jù)貝葉斯規(guī)則更新:(w/D

M)

(D/,)(w/(D/M)

M)

（3）在方程(3),D代表數(shù)據(jù)集,M是使用于特定的神經(jīng)網(wǎng)絡(luò)模型和w是網(wǎng)絡(luò)權(quán)值向量。(w/

,)是先驗率,代表我們的知識收集數(shù)據(jù)之前權(quán)重P(D/w

M)是似然函數(shù),這是數(shù)據(jù)的可能性發(fā)生,由于重量w.(/它可以保證總概率是1。

)一個標(biāo)準(zhǔn)化的因素,在這項研究中,我們使用矩陣神經(jīng)微機網(wǎng)絡(luò)化工具箱函數(shù)”trainbr”,這是一個公司麥夸特法和貝葉斯正則化定理(或貝葉斯學(xué)習(xí)融入到反向傳播來訓(xùn)練神經(jīng)網(wǎng)絡(luò)來減少計算超過海森矩陣的近似值和產(chǎn)生好的的泛化能力種算法提供了一個網(wǎng)絡(luò)參數(shù)量度(權(quán)值和偏差)被網(wǎng)絡(luò)有效的利用不考慮在網(wǎng)絡(luò)中參數(shù)的總數(shù)，有效的參數(shù)的數(shù)量應(yīng)該保持相同的。這就消除了猜測需要在確定最優(yōu)網(wǎng)絡(luò)大小。測試實驗數(shù)據(jù)可從立達(dá),戈爾的制造商（并條RSB-D40）得到。對于這些實,在紡織行業(yè)材料的選擇和實驗設(shè)計是基于特定的材料使用的頻率。例如粗梳棉紗是使用最為頻繁的材料,所以它被作為標(biāo)準(zhǔn)和實驗研究在棉花的所有可能的設(shè)置,它不同于其他材料時,認(rèn)清這個事實,所有的材料無法進(jìn)行輥壓力和牽伸裝置,不同紡紗是事先是設(shè)置好的。材料的處理方案計在表中給出。標(biāo)準(zhǔn)的程序適用于所有的材料和標(biāo)準(zhǔn)程序?qū)τ谧哉{(diào)勻整裝置棉條線密度、強度)也被采納手動和自動搜索進(jìn)行比較從SBD-40自動搜索搜索功能獲得了了更好的CV%結(jié)果。因此，勻整作用點搜索時候由立達(dá)質(zhì)量監(jiān)控器（）完成的。一個實驗?zāi)Ｐ偷某橄竺枋鋈鐖D2所示。這里主要考慮的一點是在機械上調(diào)整勻整作用點主要影響因素是不可能的，例如，喂棉速度，因此給棉速度被認(rèn)