高精度溫度測量裝置的設(shè)計-外文翻譯及論大學(xué)生寫作能力

附錄ADesignofaHighPrecisionTemperatureMeasurementSystem1IntroductionSensorsareoneofthemostimportantelementsusedinmanyinstrumentationcircuits.Theyareusedinmanyindustrialapplicationsandtakeacertainformofinput(temperature,pressure,altitude,etc.)andconvertitintoreadingsthatcanbeinterpreted.Manytypesofsensorsarenonlinearinnaturefromwhichalinearoutputisdesired.Therearemanydifferentsensorsfortemperaturemeasurementandthermocouplesarethemostcommonlyused.Theyarepreferredinindustrialapplicationsduetotheirlowcost,wideoperationrange,fastresponsetimeandaccuratewhentheirpeculiaritiesareunderstood.Thermocoupleshavealsooutputsnonlinearlyrelatedtotemperature.Therefore,sensormodelingandlinearizationtechniquesarenecessary.Tosolvethelinearizationproblemofasensor,therearegenerallytwomethodsproposed.Thefirstonerequiresnonlinearanalogcircuitandthesecondusesnumericalmethodsthatarecomputedbymicroprocessororcomputer.Analogcircuitsarefrequentlyusedforimprovingthelinearityofthesensorcharacteristics,whichimpliesadditionalanaloghardwareandtypicalproblemsassociatedtoanalogcircuitssuchastemperaturedrift,gainandoffseterror.Usingthesecondmethod,sensornonlinearitiescanbecompensatedbymeansofarithmeticoperations,ifanaccuratesensormodelisavailable(directcomputationofthepolynomials),oruseofamultidimensionallook-uptable.Directcomputationofthepolynomialmethodismoreaccuratebuttakesalongtimeforcomputation,whilethelook-uptablemethod,thoughfaster,isnotveryaccurate.Inrecentyears,applicationofANNshasemergedasapromisingareaofresearchinthefieldofinstrumentationandmeasurement.Itprovidesaneurocomputingapproachforsolvingcomplexproblemsespeciallyinnonlinearsystemmodelingwhichthenetworkitselfisanonlinearsystem.Thisisextremelyusefulwhentheareaofinterestisabsolutelynonlinearincludingtheexperimentaldatathatisusedfortraining.OneofthemostpowerfulusesofANNsisinfunctionapproximation(curvefitting).InterpolationbasedonANNprovideslowerinterpolationerrorswhencomparedwithconventionalnumericalinterpolation.Intheworkwepresenthere,highprecisiontemperaturemeasurementsystembasedonANNapproachisproposed.ThecalibratingdataisobtainedbyWavetek9100calibrationunitthatisnecessaryforthetrainingandthetestingphasesoftheANN.ThehardwareandsoftwarepartsofthesystemareintegratedinaVIusedforsystemoperationandcalibration.TheANNismatchedtothecalibratingdatabyprovidingadesiredfinalerror.ThemeansquareerrorbetweencalibrationandtheANNmodeleddataisminimizedintermsofthestructure,numberoflayers,andnumberofneuronsbythedevelopedsoftware.2SystemHardwareAthermocouplegeneratesavoltageproportionaltothemeasurementjunctiontemperatureatmVlevelswhilethecoldjunctiontemperatureisconstant.Inordertomakeanaccuratemeasurementthecoldjunctiontemperaturemustbeknown.Figure1(a)showstheblockdiagramofthetemperaturemeasurementsystemdesignedviaanANNintheoperationphase.Itconsistsofathermocouple(typeE)exposedtoadesiredtemperature,includingsignalconditioningcircuitwith16-bitanalogtodigitalconverter(ADC)andInput/Outputinterfacecardinterfacingwithacomputer.Thedesignedsignalconditioningcircuithasaprogrammablegaininstrumentationamplifier(PGA204BP)withthegainof1,10,100and1000,a16-bitADC(AD976A),anAD595monolithicthermocoupleamplifierwithcoldjunctioncompensationwhichisconfiguredasastand-aloneCelsiusthermometeranda4channelanalogmultiplexer(ADG529A)whichselectthethermocoupleoroutputofCelsiusthermometer.TheAD976Aisahighspeed,lowpower16-bitA/Dconverterthatoperatesfromasingle5Vsupply.Thispartcontainsasuccessiveapproximation,switchedcapacitorADC,aninternal2.5Vreferenceandahighspeedparallelinterface.AccuracyofthesystemdependsdirectlyonstepsizeofADC.Witha±10Vinputs,oneLSBofAD976Ais305μV.WhenAD595isusedasaCelsiusthermometer,thethermocoupleisomitted,andthedifferentialinputsareshuntedtogethertocommon.Inthismode,AD595generateavoltagewithascalefactorof10mV/°Canditsoutputisusedforcoldjunctiontemperaturedatathatthewrittensoftwareisused.SomeimportantcharacteristicsoftheAD595are:operationtemperaturerange-55to125°C;stabilityvs.temperature:±0.05°C/°Candsensitivity:10mV/°C.OutputsignalofPGA204BPisdigitizedbyAD976AwhichitsoutputisconnectedtotheI/Ointerfacecardandtransferredtoapersonalcomputerwheredatareductionandoptimizationareimplemented.Fig.1.Measurementsystemblockdiagram:(a)operationalphase,(b)calibrationphaseToestablishtheANNsweightsandbiases,duringthecalibratingphase(ANNtrainingphase),Wavetek9100calibrationunit,withtheaccuracyof±0.006%+4.16μVintherangeof000.000mVto320.000mV,isconnectedtotheterminalsofanalogmultiplexertogeneratetabledthermocouplevoltagesasshowninFigure1(b).ThisvoltageisusedastheinputoftheANN,andthermocoupletemperaturewithoutcoldjunctioncompensationistheoutputoftheANN.Intheoperationphase(Figure1(a)),inordertomakethecoldjunctioncompensation,datatakenfromCelsiusthermometeroutputisused.TheoutputvalueofANNisshiftedbytheenvironmenttemperaturethatisobtainedbyCelsiusthermometer.ThenthisvalueisdisplayedontheVIasthethermocoupletemperature.ThedevelopedVIisusedtoacquirethedataforANNtrainingphaseandtoshowthecalculatedtemperatureintheoperationphase.Figure2showsthefrontpaneloftheVI.Themainfeaturesassociatedwiththisinstrumentare:displayofthemeasuredtemperatureandcorrespondingoutputvoltagefromconditioningcircuitforcollectingthedatainthecalibratingphaseandactualtemperaturewithcoldjunctioncompensationintheoperationphase.Thesystemiscontrolledbythesoftwarewritteninbothoperationandcalibrationphases.3ArtificialNeuralNetworkANNsarebasedonthemechanismofthebiologicallyinspiredbrainmodel.ANNsarefeed-forwardnetworksanduniversalapproximators.Theyaretrainedandlearnedthroughexperiencenotfromprogramming.Theyareformedbyinterconnectionsofsimpleprocessingelements,orneuronswithadjustableweights,whichconstitutetheneuralstructureandareorganizedinlayers.Eachartificialneuronhasweightedinputs,summationandactivationfunctionsandoutput.ThebehaviouroftheoverallANNdependsupontheoperationsmentionedontheartificialneurons,thelearningruleandthearchitectureofthenetwork.Duringthetraining(learning),theweightsbetweentheneuronsareadjustedaccordingtosomecriterion(Themeansquareerrorbetweenthetargetoutputandthemeasuredvalueforallthetrainingsetfallsbelowapredeterminedthreshold)orthemaximumallowablenumberofepochsisreached.Althoughthetrainingisatimeconsumingprocess,itcanbedonebeforehand,offline.Thetrainedneuralnetworkisthentestedusingdatawaspreviouslyunseenduringtraining.MLPsarethesimplestandmostcommonlyusedneuralnetworkarchitectures.Theyconsistsofinput,outputandoneormorehiddenlayerswithapredefinednumberofneurons.Theneuronsintheinputlayeronlyactasbuffersfordistributingtheinputsignalsxitoneuronsinthehiddenlayer.Eachneuronjinthehiddenlayersumsupitsinputsignalsxi,afterweightingthemwiththestrengthsoftherespectiveconnectionswjifromtheinputlayerandcomputesitsoutputyjasafunctionfofthesum,namelywherefisoneoftheactivationfunctionsusedinANNarchitecture.Traininganeuralnetworkconsistsofadjustingthenetworkweightsusingdifferentlearningalgorithms.Alearningalgorithmgiveswji(t)intheweightofaconnectionbetweenneuronsiandjattimet.Theweightsarethenupdatedaccordingtothefollowingformula:Therearemanyavailablelearningalgorithmsintheliterature.ThealgorithmsusedtotrainANNsinthisstudyareLevenberg–Marquardt(LM),Broyden–Fletcher–Goldfarb–Shanno(BFGS),BayesianRegularization(BR),ConjugategradientbackpropagationwithFletcher-Reevesupdates(CGF),andResilientback-propagation(RP)algorithms.NeuralLinearizationInthispaper,themultilayeredperceptron(MLP)neuralnetworkarchitectureisusedasaneurallinearizer.TheproposedtechniqueinvolvesanANNtoevaluatethethermocoupletemperature(ANNoutput)whenthermocoupleoutputvoltageisgivenasinput.TrainingtheANNwiththeuseofmentionedlearningalgorithmtocalculatethetemperatureinvolvespresentingitwithdifferentsetsofinputvaluesandcorrespondingmeasuredvalues.DifferencesbetweenthetargetoutputandtheactualoutputoftheANNareevaluatedbythelearningalgorithmtoadapttheweightsusingequations(1)and(2).Theexperimentaldatatakenfromthermocoupledatasheetsareusedinthisinvestigation.Thesedatasheetsarepreparedforaparticularjunctiontemperature(usually0°C).TheANNistrainedwith80thermocoupletemperaturesthatisuniformlydistributedbetween-200and1000°Cwhichisobtainedinthecalibrationphase.Howevertheperformanceofthefinalnetworkwiththetrainingsetisnotanunbiasedestimateofitsperformanceontheuniverseofpossibleinputs,andanindependenttestsetisrequiredtoevaluatethenetworkperformanceaftertraining.Therefore,theotherdatasetof20thermocoupletemperaturesthatisuniformlydistributedbetween-200and1000°C,isusedinthetestprocess.Theinputandoutputdatatuplesarenormalizedbetween-1.0and1.0beforetraining.Afterseveraltrialswithdifferentlearningalgorithmsandwithdifferentnetworkconfigurations,itisfoundthatthemostsuitablenetworkconfigurationis1X7X3X1withtheLMalgorithm.Thismeansthatthenumberofneuronsis7forthefirsthiddenlayerand3forthesecondhiddenlayerrespectively.Theinputandoutputlayershavethelinearactivationfunctionandthehiddenlayershavethehyperbolictangentsigmoidactivationfunction.Thenumberofepochis1000fortraining.Itisimportanttonotethatthecriteriafortoosmallandtoobighiddenlayerneuronnumbersdependonalargenumberoffactors,likeANNtype,trainingsetcharacteristicsandtypeofapplication.Thistopicisstillunderspecialattentionofartificialintelligenceresearcherstoday.4ResultsandConclusionThedevelopedANNmodelsaretrainedandtestedwiththeuseofdifferentlearningalgorithmscalledLM,BR,CGF,RPandBFGStoobtainbetterperformanceandfasterconvergencewithsimplerstructure.Table1showstheerrorsfromthecompletelearningalgorithmsusedintheanalysisforthesamenetworkconfigurationmentionedabove.Whentheperformancesoftheneuralmodelsarecomparedwitheachother,thebestresultforthetrainingandthetestareobtainedfromthemodeltrainedwiththeLMalgorithm.Thetrainingandtesterrors(MSE,meansquareerror)ofthenetworkfortheLMalgorithmare0.7x10-9and1.3x10-4respectively.AsitisclearlyseenfromTable1,thenextsolutionwhichisclosertoLMisobtainedfromBRalgorithm.Amongneuralmodelspresentedhere,theworstresultsareobtainedfromtheRPmethodforthisparticularapplication.Itshouldbeemphasizedthattheaccuracyoflearningalgorithmsingeneraldependsonselectingappropriatelearningparameters,networkconfigurationsandinitializations.Figure3representsthepercentagetesterrorofthenetworktrainedwithLMfortypeEthermocouple.AsitisclearlyseenfromFigure3,themaximumpercentageerrorbecomeslowerthan0.3%.Theaveragepercentageerrorisgreaterthan0.1%fortemperaturesbetween-200and200°C,thereasonbeingthatinthisrangethethermocouplesarestronglynonlinear.However,itisobviousforbestfitintherange-200to200°Cthatthenumberoftrainingdatasetmustbeincreased.Thenormalizederrorconvergencecurvesinthelearningalgorithmsusedintheanalysisfor1000Fig.4.Learning(convergence)characteristicsoftheANNfordifferentlearningalgorithmsusedintheanalysisfor1000epoch.epochsaregraphicallyshowninFigure4.Forsimulatingtheneuralmodel,theANNistrained,tominimizetheMSE.Asthelearningproceeds,themeansquareerrorprogressivelydecreasesandfinallyattainsasteadystateminimumvalueasshowninFigure4.Asaconclusion,atechniqueforhighprecisiontemperaturemeasurementbasedonanANNmodelisproposedinthispaper.ThetrainingprocessforMLPANNsisperformedsuccessfullyinthisstudywiththeuseofLMalgorithmwhichgivesthebestresultamongotherlearningalgorithms.GainandoffseterrorsofthesignalconditioningcircuitareautomaticallycancelledasaconsequenceoftheusageoftheANNtechnique.Theproposedmethodhasalargeareaofapplicationsinallsensorbasedmeasurementsystemswherethesensornonlinearityisthemainfactortobeconsidered.Thetechniquehasapotentialfutureinthefieldofinstrumentationandmeasurement.附錄B高精度溫度測量裝置的設(shè)計1引言傳感器是許多設(shè)備電路中最重要的元素之一。許多工業(yè)都應(yīng)用傳感器，傳感器采用某一形勢的輸入（例如溫度、壓力、幅度等），并轉(zhuǎn)換成能夠解釋的儀器指示數(shù)。在本質(zhì)上許多傳感器都是非線性的，但輸出卻要求是線性的。有許多種傳感器可以進(jìn)行溫度的測量，其中熱電偶的應(yīng)用最廣泛。由于熱電偶具有低成本、寬操作范圍、響應(yīng)迅速、精確度高的優(yōu)點(diǎn)，更適合工業(yè)應(yīng)用。對于溫度，熱電偶也具有非線性輸出。因此，傳感器的仿真和線性化技術(shù)非常必要。為了解決傳感器的線性化問題，提出了兩種方法。第一種方法需要非線性的模擬電路，第二種方法應(yīng)用能夠用微處理器或計算機(jī)計算的數(shù)值方法。模擬電路經(jīng)常被用來提高傳感器的線性特性，但模擬電路需要額外的模擬硬件，還具有對于模擬電路的典型問題，如溫度漂移，增益和補(bǔ)償誤差。應(yīng)用第二種方法，如果一個精確的傳感器是可用的（通過多項(xiàng)式直接計算），或應(yīng)用查詢多維的表格，傳感器的非線性能夠通過操作算法得到補(bǔ)償。多項(xiàng)式方法的直接計算更加精確但需要較長的計算時間，而查表的方法盡管快但并不十分精確。最近幾年，在使用儀器和測量的領(lǐng)域，人工神經(jīng)網(wǎng)絡(luò)作為一種有前景的研究領(lǐng)域已經(jīng)興起。人工神經(jīng)網(wǎng)絡(luò)對于解決復(fù)雜問題特別是非線性系統(tǒng)的仿真提供了神經(jīng)計算方法，而網(wǎng)絡(luò)本身卻是一個非線性系統(tǒng)。當(dāng)測量系統(tǒng)是非線性時包括用來訓(xùn)練的試驗(yàn)數(shù)據(jù)也是非線性時，人工神經(jīng)網(wǎng)絡(luò)是非常有用的。人工神經(jīng)網(wǎng)絡(luò)的最廣泛應(yīng)用是數(shù)值逼近（曲線擬和）。與傳統(tǒng)的數(shù)值插補(bǔ)比較，基于人工神經(jīng)網(wǎng)絡(luò)的插補(bǔ)提供低的插補(bǔ)誤差。在本文中，我們提出了基于人工神經(jīng)網(wǎng)絡(luò)方法的高精度溫度測量系統(tǒng)。校正數(shù)據(jù)是通過在人工神經(jīng)網(wǎng)絡(luò)的訓(xùn)練和測試階段必須具有的調(diào)幅調(diào)頻信號源9100校正單元獲得的。系統(tǒng)的硬件和軟件部分被綜合在用于系統(tǒng)測量和校正的虛擬設(shè)備上。人工神經(jīng)網(wǎng)絡(luò)通過提供一個理想的最終誤差來校正數(shù)據(jù)。這就是按照神經(jīng)元的結(jié)構(gòu)和層數(shù)和神經(jīng)元的數(shù)量通過軟件，將校正和人工神經(jīng)網(wǎng)絡(luò)仿真數(shù)據(jù)之間的均方誤差最小化。系統(tǒng)硬件熱電偶產(chǎn)生一和測量溫度點(diǎn)成比例的在mV數(shù)量級的電壓值，，而在零點(diǎn)的溫度值是一個常數(shù)。為了精確的測量，我們必須知道零點(diǎn)的溫度值。通過人工神經(jīng)網(wǎng)絡(luò)的操作階段，圖1（a）顯示了溫度測量系統(tǒng)模塊。它組成了放置在理想溫度條件下的熱電偶（E型號），包括帶有16位的模擬數(shù)字轉(zhuǎn)換器和接入到計算機(jī)的輸入輸接口卡的信號調(diào)節(jié)電路，設(shè)計的信號調(diào)節(jié)電路具有可設(shè)計增益的放（PGA204BP），它的增益是1，10，100，1000倍，16位的A/D轉(zhuǎn)換器（AD976A），帶有零點(diǎn)溫度補(bǔ)償?shù)腁D595單片電路熱電偶放大器，并把它作為攝氏溫度計的標(biāo)準(zhǔn)，用來選擇熱電偶和攝氏溫度計的輸出的4路模擬多路器（ADG529A）。AD976A具有高速度、低功耗、16位A/D轉(zhuǎn)換器的特點(diǎn)，采用5V工作電壓。這一部分提供連續(xù)的逼近，轉(zhuǎn)換電容ADC，間隔的2.5V干擾和一高速度的平行界面。系統(tǒng)的精度直接依靠ADC每步的大小。具有的輸入，AD976A的LSB是。當(dāng)AD595作為攝氏溫度應(yīng)用時，熱電偶被忽略，微分的輸入被匯集到一起。在這種模式中，AD595產(chǎn)生一個比例因子為10mV/°C，它的輸出應(yīng)用在編寫軟件的零點(diǎn)溫度數(shù)據(jù)中。AD595的一些重要特性如下：電壓范圍是-55°C;到125°C;相對于溫度的穩(wěn)定性是±0.05°C/°C；相對于溫度的敏感性是：10mV/°C。PGA204BP的輸出信號通過AD976A被數(shù)化；AD976A的輸出連接到I/O接口卡并傳遞到個人計算機(jī)上，執(zhí)行數(shù)據(jù)的削減和優(yōu)化。為了設(shè)定人工神經(jīng)網(wǎng)絡(luò)的網(wǎng)絡(luò)權(quán)值和閾值，在校正階段（人工神經(jīng)網(wǎng)絡(luò)的訓(xùn)練階段），在000.000mV到320.000mV的范圍內(nèi)，精確度為±0.006%+4.16μA的調(diào)幅調(diào)頻信號源9100校正單元，連接到模擬多路器的終端，用來產(chǎn)生如圖1（b）所示的熱電偶電壓。此電壓作為人工神經(jīng)網(wǎng)絡(luò)的輸入值，沒有零點(diǎn)溫度補(bǔ)償?shù)臒犭娕茧妷菏侨斯ど窠?jīng)網(wǎng)絡(luò)的輸出值。在操作階段（如圖1a），為了實(shí)現(xiàn)零點(diǎn)溫度補(bǔ)償，我們使用了從攝氏溫度計輸出的數(shù)據(jù)值。人工神經(jīng)網(wǎng)絡(luò)的輸出值隨著由攝氏溫度計獲得的環(huán)境溫度值的變化而變化。然后這個值顯示在虛擬機(jī)上作為熱電偶的溫度值。虛擬機(jī)被用來獲得人工神經(jīng)網(wǎng)絡(luò)的訓(xùn)練過程中的數(shù)據(jù)和顯示操作階段的計算溫度值。圖2顯示了虛擬機(jī)的前端電路板。此設(shè)備的主要特征是：顯示測量的溫度值和從調(diào)理電路輸出的相應(yīng)電壓值，調(diào)理電路是用來收集在校正階段的數(shù)據(jù)值和在操作階段具有零點(diǎn)溫度補(bǔ)償?shù)膶?shí)際電壓值。此系統(tǒng)是由校正和操作階段編寫的軟件來控制的。3人工神經(jīng)網(wǎng)絡(luò)人工神經(jīng)網(wǎng)絡(luò)是基于生物機(jī)理的人腦的模擬。人工神經(jīng)網(wǎng)絡(luò)是前向反饋網(wǎng)絡(luò)和全局逼近網(wǎng)絡(luò)。它是通過經(jīng)驗(yàn)而不是程序來訓(xùn)練和學(xué)習(xí)的。它們是通過簡單的進(jìn)程元素相互連接而形成的，或是可調(diào)節(jié)的閾值相互連接而形成的，閾值組成神經(jīng)的結(jié)構(gòu)組成層。每一個人工神經(jīng)都有閾值輸入，求和單元、功能函數(shù)和輸出。整個人工神經(jīng)網(wǎng)絡(luò)的特性依靠所提及的人工神經(jīng)元、學(xué)習(xí)算法和網(wǎng)絡(luò)的構(gòu)造。在訓(xùn)練的過程中，神經(jīng)元之間的閾值可以通過一些標(biāo)準(zhǔn)（對于所有的訓(xùn)練集目標(biāo)輸出值和測量值之間的均方誤差達(dá)到預(yù)先設(shè)定好的極限值）進(jìn)行調(diào)整，或達(dá)到最大的允許步數(shù)。雖然訓(xùn)練過程非常耗費(fèi)時間，但這能提前完成，并能脫機(jī)運(yùn)行。訓(xùn)練完的神經(jīng)網(wǎng)絡(luò)用來測試數(shù)據(jù)但這一過程是看不見的。機(jī)器語言程序是構(gòu)建神經(jīng)網(wǎng)絡(luò)中最簡單的和應(yīng)用最廣泛的。機(jī)器語言程序由輸入、輸出、事先定義好神經(jīng)元的數(shù)量的一層或多個隱層。輸入層的神經(jīng)元只是作為緩沖器，用來運(yùn)輸輸入信號到神經(jīng)元的隱層。在應(yīng)用分別的連接強(qiáng)度對輸入信號進(jìn)行加權(quán)后，隱層的每一個神經(jīng)元加和他的輸入信號,有求和函數(shù)計算輸出值，可以表示為此式中是人工神經(jīng)網(wǎng)絡(luò)的功能函數(shù)之一。訓(xùn)練的神經(jīng)網(wǎng)絡(luò)由應(yīng)用不同的學(xué)習(xí)算法可調(diào)整的閾值構(gòu)成。學(xué)習(xí)算法給出了在時間t內(nèi)神經(jīng)元i與j之間閾值的連接強(qiáng)度。閾值的遞推公式如下：有許多應(yīng)用學(xué)習(xí)算法的文字描述。在我們的學(xué)習(xí)過程中應(yīng)用的神經(jīng)網(wǎng)絡(luò)算法有LM算法，BFGS算法，BR算法，CGF算法，和RP算法。再此文章中，多層感知器神經(jīng)網(wǎng)絡(luò)作為神經(jīng)網(wǎng)絡(luò)的線性化電路。這涉及的技術(shù)包括應(yīng)用神經(jīng)網(wǎng)絡(luò)估算熱電偶溫度值（神經(jīng)網(wǎng)絡(luò)的輸出值），此時給出的熱電偶輸出電壓作為輸入值。應(yīng)用之前所提及的學(xué)習(xí)算法訓(xùn)練人工神經(jīng)網(wǎng)絡(luò)，進(jìn)而估算應(yīng)用不同的輸入值和相應(yīng)的測量值而訓(xùn)練出的網(wǎng)絡(luò)的溫度輸出值。人工神經(jīng)網(wǎng)絡(luò)的目標(biāo)輸出值和實(shí)際輸出值之間的差別通過學(xué)習(xí)算法悲劇算出來，并應(yīng)用方程式1和2調(diào)整網(wǎng)絡(luò)閾值。我們應(yīng)用從熱電偶數(shù)據(jù)表得到的實(shí)驗(yàn)數(shù)據(jù)進(jìn)行實(shí)驗(yàn)。這些數(shù)據(jù)表是相對于特殊的溫度點(diǎn)而言的（通常是零攝氏度）。人工神經(jīng)網(wǎng)絡(luò)是應(yīng)用熱電偶的80個溫度點(diǎn)進(jìn)行訓(xùn)練的，這80個溫度點(diǎn)分布在-200°C到1000°C的范圍內(nèi)，是在校正階段獲得的。然而，對于可能的輸入值，由訓(xùn)練得到的最終神經(jīng)網(wǎng)絡(luò)的特性并不是無偏差的。在訓(xùn)練之后，我們必須應(yīng)用獨(dú)立的測試系統(tǒng)來估計訓(xùn)練后網(wǎng)絡(luò)的特性。因此，熱電偶的分布在-200°C到1000°C的范圍內(nèi)其他20個數(shù)據(jù)被用來測試。在訓(xùn)練之前，在-1.0到1.0的范圍內(nèi)，輸入與輸出數(shù)據(jù)組被標(biāo)準(zhǔn)化。在采用不同的學(xué)習(xí)算法和和不同的網(wǎng)絡(luò)構(gòu)造后，我們發(fā)現(xiàn)最適合的網(wǎng)絡(luò)構(gòu)造是1X7X3X1。采用LM算法。也就是說第一隱層神經(jīng)元的數(shù)量是7，第二隱層神經(jīng)元的數(shù)量是3。輸入和輸出層采用線性函數(shù)。隱層采用雙曲線正切，對數(shù)的功能函數(shù)。訓(xùn)練的最大步數(shù)是1000步。對于與許多因素有關(guān)的、過大或過小的隱層神經(jīng)元數(shù)量，標(biāo)注這些標(biāo)準(zhǔn)是很重要的，像人工神經(jīng)網(wǎng)絡(luò)的類型，訓(xùn)練數(shù)集的特性和應(yīng)用的類型，如今，人工智能研究更加關(guān)注人工神經(jīng)網(wǎng)絡(luò)。結(jié)果和結(jié)論人工神經(jīng)網(wǎng)絡(luò)模型可以通過不同的學(xué)習(xí)算法如LM算法，BR算法，CGF算法，BP算法，BFGS算法進(jìn)行訓(xùn)練和測試，簡單的結(jié)構(gòu)就可以獲得更高的特性和更快的收斂速度。對于以上所提及的相同的網(wǎng)絡(luò)構(gòu)造，表1顯示了應(yīng)用學(xué)習(xí)算法進(jìn)行分析的誤差。當(dāng)我們將神經(jīng)模型的性質(zhì)相互進(jìn)行比較時，我們發(fā)現(xiàn)應(yīng)用LM算法訓(xùn)練的模型可以獲得最好的實(shí)驗(yàn)結(jié)果。采用LM算法進(jìn)行訓(xùn)練的網(wǎng)絡(luò)他的訓(xùn)練和測試誤差（最小均方誤差）分別為和。從表1我們可以清楚地看到，更接近LM的解可以從BR算法得到。在所提及的神經(jīng)網(wǎng)絡(luò)模型中，對于特殊應(yīng)用的RP的實(shí)驗(yàn)效果最差。他所強(qiáng)調(diào)的學(xué)習(xí)算法的精確度依靠寓所選擇恰當(dāng)?shù)膶W(xué)習(xí)參數(shù)、網(wǎng)絡(luò)的構(gòu)造和初始化值。圖3代表了對于E類型的熱電偶采用LM算法訓(xùn)練網(wǎng)絡(luò)的誤差百比。從圖3可以看出，誤差的最大百分比低于0.3%。在-200°C到200°C之間的誤差百分比大于0.1%，這就是在此溫度范圍內(nèi)熱電偶非線性的原因。然而，200°C~200°C溫度范圍內(nèi)，得到好的擬和曲線顯然要增加訓(xùn)練的數(shù)據(jù)數(shù)量。訓(xùn)練步數(shù)為1000步的學(xué)習(xí)算法仿真神經(jīng)模型，訓(xùn)練人工神經(jīng)網(wǎng)絡(luò)，最小化最小均方誤差。隨著學(xué)習(xí)的進(jìn)程，最小均方誤差逐漸減少，最終穩(wěn)定在某一個最小值如圖4所示。結(jié)論，本論文提出了一種基于人工神經(jīng)網(wǎng)絡(luò)的高精度溫度測量技術(shù)。應(yīng)用LM

論大學(xué)生寫作能力寫作能力是對自己所積累的信息進(jìn)行選擇、提取、加工、改造并將之形成為書面文字的能力。積累是寫作的基礎(chǔ)，積累越厚實(shí)，寫作就越有基礎(chǔ)，文章就能根深葉茂開奇葩。沒有積累，胸?zé)o點(diǎn)墨，怎么也不會寫出作文來的。寫作能力是每個大學(xué)生必須具備的能力。從目前高校整體情況上看，大學(xué)生的寫作能力較為欠缺。一、大學(xué)生應(yīng)用文寫作能力的定義那么，大學(xué)生的寫作能力究竟是指什么呢？葉圣陶先生曾經(jīng)說過，“大學(xué)畢業(yè)生不一定能寫小說詩歌，但是一定要寫工作和生活中實(shí)用的文章，而且非寫得既通順又扎實(shí)不可。”對于大學(xué)生的寫作能力應(yīng)包含什么，可能有多種理解，但從葉圣陶先生的談話中，我認(rèn)為：大學(xué)生寫作能力應(yīng)包括應(yīng)用寫作能力和文學(xué)寫作能力，而前者是必須的，后者是“不一定”要具備，能具備則更好。眾所周知，對于大學(xué)生來說，是要寫畢業(yè)論文的，我認(rèn)為寫作論文的能力可以包含在應(yīng)用寫作能力之中。大學(xué)生寫作能力的體現(xiàn)，也往往是在撰寫畢業(yè)論文中集中體現(xiàn)出來的。本科畢業(yè)論文無論是對于學(xué)生個人還是對于院系和學(xué)校來說，都是十分重要的。如何提高本科畢業(yè)論文的質(zhì)量和水平