李偉平+竇現(xiàn)東+王振興+柳超
文章編號(hào):16742974(2014)05003207
收稿日期:20130719
基金項(xiàng)目:國(guó)家高新技術(shù)研究發(fā)展計(jì)劃(863計(jì)劃)資助項(xiàng)目(2012AA111802)
作者簡(jiǎn)介:李偉平(1971-),男,湖南邵陽(yáng)人,湖南大學(xué)副教授,博士
通訊聯(lián)系人,Email: lwpzlbb@yeah.cn
摘 要: 提出基于誤差反向傳播神經(jīng)網(wǎng)絡(luò)(Back Propagation Neural Network, BPNN)的高維模型表示(high dimensional model representation, HDMR)方法,即BPNNHDMR方法.BPNNHDMR方法的顯著優(yōu)勢(shì)是將BP神經(jīng)網(wǎng)絡(luò)的非線性函數(shù)逼近能力與高維模型的層級(jí)結(jié)構(gòu)理論相結(jié)合來(lái)構(gòu)建近似模型,并能夠揭示輸入變量之間固有的線性或非線性關(guān)系及其相關(guān)性,將構(gòu)造模型復(fù)雜度由指數(shù)級(jí)增長(zhǎng)降階為多項(xiàng)式級(jí),有效地解決了高維建模問(wèn)題.通過(guò)測(cè)試和對(duì)比驗(yàn)證了該算法的效率和建模能力,并將該算法應(yīng)用于礦用自卸車安全駕駛室翻車保護(hù)裝置(RollOver Protective Structure, ROPS)的優(yōu)化設(shè)計(jì).通過(guò)優(yōu)化結(jié)果驗(yàn)證了所提方法的可行性和有效性.
關(guān)鍵詞:近似模型;高維模型;誤差反向傳播神經(jīng)網(wǎng)絡(luò);非線性;結(jié)構(gòu)優(yōu)化
中圖分類號(hào):U463.4 文獻(xiàn)標(biāo)識(shí)碼:A
BPNNHDMR Nonlinear Metamodeling Technique
and Its Application
LI Weiping, DOU Xiandong, WANG Zhenxing, LIU Chao
(State Key Laboratory of Advanced Designed and Manufacture for Vehicle Body,
Hunan Univ, Changsha, Hunan 410082,China)
Abstract:This paper proposed a new highdimension model representation (HDMR) based on back propagation neural network (BPNN), which is called BPNNHDMR. The most remarkable advantage of this method lies in its ability to integrate the nonlinear function approximation capability of BP neural network and the hierarchy structure theory of high dimensional model to build an approximation model. Moreover, this method can reveal the inherent linearity or nonlinearity relationship as well as correlation with respect to input variables. The problem of modeling high dimension model is effectively tackled by reducing the computation cost from exponential growing to polynomial. Testing and comparative analysis confirm the efficiency and capability of BPNNHDMR for high dimension nonlinear problems. Furthermore, the algorithm was applied to optimize the ROPS of Mining Dump Truck's Safety Cab. The optimized results verify the feasibility and effectiveness of the method proposed.
Key words: metamodel; high dimensional model representation(HDMR); back propagation neural network(BPNN); nonlinearity; structural optimization
工程優(yōu)化問(wèn)題中,對(duì)于基于真實(shí)模型的嵌套優(yōu)化,每次計(jì)算目標(biāo)函數(shù)值都要調(diào)用費(fèi)時(shí)的仿真計(jì)算模型,其計(jì)算代價(jià)不可小視.而利用近似模型可以有效地解決這一問(wèn)題[1],即通過(guò)對(duì)近似模型的求優(yōu)近似得到真實(shí)模型的優(yōu)化值.
目前廣泛應(yīng)用的一些近似方法,如響應(yīng)面法、Kriging插值、人工神經(jīng)網(wǎng)絡(luò)等,在處理較低維問(wèn)題時(shí)有很好的效果.而對(duì)于工程中復(fù)雜的高維非線性模型,隨著維數(shù)和非線性程度的增加,構(gòu)造近似模型所需的樣本點(diǎn)數(shù)量和計(jì)算花費(fèi)呈指數(shù)增長(zhǎng),使解決此類問(wèn)題的計(jì)算效率大大降低[2] .針對(duì)這一問(wèn)題,Sobol證明了可積函數(shù)可以分解為不同維數(shù)函數(shù)的疊加理論[3].該理論表明,對(duì)于任意一個(gè)可以積分的函數(shù),在其積分空間內(nèi)存在唯一的、可以擴(kuò)展的高維模型(High Dimensional Model Representation,HDMR).這一模型是精確的,有確定的階數(shù),并包含一個(gè)層級(jí)結(jié)構(gòu).高維模型可以將計(jì)算時(shí)間隨非線性程度和維數(shù)增加按指數(shù)增長(zhǎng)的隱函數(shù),轉(zhuǎn)化為可以忽略高階耦合項(xiàng)的多項(xiàng)式函數(shù),并揭示了每個(gè)設(shè)計(jì)變量對(duì)近似函數(shù)的貢獻(xiàn)量,大大減少了計(jì)算時(shí)間.同時(shí)反映了輸入變量之間固有的線性或非線性關(guān)系及其相關(guān)性,在近似高維非線性問(wèn)題時(shí)非常有效.由此,一系列不同特性的高維模型開(kāi)始發(fā)展起來(lái),被研究和應(yīng)用于不同的領(lǐng)域.其中,Shan等提出了基于徑向基的高維模型(RBFHDMR)[4],湯龍等提出了基于Kriging的高維模型(KrigingHDMR)[5].
本文采用誤差反向傳播神經(jīng)網(wǎng)絡(luò)(Back Propagation Neural Network, BPNN)與中心切面高維模型(CutHDMR[6])方法相結(jié)合的BPNNHDMR.BP神經(jīng)網(wǎng)絡(luò)是一種多層前饋型神經(jīng)網(wǎng)絡(luò),其神經(jīng)元的傳遞是S型函數(shù),它可以實(shí)現(xiàn)從輸入到輸出的任意非線性映射.對(duì)于非線性問(wèn)題,在精度表達(dá)上具有一定優(yōu)勢(shì).CutHDMR方法是用過(guò)指定點(diǎn)的特定的直線、平面和超平面上的信息來(lái)建立模型,計(jì)算效率高,方便易行.
1 高維模型(HDMR)基本理論
設(shè)待求問(wèn)題的設(shè)計(jì)變量可行域?yàn)楠n(n維實(shí)數(shù)空間),那么多元函數(shù)f(x)∈R與輸入變量x∈An之間的映射關(guān)系可以用HDMR[6-7]來(lái)表示為:
f(x)=f0+∑ni=1fi(xi)+∑1≤i
其中f0為函數(shù)在中心點(diǎn)的函數(shù)值,后面依次為不同階耦合項(xiàng)對(duì)近似函數(shù)的貢獻(xiàn)量.
為了方便計(jì)算,本文引入CutHDMR.與其他類型的高維模型相比,CutHDMR用少量簡(jiǎn)單的算術(shù)運(yùn)算來(lái)表達(dá)計(jì)算花費(fèi)高昂的真實(shí)模型,并達(dá)到了其他類型高維模型相似的精度,計(jì)算效率高.CutHDMR展開(kāi)式的各分項(xiàng)表達(dá)請(qǐng)參考文獻(xiàn)[6].
2 BPNNHDMR
2.1 BP神經(jīng)網(wǎng)絡(luò)
BP網(wǎng)絡(luò)是一種多層前饋型神經(jīng)網(wǎng)絡(luò),它由一個(gè)輸入層、一個(gè)輸出層和至少一層隱含層組成[8].該網(wǎng)絡(luò)的主要特點(diǎn)是信號(hào)向前傳遞,誤差反向傳播.在前向傳遞中,輸入信號(hào)從輸入層經(jīng)隱含層逐層處理,直至輸出層.每一層的神經(jīng)元狀態(tài)只影響下一層神經(jīng)元狀態(tài).如果輸出層得不到期望輸出,則轉(zhuǎn)入反向傳播,根據(jù)預(yù)測(cè)誤差調(diào)整網(wǎng)絡(luò)權(quán)值和閾值,從而使BP神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)輸出不斷逼近期望輸出.因此,BP網(wǎng)絡(luò)可以看作是解決函數(shù)逼近的工具.
本文中,隱層神經(jīng)元的傳遞函數(shù)采用雙曲正切S型函數(shù)h,它類似于一個(gè)平滑的階梯函數(shù):
h(x)=(ex-e-x)/(ex+e-x) (2)
考慮一個(gè)分層神經(jīng)網(wǎng)絡(luò)(x,W),其中x是輸入矢量,W是可變權(quán)系數(shù)矢量.在這里要訓(xùn)練(x,W)來(lái)近似函數(shù)f:[a,b]R→R,其中f([a,b])是一個(gè)有界集.在區(qū)間[a, b]中隨機(jī)選取xk作為神經(jīng)網(wǎng)絡(luò)的輸入,(xk,Wk)為網(wǎng)絡(luò)的輸出,f(xk)為在區(qū)間f([a,b])中的期望輸出值.神經(jīng)網(wǎng)絡(luò)的任務(wù)是調(diào)整網(wǎng)絡(luò)中的可變權(quán)系數(shù)來(lái)減小誤差Ek,Ek定義為:
Ek=12[xk,W(k)]-f(xk)2(3)
假設(shè)wi(k)為W(k)中的任意元素.調(diào)整wi(k)對(duì)誤差Ek的影響直接取決于偏導(dǎo)數(shù)(Ek)/[wi(k)].BP算法的作用在于如何評(píng)價(jià)(Ek)/[wi(k)].詳細(xì)內(nèi)容可參考文獻(xiàn)[9].這樣,wi(k)就向著減小誤差Ek的方向調(diào)整:
wi(k+1)=wi(k)-μ(Ek)/[wi(k)],
wi(k)∈W(k)(4)
其中μ為常量,是指定的更新率.可以看出,如果μ足夠小時(shí),則:
{[xk,W(k+1)]-f(xk)}2<
{[xk,W(k)]-f(xk)}2(5)
其中假設(shè)Ek>0.由式(5)可知,對(duì)于同樣的輸入?yún)?shù)x1,x2,…,xk,更新后的神經(jīng)網(wǎng)絡(luò)的輸出更接近于真實(shí)函數(shù)輸出值.
HechtNielsen[9]研究了分層神經(jīng)網(wǎng)絡(luò)近似非線性函數(shù)的能力.他在文獻(xiàn)[9]中表明,在確定的條件下,對(duì)于任意ε>0,存在一個(gè)三層神經(jīng)網(wǎng)絡(luò)(包含一個(gè)輸入層,一個(gè)輸出層和一個(gè)非線性隱含層)可以在均方誤差精確度ε范圍內(nèi)近似函數(shù)f.因此,BP神經(jīng)網(wǎng)絡(luò)可以應(yīng)用于工程實(shí)際中的大多數(shù)情況.
2.2 BPNNHDMR的構(gòu)建
對(duì)大多數(shù)工程問(wèn)題而言,非耦合項(xiàng)和低階耦合項(xiàng)對(duì)響應(yīng)函數(shù)影響較大.為此,BPNNHDMR方法只考慮到一階耦合項(xiàng),表達(dá)式如下:
f(x)≈f0+∑ni=1i(xi)+∑1≤i 式中帶頂標(biāo)^的項(xiàng)表示BPNN近似項(xiàng),∑ni=1i(xi),∑1≤i 高維模型的一般構(gòu)建流程[4-5]如下: 1)選取各設(shè)計(jì)變量中心位置的點(diǎn)x0=[x10,x20,…,xn0]作為中心點(diǎn),計(jì)算得到f0. 2)在每個(gè)變量xi([x10,x20,…,xi,…,xn0])的取值區(qū)間上為非耦合項(xiàng)fi(xi)=f([x10,…,xi,…,xn0]T)-f0布點(diǎn).選取xi取值區(qū)間上下界的兩個(gè)端點(diǎn),并計(jì)算這兩點(diǎn)的函數(shù)值,用BPNN構(gòu)建近似函數(shù)i(xi). 3)檢驗(yàn)i(xi)的線性關(guān)系,如果(x0)-f0f0≤10-3,則認(rèn)為i(xi)是線性的,程序終止;否則,繼續(xù)采樣和xi上下界三個(gè)點(diǎn)構(gòu)建一個(gè)新的i(xi),看i(xi)是否滿足給定的精度要求,如果滿足,則程序終止;否則,繼續(xù)采樣構(gòu)建一個(gè)新的近似函數(shù)i(xi),直至滿足精度要求.循環(huán)執(zhí)行第2步和第3步直到所有的非耦合項(xiàng)構(gòu)造完畢. 4)判斷模型中是否存在一階耦合項(xiàng).創(chuàng)建新檢驗(yàn)點(diǎn)(xi,xj,xij0)=[x10,x20,…,xi,…,xj,…,xn0],不失一般性,選取構(gòu)造非耦合項(xiàng)時(shí)用的采樣點(diǎn)分量的上下界中的一個(gè)作為新樣本的第i維分量.在精確度準(zhǔn)則所允許的誤差范圍內(nèi),若f0+∑ni=1fi(xi)=f0+∑ni=1i(xi),就認(rèn)為不存在一階耦合項(xiàng),程序終止;否則轉(zhuǎn)入第5步. 5)識(shí)別相互耦合的變量.構(gòu)造新樣本點(diǎn)[x10,…,xei,…,xej,…,xn0]T(1≤i 6)將上面構(gòu)建的各階BPNN近似函數(shù)代入式(6)就得到了高維模型近似響應(yīng)函數(shù). 在用BP神經(jīng)網(wǎng)絡(luò)構(gòu)建近似模型過(guò)程中,一般隱含層結(jié)點(diǎn)數(shù)取2P+1[10] (P為輸入結(jié)點(diǎn)數(shù)),由于本文樣本數(shù)據(jù)準(zhǔn)確,對(duì)于非耦合項(xiàng)和一階耦合項(xiàng)我們選擇的隱結(jié)點(diǎn)數(shù)分別為3和5.根據(jù)文獻(xiàn)[11],只有當(dāng)學(xué)習(xí)率為0<η<43N+1時(shí),神經(jīng)網(wǎng)絡(luò)算法是收斂的,這里η為學(xué)習(xí)率,N為神經(jīng)網(wǎng)絡(luò)訓(xùn)練樣本點(diǎn)數(shù).同時(shí),本文中神經(jīng)網(wǎng)絡(luò)最大訓(xùn)練步數(shù)設(shè)定為100次,訓(xùn)練目標(biāo)為0.000 04,單隱含層.其他更多BPNN具體參數(shù)設(shè)置和程序設(shè)計(jì)請(qǐng)參看文獻(xiàn)[10]和[12]. 本文高維模型構(gòu)建時(shí)的精確度準(zhǔn)則和收斂準(zhǔn)則都是通過(guò)相對(duì)誤差來(lái)定義的.精確度準(zhǔn)則主要用于輸入變量之間耦合性的識(shí)別,所允許的相對(duì)誤差一般不超過(guò)10-4.由于本文方法具有較高精確度,所以本文針對(duì)測(cè)試函數(shù)收斂準(zhǔn)則所允許的相對(duì)誤差均取0.001;而對(duì)于工程問(wèn)題,考慮到計(jì)算效率和一般工程要求的5%的近似精度,設(shè)置的相對(duì)誤差均取0.01.通過(guò)以上設(shè)置,在樣本點(diǎn)數(shù)一定時(shí),由于本文方法將BP神經(jīng)網(wǎng)絡(luò)的非線性函數(shù)逼近能力與高維模型的層級(jí)結(jié)構(gòu)理論相結(jié)合,所以本文近似模型方法在近似非線性問(wèn)題時(shí)較傳統(tǒng)方法具有更高的精度,建模效率也有所提高. 3 數(shù)值算例 3.1 評(píng)價(jià)指標(biāo) 為了測(cè)試近似模型近似效果,采用3種比較常用的評(píng)價(jià)指標(biāo),R2(R square),相對(duì)平均絕對(duì)誤差(relative average absolute error, RAAE)和相對(duì)最大絕對(duì)誤差(relative maximum absolute error, RMAE).這些指標(biāo)反映了BPNNHDMR在新樣本點(diǎn)上的預(yù)測(cè)能力,其具體表達(dá)式請(qǐng)參考文獻(xiàn)[4].其中,R2是從整體上反映近似模型的精度,其值越接近1越好;RAAE也是從整體上反映近似模型的精度,其值越小越好;RMAE是一個(gè)局部指標(biāo),描述了設(shè)計(jì)空間中某個(gè)局部區(qū)域的誤差,其值越小越好.
3.2 函數(shù)算例
首先選擇一個(gè)高維非線性測(cè)試函數(shù):
f(x)=∑9i=1[(x2i)(x2i+1+1)+(x2i+1)(x2i+1)],
xi∈[0,1](7)
采用相同數(shù)量的訓(xùn)練樣本點(diǎn)(計(jì)算費(fèi)用相同),分別采用BPNN,BPNNHDMR和Kriging 3種方法進(jìn)行建模并比較它們的精度,計(jì)算結(jié)果如表1所示(表中數(shù)據(jù)是計(jì)算100次的平均值).
表1 BPNN,BPNNHDMR和Kriging方法比較
Tab.1 Comparison between BPNN, BPNNHDMR
and Kriging algorithm
方法
R2
RAAE
RMAE
BPNN
0.886 0
0.221 6
0.225 1
BPNNHDMR
0.993 9
0.058 2
0.049 5
Kriging
0.326 4
2.975 8
0.836 6
分析表中數(shù)據(jù),對(duì)于同一測(cè)試函數(shù),3種近似方法中,BPNNHDMR方法的R2值最接近1,RAAE和RMAE也都最小.由對(duì)比可知,對(duì)于高維非線性問(wèn)題,基于同樣數(shù)量的一組訓(xùn)練樣本,采用BPNNHDMR方法得到的近似模型精度更高.
其次,增加高維非線性測(cè)試函數(shù)式(7)的維數(shù)
f(x)=∑d-1i=1[(x2i)(x2i+1+1)+(x2i+1)(x2i+1)],
0≤xi≤1(8)
來(lái)測(cè)試BPNNHDMR方法的效率.式(8)中,維數(shù)d分別取d=10,30,50等不同值,假設(shè)每一維的訓(xùn)練樣本點(diǎn)數(shù)為9(經(jīng)測(cè)試,基本可以滿足精度要求),表2中列出了各階HDMR計(jì)算費(fèi)用的比較.
表2 各階HDMR建模費(fèi)用的比較
Tab.2 Comparison of modeling cost among variousorder
HDMRs for the study problem
維數(shù)
BPNNHDMR
高維模型二階項(xiàng)全部
展開(kāi)(多項(xiàng)式級(jí)增長(zhǎng))
全因子設(shè)計(jì)sd
(指數(shù)級(jí)增長(zhǎng))
10
172
2 961
3.486 8×109
30
532
28 081
4.239 1×1028
50
892
78 801
5.153 8×1047
4 工程應(yīng)用
為了驗(yàn)證BPNNHDMR在處理工程實(shí)際問(wèn)題時(shí)的可行性,本文以某礦用自卸車全駕駛室式翻車保護(hù)裝置(ROPS)的優(yōu)化設(shè)計(jì)為例.根據(jù)標(biāo)準(zhǔn)ISO 3471:2008[13]對(duì)ROPS進(jìn)行非線性有限元分析,確保ROPS滿足標(biāo)準(zhǔn)ISO 3164:1992[14].通過(guò)仿真分析得到ROPS的變形量,為進(jìn)一步優(yōu)化打下基礎(chǔ).安全駕駛室ROPS的有限元模型如圖1所示.
圖1 ROPS加載分析模型
Fig.1 Loads and constraints of the ROPS model
ROPS框架采用殼單元模擬,焊接使用剛性殼單元模擬為一個(gè)載荷變換器[15].焊接連接處網(wǎng)格局部加密,計(jì)算時(shí)假設(shè)焊縫與母材材料特性相同(實(shí)際中焊縫一般不先破壞).單元大小設(shè)定為20 mm,其中四邊形單元54 951個(gè),三角形單元1 389個(gè),單元類型分別為S4R和S3R.本文ROPS材料為進(jìn)口鋼板A710,屈服極限為690 MPa,斷裂極限為792 MPa,材料模型選擇金屬塑性材料模型,定義材料的應(yīng)力應(yīng)變曲線.依據(jù)標(biāo)準(zhǔn)ISO 3471:2008進(jìn)行載荷和約束的施加.其中,約束ROPS底部與車架連接處3個(gè)平動(dòng)方向的自由度(UX,UY,UZ),F(xiàn)C為側(cè)向力加載,FV為垂向力加載,FL為縱向力加載.標(biāo)準(zhǔn)撓曲極限量(Deflectionlimiting volume ,DLV)與ROPS的側(cè)向間距為200 mm,垂向間距為110 mm,縱向間距為320 mm.分析時(shí)考慮幾何非線性和材料非線性.
4.1 優(yōu)化變量的選取
本文優(yōu)化的ROPS主要由一些不同厚度的矩形管和加強(qiáng)筋焊接而成.考慮對(duì)稱性和加工可行性,對(duì)稱的選取ROPS不同部位的矩形管厚度作為設(shè)
計(jì)變量.各設(shè)計(jì)變量選取見(jiàn)圖1,取值范圍如表3所示.
表3 各優(yōu)化變量取值范圍
Tab.3 Range of optimization variablesmm
參數(shù)
x1
x2
x3
x4
x5
x6
x7
x8
x9
x10
取值范圍
[8,12]
[6,14]
[8,12]
[6,14]
[6,10]
[6,10]
[5,7]
[5,7]
[6,10]
[6,10]
4.2 目標(biāo)函數(shù)的建立
通過(guò)非線性有限元分析可知,由于要考慮ROPS側(cè)向吸能問(wèn)題,ROPS側(cè)向變形最接近DLV,因此側(cè)向變形為危險(xiǎn)工況.根據(jù)企業(yè)要求,本文選取側(cè)向變形作為一個(gè)優(yōu)化目標(biāo),以垂向和縱向變形為約束.因此以ROPS側(cè)向變形最小和ROPS質(zhì)量最小建立多目標(biāo)優(yōu)化問(wèn)題,ROPS的多目標(biāo)優(yōu)化模型可以描述為:
min XUC(X)min Xm(X)s.t.UV(X)
式(9)中UC為ROPS的側(cè)向變形量;UV為ROPS的垂直變形量;UL為ROPS的縱向變形量;B1,B2分別為ROPS垂向和縱向的最大允許變形量;m為ROPS的總質(zhì)量.
4.3 近似模型的構(gòu)建
本文采用BPNNHDMR構(gòu)建ROPS變形量和質(zhì)量響應(yīng)近似模型.對(duì)每個(gè)設(shè)計(jì)變量在其取值范圍內(nèi)等間距取點(diǎn),取點(diǎn)數(shù)量根據(jù)構(gòu)建近似模型時(shí)是否達(dá)到預(yù)先設(shè)定的精度決定,通過(guò)ABAQUS計(jì)算獲得樣本點(diǎn)處的真實(shí)響應(yīng)值.首先構(gòu)建一階項(xiàng),然后進(jìn)行精度判斷.在設(shè)計(jì)變量空間,采用拉丁超立方實(shí)驗(yàn)設(shè)計(jì)方法,選擇10個(gè)采樣點(diǎn).分別對(duì)真實(shí)模型和近似模型進(jìn)行計(jì)算,并計(jì)算兩者的相對(duì)誤差.如果相對(duì)誤差滿足給定精度要求(設(shè)定為5%),則構(gòu)建近似模型成功,若精度不滿足要求,則繼續(xù)構(gòu)建高階項(xiàng).
通過(guò)計(jì)算,本文所求近似模型只需計(jì)算到一階項(xiàng),就可以得到比較精確的近似結(jié)果,可以省略高階耦合項(xiàng).構(gòu)建過(guò)程中只需56個(gè)樣本點(diǎn)即得到滿足工程要求的近似模型.
為了客觀、全面地反映所建近似模型在設(shè)計(jì)域內(nèi)的精確程度,在設(shè)計(jì)域內(nèi)采用拉丁超立方實(shí)驗(yàn)設(shè)計(jì)方法隨機(jī)生成10個(gè)測(cè)試樣本點(diǎn),用它們分別對(duì)BPNN,BPNNHDMR和Kriging模型進(jìn)行精度測(cè)試,并以側(cè)向變形響應(yīng)精度為例,測(cè)試結(jié)果如表4所示.
表4 采用BPNN,BPNNHDMR和Kriging
方法所得側(cè)向變形近似模型的比較
Tab.4 Comparison of the lateral deformation between
BPNN, BPNNHDMR and Kriging algorithm
方法
樣本點(diǎn)數(shù)
R2
RAAE
BPNN
56
0.603 3
0.298 9
BPNNHDMR
56
0.984 0
0.085 1
Kriging
56
0.238 3
0.716 7
由表4中數(shù)據(jù)可知,BPNNHDMR解決此類工程問(wèn)題具有高效準(zhǔn)確的優(yōu)點(diǎn).同時(shí),可以看出BPNN和Kriging方法在構(gòu)建高維非線性有限元模型時(shí)的局限性,本文需要增加采樣點(diǎn)數(shù)量才能用這兩種方法獲得比較精確的近似模型,這樣就增加了真實(shí)有限元模型的計(jì)算次數(shù),而真實(shí)非線性有限元模型一次計(jì)算要數(shù)小時(shí),導(dǎo)致計(jì)算花費(fèi)大大增加.
4.4 優(yōu)化過(guò)程及結(jié)果分析
本文采用基于Pareto概念的多目標(biāo)優(yōu)化遺傳算法,該方法是求解多目標(biāo)問(wèn)題非劣最優(yōu)解的有效途徑之一 [16].在基于Pareto最優(yōu)概念的遺傳算法中,NSGAⅡ[17-18]是最有效的.因此,本文采用NSGAⅡ,在近似模型的基礎(chǔ)上對(duì)ROPS進(jìn)行多目標(biāo)優(yōu)化.
初始種群設(shè)為100,最大迭代次數(shù)設(shè)為200,交叉概率0.9,變異概率0.1.收斂規(guī)則為:達(dá)到最大迭代次數(shù)作為終止條件.得到Pareto最優(yōu)解集如圖2所示.
圖2中星號(hào)表示多目標(biāo)優(yōu)化非劣解.在優(yōu)化解集中取有代表性的10組解,多目標(biāo)優(yōu)化的Pareto最優(yōu)解集見(jiàn)表5.ROPS優(yōu)化前側(cè)向變形和質(zhì)量如表6所示.對(duì)比表6與表5中第4組和第6組數(shù)據(jù)可以看出,在質(zhì)量相當(dāng)?shù)那闆r下,優(yōu)化后的側(cè)向變形大約減小了26.7 mm,在側(cè)向變形相當(dāng)?shù)那闆r下,優(yōu)化后的質(zhì)量大約減小了75.1 kg.因此,通過(guò)多目標(biāo)優(yōu)化,ROPS的質(zhì)量和變形情況明顯改善.同時(shí),根據(jù)得到的Pareto最優(yōu)解集,可以根據(jù)設(shè)計(jì)者經(jīng)驗(yàn)和需求,高效率地實(shí)現(xiàn)ROPS各矩形管厚度的選取,以滿足不同性能需要.
把優(yōu)化結(jié)果代入有限元計(jì)算模型驗(yàn)證,計(jì)算結(jié)果如表7所示.
由ABAQUS計(jì)算驗(yàn)證可知,所選優(yōu)化解都符合標(biāo)準(zhǔn)ISO 3164:1992的要求.這也證明了本文所提方法在工程優(yōu)化實(shí)際應(yīng)用中的可行性和有效性.同時(shí),由于模型簡(jiǎn)化掉了一些蒙皮和附屬部件,以及真實(shí)情況下車架也有一部分吸能作用,所以本文的分析結(jié)果是偏向安全的.
側(cè)向變形/mm
圖2 多目標(biāo)優(yōu)化的Pareto最優(yōu)解集
Fig.2 Feasible Pareto optimal solutions
表5 雙目標(biāo)優(yōu)化的Pareto最優(yōu)解
Tab.5 The Pareto optimal solutions of twoobjective uncertain optimization
Pareto解
x1
/mm
x2
/mm
x3
/mm
x4
/mm
x5
/mm
x6
/mm
x7
/mm
x8
/mm
x9
/mm
x10
/mm
質(zhì)量
/kg
側(cè)向變形
/mm
1
12.0
6.0
12.0
13.2
10.0
6.7
9.9
5.8
7.0
9.8
1 508.5
88.0
2
11.9
6.0
11.5
10.3
10.0
6.6
9.8
5.8
6.8
9.9
1 453.1
94.5
3
11.8
6.0
10.7
10.6
10.0
6.5
9.6
5.9
6.6
9.9
1 421.2
100.9
4
12.0
8.4
8.5
10.1
10.0
6.9
7.8
6.4
5.0
9.9
1 347.9
116.4
5
11.0
8.5
8.0
9.6
9.3
6.4
7.5
6.4
5.0
9.5
1 284.3
136.5
6
9.9
6.0
9.4
8.6
9.5
7.0
9.0
6.0
5.0
7.5
1 269.8
143.9
7
9.6
6.2
8.8
6.3
9.3
7.7
6.4
5.3
5.4
7.1
1 215.1
152.6
8
10.3
6.0
8.9
6.8
7.6
6.5
6.9
5.3
5.1
8.3
1 188.2
164.2
9
10.0
6.1
8.7
6.0
7.5
6.5
7.0
5.6
5.2
8.6
1 168.5
171.5
10
10.0
6.0
8.0
6.3
6.7
6.5
7.2
5.3
5.0
8.3
1 125.0
186.3
表6 ROPS側(cè)向變形優(yōu)化前數(shù)據(jù)
Tab.6 Finite element analysis results of ROPS before optimized
變量
x1
/mm
x2
/mm
x3
/mm
x4
/mm
x5
/mm
x6
/mm
x7
/mm
x8
/mm
x9
/mm
x10
/mm
質(zhì)量
/kg
側(cè)向變形
/mm
初始值
10
10
10
10
8
8
8
6
6
8
1 344.9
143.1
表7 優(yōu)化后的響應(yīng)值A(chǔ)BAQUS計(jì)算驗(yàn)證
Tab.7 Check the optimization results
by ABAQUS calculation mm
優(yōu)化后最大
側(cè)向變形量
ABAQUS
計(jì)算的側(cè)向
最大變形
ABAQUS
計(jì)算的垂向
最大變形
ABAQUS
計(jì)算的縱向
最大變形
解一
88.0
92.22
35.70
33.64
解二
94.5
98.99
34.77
35.85
解三
100.9
102.9
36.47
39.74
解四
116.4
112.8
46.09
51.62
解五
136.5
134.1
60.10
68.26
解六
143.9
142.9
62.63
74.29
解七
152.6
159.0
73.41
88.32
解八
164.2
167.1
69.32
82.35
解九
171.5
174.0
70.30
85.04
解十
186.3
192.1
78.05
100.6
5 結(jié) 論
本文提出的BPNNHDMR建模方法,很好地利用了BP神經(jīng)網(wǎng)絡(luò)的非線性函數(shù)逼近能力與高維模型的層級(jí)結(jié)構(gòu)理論,并能夠反映輸入變量之間固有的線性關(guān)系和耦合性.尤其對(duì)于高維問(wèn)題,它可以將構(gòu)造模型計(jì)算花費(fèi)由維數(shù)的指數(shù)級(jí)增長(zhǎng)降解為多項(xiàng)式級(jí),有效地解決了高維建模問(wèn)題,且與傳統(tǒng)算法相比,具有更高的精度.通過(guò)數(shù)值算例和工程優(yōu)化問(wèn)題的對(duì)比,驗(yàn)證了BPNNHDMR近似方法的精度和效率.另外,該算法對(duì)于高度非線性問(wèn)題精確的數(shù)學(xué)建模還需進(jìn)一步發(fā)展和完善,BP神經(jīng)網(wǎng)絡(luò)的計(jì)算效率也有待進(jìn)一步提高.
同時(shí),本文所采用的結(jié)構(gòu)分析與優(yōu)化方法在工程領(lǐng)域的其他方面也可以廣泛應(yīng)用,具有一定的理論和工程實(shí)際意義.
參考文獻(xiàn)
[1] 穆雪峰. 多學(xué)科設(shè)計(jì)優(yōu)化代理模型技術(shù)的研究和應(yīng)用[D].南京:南京航空航天大學(xué)航空宇航學(xué)院,2004.
MU Xuefeng. The research and application of multidiscipline design optimization surrogate model technology[D]. Nanjing: College of Aerospace Engineering, Nanjing University of Aeronautics and Astronautics, 2004.(In Chinese)
[2] 趙子衡,韓旭,姜潮.基于近似模型的非線性區(qū)間數(shù)優(yōu)化方法及其應(yīng)用[J].計(jì)算力學(xué)學(xué)報(bào),2010,27(3):451-456.
ZHAO Ziheng, HAN Xu, JIANG Chao. Approximation model based nonlinear interval number optimization method and its application [J]. Chinese Journal of Computational Mechanics, 2010, 27(3):451-456. (In Chinese)
[3] SOBOL I M. Sensitivity estimates for nonlinear mathematical models [J]. Mathematical Modeling & Computational Experiment, 1993, 1(4): 407-414.
[4] SHAN Songqing,WANG G Gary. Metamodeling for high dimensional simulationbased design problems[J]. Journal of Mechanical Design, 2010, 132:051009.
[5] 湯龍,李光耀,王琥. KrigingHDMR非線性近似模型方法[J].力學(xué)學(xué)報(bào),2011,43(4):780-784.
TANG Long, LI Guangyao, WANG Hu. KrigingHDMR metamodeling technique for nonlinear problems [J]. Chinese Journal of Theoretical and Applied Mechanics, 2011, 43(4):780-784. (In Chinese)
[6] RABITZ H, AL1S F. General foundations of highdimensional model representations[J]. Journal of Mathematical Chemistry,1999,25(2/3):197-233.
[7] LI G, WANG S W, ROSENTHAL C, et al. High dimensional model representations generated from low dimensional data samples. I. MpCutHDMR[J]. Journal of Mathematical Chemistry, 2001,30(1): 1-30.
[8] CHEN Fuchang .Backpropagation neural networks for nonlinear self tuning adaptive control[J]. Control Systems Magazine, IEEE April, 1990, 10(3): 44-48.
[9] HECHTNIELSEN R.Theory of the backpropagation neural network[C]//Proc IEEE 2002 Intl Conf Neural Networks.Califormia:International Toint Conference in Nemal Networks, 2002.
[10]謝慶生,尹健,羅延科,等.機(jī)械工程中的神經(jīng)網(wǎng)絡(luò)方法[M].北京:機(jī)械工業(yè)出版社,2003: 39-51.
XIE Qingsheng, YIN Jian, LUO Yanke,et al. Neural network method in mechanical engineering [M].Beijing:China Machine Press, 2003:39-51. (In Chinese)
[11]曾喆昭,李仁發(fā).高階多通帶濾波器優(yōu)化設(shè)計(jì)研究[J].電子學(xué)報(bào),2002,30(1):87-89.
ZENG Zhezhao, LI Renfa. Study on the optimum design of the highorder multibandpass filters [J].Acta Electronica Sinica, 2002, 30(1):87-89. (In Chinese)
[12]張德豐. MATLAB 神經(jīng)網(wǎng)絡(luò)應(yīng)用設(shè)計(jì)[M]. 北京:機(jī)械工業(yè)出版社,2012:49-75.
ZHANG Defeng. MATLAB neural network application design [M]. Beijing:China Machine Press, 2012:49-75. (In Chinese)
[13]ISO 3471:2008 Earthmoving machineryrollover protective structureslaboratory tests and performance requirements[S]. British: Standards Policy and Strategy Committee,2009.
[14]ISO 3164: 1992Earthmoving machinery —Laboratory evaluation of rollover and fallingobject protective structuresspecifications for deflectionlimiting volume[S]. British: The Authority of the Standards Board, 1992.
[15]周傳月. MSC.Fatigue疲勞分析應(yīng)用與實(shí)例[M]. 北京:科學(xué)出版社,2005:86-91.
ZHOU Chuanyue. Fatigue analysis of application and examples with MSC.Fatigue[M]. Beijing:Science Press, 2005:86-91. (In Chinese)
[16]謝濤,陳火旺,康立山.多目標(biāo)優(yōu)化的演化算法[J].計(jì)算機(jī)學(xué)報(bào),2003,26(8):997-1003.
XIE Tao, CHEN Huowang, KANG Lishan. Evolutionary algorithms of multiobjective optimization problems [J].Chinese Journal of Computers, 2003, 26(8):997-1003. (In Chinese)
[17]DEB K, PRATAP A, AGARWAL S, et al. A fast elitist nondominated sorting genetic algorithm for multiobjective optimization: NSGAII,Kan GAL Report No.200001[R]. Kanpur: Indian Istitute of Technology, 2000.
[18]李偉平,王世東,周兵,等. 基于響應(yīng)面法和NSGAⅡ算法的麥弗遜懸架優(yōu)化[J]. 湖南大學(xué)學(xué)報(bào):自然科學(xué)版, 2011,38(6):27-32.
LI Weiping, WANG Shidong, ZHOU Bin, et al. Macpherson suspension parameter optimization based on response surface method and NSGAⅡ algorithm [J]. Journal of Hunan University:Natural Sciences, 2011,38(6):27-32. (In Chinese)
同時(shí),本文所采用的結(jié)構(gòu)分析與優(yōu)化方法在工程領(lǐng)域的其他方面也可以廣泛應(yīng)用,具有一定的理論和工程實(shí)際意義.
參考文獻(xiàn)
[1] 穆雪峰. 多學(xué)科設(shè)計(jì)優(yōu)化代理模型技術(shù)的研究和應(yīng)用[D].南京:南京航空航天大學(xué)航空宇航學(xué)院,2004.
MU Xuefeng. The research and application of multidiscipline design optimization surrogate model technology[D]. Nanjing: College of Aerospace Engineering, Nanjing University of Aeronautics and Astronautics, 2004.(In Chinese)
[2] 趙子衡,韓旭,姜潮.基于近似模型的非線性區(qū)間數(shù)優(yōu)化方法及其應(yīng)用[J].計(jì)算力學(xué)學(xué)報(bào),2010,27(3):451-456.
ZHAO Ziheng, HAN Xu, JIANG Chao. Approximation model based nonlinear interval number optimization method and its application [J]. Chinese Journal of Computational Mechanics, 2010, 27(3):451-456. (In Chinese)
[3] SOBOL I M. Sensitivity estimates for nonlinear mathematical models [J]. Mathematical Modeling & Computational Experiment, 1993, 1(4): 407-414.
[4] SHAN Songqing,WANG G Gary. Metamodeling for high dimensional simulationbased design problems[J]. Journal of Mechanical Design, 2010, 132:051009.
[5] 湯龍,李光耀,王琥. KrigingHDMR非線性近似模型方法[J].力學(xué)學(xué)報(bào),2011,43(4):780-784.
TANG Long, LI Guangyao, WANG Hu. KrigingHDMR metamodeling technique for nonlinear problems [J]. Chinese Journal of Theoretical and Applied Mechanics, 2011, 43(4):780-784. (In Chinese)
[6] RABITZ H, AL1S F. General foundations of highdimensional model representations[J]. Journal of Mathematical Chemistry,1999,25(2/3):197-233.
[7] LI G, WANG S W, ROSENTHAL C, et al. High dimensional model representations generated from low dimensional data samples. I. MpCutHDMR[J]. Journal of Mathematical Chemistry, 2001,30(1): 1-30.
[8] CHEN Fuchang .Backpropagation neural networks for nonlinear self tuning adaptive control[J]. Control Systems Magazine, IEEE April, 1990, 10(3): 44-48.
[9] HECHTNIELSEN R.Theory of the backpropagation neural network[C]//Proc IEEE 2002 Intl Conf Neural Networks.Califormia:International Toint Conference in Nemal Networks, 2002.
[10]謝慶生,尹健,羅延科,等.機(jī)械工程中的神經(jīng)網(wǎng)絡(luò)方法[M].北京:機(jī)械工業(yè)出版社,2003: 39-51.
XIE Qingsheng, YIN Jian, LUO Yanke,et al. Neural network method in mechanical engineering [M].Beijing:China Machine Press, 2003:39-51. (In Chinese)
[11]曾喆昭,李仁發(fā).高階多通帶濾波器優(yōu)化設(shè)計(jì)研究[J].電子學(xué)報(bào),2002,30(1):87-89.
ZENG Zhezhao, LI Renfa. Study on the optimum design of the highorder multibandpass filters [J].Acta Electronica Sinica, 2002, 30(1):87-89. (In Chinese)
[12]張德豐. MATLAB 神經(jīng)網(wǎng)絡(luò)應(yīng)用設(shè)計(jì)[M]. 北京:機(jī)械工業(yè)出版社,2012:49-75.
ZHANG Defeng. MATLAB neural network application design [M]. Beijing:China Machine Press, 2012:49-75. (In Chinese)
[13]ISO 3471:2008 Earthmoving machineryrollover protective structureslaboratory tests and performance requirements[S]. British: Standards Policy and Strategy Committee,2009.
[14]ISO 3164: 1992Earthmoving machinery —Laboratory evaluation of rollover and fallingobject protective structuresspecifications for deflectionlimiting volume[S]. British: The Authority of the Standards Board, 1992.
[15]周傳月. MSC.Fatigue疲勞分析應(yīng)用與實(shí)例[M]. 北京:科學(xué)出版社,2005:86-91.
ZHOU Chuanyue. Fatigue analysis of application and examples with MSC.Fatigue[M]. Beijing:Science Press, 2005:86-91. (In Chinese)
[16]謝濤,陳火旺,康立山.多目標(biāo)優(yōu)化的演化算法[J].計(jì)算機(jī)學(xué)報(bào),2003,26(8):997-1003.
XIE Tao, CHEN Huowang, KANG Lishan. Evolutionary algorithms of multiobjective optimization problems [J].Chinese Journal of Computers, 2003, 26(8):997-1003. (In Chinese)
[17]DEB K, PRATAP A, AGARWAL S, et al. A fast elitist nondominated sorting genetic algorithm for multiobjective optimization: NSGAII,Kan GAL Report No.200001[R]. Kanpur: Indian Istitute of Technology, 2000.
[18]李偉平,王世東,周兵,等. 基于響應(yīng)面法和NSGAⅡ算法的麥弗遜懸架優(yōu)化[J]. 湖南大學(xué)學(xué)報(bào):自然科學(xué)版, 2011,38(6):27-32.
LI Weiping, WANG Shidong, ZHOU Bin, et al. Macpherson suspension parameter optimization based on response surface method and NSGAⅡ algorithm [J]. Journal of Hunan University:Natural Sciences, 2011,38(6):27-32. (In Chinese)
同時(shí),本文所采用的結(jié)構(gòu)分析與優(yōu)化方法在工程領(lǐng)域的其他方面也可以廣泛應(yīng)用,具有一定的理論和工程實(shí)際意義.
參考文獻(xiàn)
[1] 穆雪峰. 多學(xué)科設(shè)計(jì)優(yōu)化代理模型技術(shù)的研究和應(yīng)用[D].南京:南京航空航天大學(xué)航空宇航學(xué)院,2004.
MU Xuefeng. The research and application of multidiscipline design optimization surrogate model technology[D]. Nanjing: College of Aerospace Engineering, Nanjing University of Aeronautics and Astronautics, 2004.(In Chinese)
[2] 趙子衡,韓旭,姜潮.基于近似模型的非線性區(qū)間數(shù)優(yōu)化方法及其應(yīng)用[J].計(jì)算力學(xué)學(xué)報(bào),2010,27(3):451-456.
ZHAO Ziheng, HAN Xu, JIANG Chao. Approximation model based nonlinear interval number optimization method and its application [J]. Chinese Journal of Computational Mechanics, 2010, 27(3):451-456. (In Chinese)
[3] SOBOL I M. Sensitivity estimates for nonlinear mathematical models [J]. Mathematical Modeling & Computational Experiment, 1993, 1(4): 407-414.
[4] SHAN Songqing,WANG G Gary. Metamodeling for high dimensional simulationbased design problems[J]. Journal of Mechanical Design, 2010, 132:051009.
[5] 湯龍,李光耀,王琥. KrigingHDMR非線性近似模型方法[J].力學(xué)學(xué)報(bào),2011,43(4):780-784.
TANG Long, LI Guangyao, WANG Hu. KrigingHDMR metamodeling technique for nonlinear problems [J]. Chinese Journal of Theoretical and Applied Mechanics, 2011, 43(4):780-784. (In Chinese)
[6] RABITZ H, AL1S F. General foundations of highdimensional model representations[J]. Journal of Mathematical Chemistry,1999,25(2/3):197-233.
[7] LI G, WANG S W, ROSENTHAL C, et al. High dimensional model representations generated from low dimensional data samples. I. MpCutHDMR[J]. Journal of Mathematical Chemistry, 2001,30(1): 1-30.
[8] CHEN Fuchang .Backpropagation neural networks for nonlinear self tuning adaptive control[J]. Control Systems Magazine, IEEE April, 1990, 10(3): 44-48.
[9] HECHTNIELSEN R.Theory of the backpropagation neural network[C]//Proc IEEE 2002 Intl Conf Neural Networks.Califormia:International Toint Conference in Nemal Networks, 2002.
[10]謝慶生,尹健,羅延科,等.機(jī)械工程中的神經(jīng)網(wǎng)絡(luò)方法[M].北京:機(jī)械工業(yè)出版社,2003: 39-51.
XIE Qingsheng, YIN Jian, LUO Yanke,et al. Neural network method in mechanical engineering [M].Beijing:China Machine Press, 2003:39-51. (In Chinese)
[11]曾喆昭,李仁發(fā).高階多通帶濾波器優(yōu)化設(shè)計(jì)研究[J].電子學(xué)報(bào),2002,30(1):87-89.
ZENG Zhezhao, LI Renfa. Study on the optimum design of the highorder multibandpass filters [J].Acta Electronica Sinica, 2002, 30(1):87-89. (In Chinese)
[12]張德豐. MATLAB 神經(jīng)網(wǎng)絡(luò)應(yīng)用設(shè)計(jì)[M]. 北京:機(jī)械工業(yè)出版社,2012:49-75.
ZHANG Defeng. MATLAB neural network application design [M]. Beijing:China Machine Press, 2012:49-75. (In Chinese)
[13]ISO 3471:2008 Earthmoving machineryrollover protective structureslaboratory tests and performance requirements[S]. British: Standards Policy and Strategy Committee,2009.
[14]ISO 3164: 1992Earthmoving machinery —Laboratory evaluation of rollover and fallingobject protective structuresspecifications for deflectionlimiting volume[S]. British: The Authority of the Standards Board, 1992.
[15]周傳月. MSC.Fatigue疲勞分析應(yīng)用與實(shí)例[M]. 北京:科學(xué)出版社,2005:86-91.
ZHOU Chuanyue. Fatigue analysis of application and examples with MSC.Fatigue[M]. Beijing:Science Press, 2005:86-91. (In Chinese)
[16]謝濤,陳火旺,康立山.多目標(biāo)優(yōu)化的演化算法[J].計(jì)算機(jī)學(xué)報(bào),2003,26(8):997-1003.
XIE Tao, CHEN Huowang, KANG Lishan. Evolutionary algorithms of multiobjective optimization problems [J].Chinese Journal of Computers, 2003, 26(8):997-1003. (In Chinese)
[17]DEB K, PRATAP A, AGARWAL S, et al. A fast elitist nondominated sorting genetic algorithm for multiobjective optimization: NSGAII,Kan GAL Report No.200001[R]. Kanpur: Indian Istitute of Technology, 2000.
[18]李偉平,王世東,周兵,等. 基于響應(yīng)面法和NSGAⅡ算法的麥弗遜懸架優(yōu)化[J]. 湖南大學(xué)學(xué)報(bào):自然科學(xué)版, 2011,38(6):27-32.
LI Weiping, WANG Shidong, ZHOU Bin, et al. Macpherson suspension parameter optimization based on response surface method and NSGAⅡ algorithm [J]. Journal of Hunan University:Natural Sciences, 2011,38(6):27-32. (In Chinese)