胡燕祝,李雷遠(北京郵電大學自動化學院,北京100876)
?
基于多層感知人工神經網絡的執行機構末端綜合定位
胡燕祝,李雷遠
(北京郵電大學自動化學院,北京100876)
摘要:非標準化執行機構的雅可比矩陣和連桿坐標系往往難以確定,導致任務空間的定位性難以分析。論文提出并證明綜合型定位方法的充分必要條件,即完成多種特別定位任務的充要條件;用反向傳播的多層感知人工神經網絡(MLP, multilayer perceptron neural network)求解逆運動學模型,在笛卡爾空間,把執行機構D-H(denavit-hartenberg)參數作為訓練集,對神經網絡進行訓練;定義一個函數,判斷執行機構定位到目標點的性能,即可定位性。經仿真驗證,神經網絡求解逆運動學模型,較傳統方法縮短了計算時間,計算效率提高20%,精度提高2.4%,可定位性最小值為0.96,最優運動學函數值4.0349×1014。
關鍵詞:機械化;控制;模型;機械臂;神經網絡;逆運動學
胡燕祝,李雷遠.基于多層感知人工神經網絡的執行機構末端綜合定位[J].農業工程學報,2016,32(01):22-29.doi:10.11975/j.issn.1002-6819.2016.01.003 http://www.tcsae.org
Hu Yanzhu, Li Leiyuan.Series actuator end integrated positioning analysis based-on multilayer perceptron neural network [J].Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2016, 32(01): 22-29.(in Chinese with English abstract)doi:10.11975/j.issn.1002-6819.2016.01.003 http://www.tcsae.org
標準化執行機構較為通用,但也不能保證執行最優的任務,工業執行機構只能重復執行一系列給定任務。由此,特殊任務或帶有優化任務的非標準化執行機構是急需的,如個性化精密制造、裝配和包裝等。若要完成這些個性化任務,則必須對執行機構的定位性或可到達。
空間性進行分析。本文分為3步,第1步,確定D-H (denavit-hartenberg)參數,建立正運動學模型[1],證明綜合定位任務的充要條件;第2步,人工神經網絡(ANNs,artificial neural network)求解逆運動學;第3步,建立定位性函數,評價定位任務性能。
非標準化執行機構定位性分析方法可分為3種,幾何分析法、參數優化分析法和基于任務分析法。幾何分析法有其局限性,不能擴展和應用到棱形連桿結構,同時不滿足多任務定位需求;參數優化法的主要缺點是受限于關節自由度和關節極限位置等;基于任務的分析方法應用先驗知識,生成齊次變換矩陣和定位動態參數;Paredis和Kholsa[2]應用基于任務分析方法產生D-H參數,并對6-DOF(degree of freedom)標準化執行機構末端進行定位分析;Kholsa等[3]在任務分析法基礎上,提出任務描述概念,進行運動學建模、定位規劃和定位控制等。本文在任務分析法的基礎上,對非結構化執行機構進行定位問題描述、建立綜合充要條件和定位性評價。
雅可比矩陣在執行機構的運動學分析中具有重要地位,機器人的分離速度控制、靜力分析和靈活性和可操作度分析等都要用到機器人的雅可比矩陣[4]。雅可比矩陣的構造方法有矢量積法、微分變換法、力和力矩遞推法和速度遞推法。基于雅可比矩陣能夠分析穩定性(特征根)、雅可比矩陣的奇異性用矩陣的秩來描述,滿秩時,可進行奇異值分解;不滿秩時,執行機構處于奇異位形。執行機構的靈活性與運動學逆解的精度與雅可比矩陣奇異值有關。同時,雅可比矩陣是由關節速度映射到執行機構末端速度,因此也是構成綜合定位[5]充要條件的關鍵因素之一。
已知執行機構末端在操作空間中的位姿,利用逆運動學解出關節角度。傳統的逆運動學求解方法有幾何法、迭代法、代數法和Particle Swarm優化法等。其封閉解也不是萬能的,只有在執行機構的雅可比矩陣為滿秩時才成立。其次,逆運動學方程通常沒有獨一無二的解,因為在關節空間中,執行機構存在多個位姿,使之定位到任務空間中的目標位置。逆運動學求解時,應避免奇異性域,奇異性即是執行機構的2個或更多個旋轉軸共線引起的不可預測的運動和速度。奇異性的存在影響執行機構末端的定位,人工神經網絡逆運動學求解方法可以盡量消除奇異域[6]。在網絡訓練和學習時,通過Levenberg-Marquart (LM)算法優化均方差(MSE)。以奇異點處的笛卡爾坐標作為測試集,以執行機構笛卡爾坐標和D-H參數作為訓練集。
為了判斷非標準化執行機構末端能否定位到給定點,及到達給定點的收斂程度,引入可到達性函數,即可定位性函數[7-8]。可定位性函數反映非標準化執行機構運動學性能。
首先,對論文中用到的一些空間參數和變量進行說明,ΓT表示任務空間,ΓQ關節空間,r1表示大臂長度,r2表示小臂長度,r3腕部長度,ΓQC受限的關節空間,qi各關節運動矢量,ξE執行機構末端位姿,ΓC配置空間,0AE基座相對于執行機構末端的變換矩陣,ΓW可到達或可定位空間,ΓCW受關節限制的可到達或可定位空間。
非標準化執行機構具有5個自由度3連桿,如圖1a。5個旋轉軸分別為腰部旋轉軸,大臂俯仰軸,小臂俯仰軸,腕部俯仰軸和腕部旋轉軸。5關節分別為腰部旋轉關節,大臂俯仰關節,小臂俯仰關節,腕部俯仰關節和腕部旋轉關節。角度儀測量各關節極限角度,以車體水平面作為參考平面,如表1所示。

圖1 非標準化執行機構及坐標系Fig.1 Non standardized execution mechanism and DH parameter coordinate system

表1 執行機構工作空間和最大轉速Table 1 Working space and maximum speed of actuator
{a,α,d}表示旋轉型連桿參數,{a,α,θ}表示棱型連桿的參數。標準D-H參數存在限制性,其關節必須繞z軸旋轉,連桿在x軸上產生位移。在配置空間ΓC中,5(N)自由度執行機構具有15(3N)個配置參數,因此D-H參數集構成了一個15維的ΓC空間,表達如下:

正運動學模型可用下列函數表示:

表示執行機構末端的運動矢量,q為各關節運動矢量。正運動學變換矩陣是1個4×4方陣。n,b和t分別表示執行機構末端在x,y和z軸的方向,p表示執行機構末端的位置(笛卡爾坐標系)。

在空間ΓCW內,執行機構末端的空間坐標為:

非標準化5軸執行機構完成綜合定位任務的充要條件為:能夠找到所有的D-H參數滿足(DH,q)=p且rank(Jacobian(q))=5。證明如下。
充分性:若非標準化5軸執行機構可以完成綜合定位任務,則能夠找到所有的D-H參數滿足?p∈ΓT,?q∈ΓQCf(DH,q)=p且rank(Jacobian(q))=5。
眾所周知,DH∈ΓC,ΓCW?ΓW,ΓQC?ΓQ,設P是ΓT內的一個點集,

ΓT內的每一個點含有6個維度,分別用執行機構末端的位置和方向進行定義,如下:

對于5自由度的非標準化執行機構,其ΓQ維數為5,所以,關節向量為:

每個關節在ΓQC內,都有其極限位姿,上限位姿和下限位姿約束為:

定義ΓW為無關節限制時非標準化執行機構末端能夠達到或定位到的世界坐標系內所有點的集合[11-15]。根據正運動學原理(2)得到映射,

同理,在ΓWS內,且ΓCW?ΓW,ΓQC?ΓQ,

再由(1)式,得

由于執行機構任務空間存在奇異性,通過雅可比矩陣予以避免。對(2)求導,


執行機構雅可比矩陣[16-17]是1個6×N矩陣,N為關節數。為了避免奇異性,rank(J(DH,q))=5。推導出結論:存在D-H參數滿足且rank(Jacobian(q))=5。
必要性:如果能夠找到D-H參數滿足?p∈ΓT,?q∈且rank(Jacobian(q))=5,則非標準化5軸執行機構能夠完成綜合定位任務。
給定任務空間執行機構末端的位置坐標(X,Y,Z),通過ANNs確定執行機構各關節角度θ1,θ2,θ3,θ4和θ5。神經網絡采用前饋多層感知(MLP,multilayer perceptron neural network),把執行機構的笛卡爾坐標和D-H參數作為訓練集[18-20],不斷更新神經網絡權值,使均方誤差參數(MSE, mean square error)達到最小;驗證集,確定網絡結構或者控制模型復雜程度的參數,當泛化值停止改善時,停止訓練;訓練后,用測試集對神經網絡性能進行獨立測試。
MLP前饋反向傳播網絡用權值(w)和偏移量(b)表示如下:

xi為網絡輸入,即(X,Y,Z);wi為每個輸入的權值;b為偏移量;S為輸出,即(θ1,θ2,θ3,θ4,θ5);神經網絡架構,如圖2所示。

圖2 神經網絡架構Fig.2 Neural network architecture
采用Levenberg-Marquardt(LM)算法[21-23]訓練MLP網絡,通過調整學習速率,獲得最優權值wi,并且使均方誤差(MSE)取得最小值。LM算法以Newton算法和梯度下降算法為基礎,取誤差函數Ei的1階導數,使全局誤差最小[24]。表達式為,

d為Ei的1階導數,ds為Ei的2階導數,e為自然對數函數,λ為阻尼因子。均方誤差函數的表達式如下,

Ei為第i個輸入數據的誤差,n為輸入數據數量。
MLP神經網絡采用有監督學習,包含3個輸入,帶有20個神經元的隱層,5個輸出。隱層具有tansigmoid激勵函數,范圍[-1,1];輸出層具有pureline線性激勵函數。
神經網絡訓練集、驗證集和測試集是歸一化后的數據,要求取值范圍在(0,1)內,所以采用歸一化函數f(x)=1/(1+e-x)。證明如下,令x∈(-∞,∞),

神經網絡訓練出來的模型,進行逆運動學求解時[25],不存在復數解,因此,不考慮復數解情況。建立一個可定位性函數,判定執行機構末端能否按照需要的方向到達任務空間中目標點[26-29]。利用歸一化模型,評價ΓT空間中任意點的可定位性,

g為點p對應所有運動學逆解的數量,函數reachability(DH)取值范圍為[0,1]。當取(0,1)時,逆運動學解中至少有一組受關節限制而無法定位到;當取0時,最優解中至少有一個關節角度超過極限位置;當取1時,最優解中所有關節都處于中間位置,無論哪一組解都可以定位到目標點[30-31]。最優逆運動學解選取,即執行機構的最優配置,函數如下,

其中,J(DH,q1)是6×5的矩陣,JT(DH,q1))J(DH,q1)為5×5矩陣,評價ΓQ內NUM個點的可定位性能,

評價ΓT內定位到NUM個點的運動學性能,有

取目標函數的最大值,是為了找到目標點處的最合適的速度變換矩陣,即雅可比矩陣。雅可比矩陣是由關節速度到執行機構末端速度的映射,能夠反映執行機構的運動學性能。
此執行機構的標準D-H參數如表2。DH參數對應的正運動學變換矩陣為,

所以,初始狀態末端空間坐標為(433.5,0,-255.4)。由正運動學推導末端運動區域,如圖3所示,圖中各個點表示執行機構末端所能達到的空間坐標,近似球形。執行機構末端運動范圍構成ΓT,所以?p∈ΓT,?q∈ΓQCf(DH,q)=p。此執行機構的雅可比矩陣是一個6*5矩陣,初始狀態下DH參數對應的雅可比矩陣為,


表2 標準D-H參數Table 2 Standard D-H parameters

圖3 末端運動范圍Fig.3 Manipulator end motion range
利用傳統方法計算逆運動學解,共2 000組數據,在ΓT內,(X,Y,Z)數據隨機生成,如圖3中的點集,逆運動學解為(Θ1,Θ2,Θ3,Θ4,Θ5),并將(X,Y,Z,Θ1,Θ2,Θ3,Θ4,Θ5)數據按列歸一化,歸一化函數為f(x)=1/(1+e-x)。1 400組(X,Y,Z,Θ1,Θ2,Θ3,Θ4,Θ5)作為訓練樣本,構成1400*8維矩陣,用試湊法確定最佳隱層節點數,先設置較少的隱節點訓練網絡,然后逐漸增加隱節點數,進行訓練,確定網絡誤差最小時對應的隱節點數為20;最大訓練次數為28 302,訓練目標誤差設置為1×10-4,動量因子常數為1×10-4,學習速率為1×10-5,學習速率增加比率為5,學習速率減少比率0.5;引起訓練結束的條件是Validation Checks為20,訓練時間47 s,迭代30次,均方誤差為5.96×10-5,梯度值為0.028 8;500組作為驗證樣本,維數500×8,100組測試樣本,維數100×8。訓練樣本和測試樣本相互獨立。θ1到θ5的網絡輸出和理論計算值之間的誤差Err=網絡估計值-理論計算值,如圖4,θ1的估計誤差最大約為3.7,θ2的最大估計誤差約為3.1,θ3的最大估計誤差約為3.5,θ4的最大估計誤差約為3.3,θ5的最大估計誤差約為4.5,所有實際誤差均在5以內,能夠滿足實際需求。
采用神經網絡解逆運動學,需要在計算時間和精度之間進行權衡,往往是縮短了計算時間,而精度卻不理想。傳統算法計算時間為1.2 s,神經網絡訓練好后,計算時間為0.9 s,效率提高了20%。Luv Aggarwal等用神經網絡求解精度為87.5%,論文神經網絡求解精度為89.9%,提高了2.4%。因此,在實時性要求不高而精度要求較高時,應該采用傳統計算方法,而在實時性要求較高而精度要求不高且系統為非線性時,可以選擇神經網絡求解方法。

表3 定位到(41.4, 89.0, 104.5)時第1組解Table 3 First group solutions of positioning to(41.4, 89.0, 104.5)

圖4 網絡估計值與理論計算值的誤差Fig.4 Error between estimated value and calculated value
在ΓT內,?p∈ΓT,設定位點p=(41.4,89.0,104.5),逆運動學解如表3和表4,共2組解。由式(18),可計算每一組逆運動學解的可定位性值,如圖5(a),用紅色’.’表示,軸表示第幾組逆運動學解,y軸表示對應的定位性值。第21組逆運動學解的可定性函數值最小,因此在p點處的可定位性為reachability(DH)=0.96,藍色’o’表示;最優運動學性能為configuration(DH)=4.034 9×1014,1~21組解的最優運動學性能如圖5(b)所示,用紅色’.’表示,x軸表示第幾組逆運動學解,y軸表示對應的最優運動學性能值,藍色’o’表示最優運動學性能值中最大值。從圖中可以看出,第12組逆運動學解(21.61,125.73,108.42,99.41,0)可以達到最優運動學。

表4 定位到(41.4, 89.0, 104.5)時第2組解Table 4 Second group solutions of positioning to(41.4, 89.0, 104.5)

圖5 點的可定位性及運動學性能分布Fig.5 Distribution of localization and kinematic performance for a point
為滿足用戶個性化需求,需對執行機構的結構重新設計,提出可定位性充要條件,在滿足精度要求的前提下,盡量提高執行機構逆運動學求解效率,并對可定位性定量分析,其實質是基于可定位性的最優配置。論文對可定位性充要條件進行證明,計算執行機構末端任務空間的運動范圍,訓練基于MLP的神經網絡并逆運動學求解,效率提高了20%,精度提高2.4%,并舉例計算p=(41.4,89.0,104.5)時,可定位性值reachability(DH)=0.96,configuration(DH)=4.039×1014。接下來工作完成一整套環形或平面多點等任務的最低功耗研究。
[參考文獻]
[1]張曉燕,孫劍橋,丁千.5自由度機械臂的運動學分析與時滯控制[J].西安交通大學學報,2013,10(47);103-108.Zhang Xiaoyan, Sun Jianqiao, Ding Qian.Kinematics analysis and time delayed control of a five-degree of freedom manipulator [J].Journal of Xi’An Jiaotong University, 2013, 10(47): 103-108.(in Chinese with English abstract)
[2] Paredis C J, Khosla P K.Kinematic design of serial link manipulators from task specifications[J].IntJ Robotics Res, 1993, 12(3): 274-87.
[3] Khosla P K.A multi-population genetic algorithm and its application to design of manipulators[J].In: Proc.IEEE/RSJ int conf on intell robots and syst, 1992, 3(24): 7-10.
[4]李誠,謝志江,倪衛,等.六自由度裝校機器人雅可比矩陣的建立及奇異性分析[J].中國機械工程,2012,5(11):1165-1169.Li Cheng, Xie Zhijiang, Ni Wei, et al.Establishment of jacobian matrix and singularity analysis of a 6-DOF installing calibrating robot[J].China Mechanical Engineering, 2012, 5(11): 1165-1169.(in Chinese with English abstract)
[5] Sarosh P, Tarek S.Task based synthesis of serial manipulators[J].Journal of Advanced Research, 2015, 39(8):2-7.
[6] Luv A, Kush A, Ruth J U.Use of artificial neural networks for the development of an inverse kinematic solution and visual identification of singularity zone [J].Variety Management in Manufacturing, 2014,8(10):812-817.
[7] Tian T X, Jiang H Z, Tong Z Z, et al.An internal parameter identification method of eliminating system damping effect for a six-degree-of-freedom parallel manipulator[J].Chinese Journal of Aeronautics, 2015, 12(24): 450-457.
[8] Monika Z F, Ana M D.Complexity analysis for calculating the jacobian matrix of 6-DOF reconfigurable machines[J].Variety Management in Manufacturing, 2014, 2(19): 218-223.
[9]任子武,王振華,李娟,等.基于混合類電磁機制算法的機械臂逆運動學解[J].機械工程學報,2014,17(48):21-27.Ren Ziwu, Wang Zhenhua, Li Juan, et al.Inverse kinematics solution for robot manipulator based on hybrid electromagnetism-like mechanism algorithm [J].Journal ofMechanical Engineering, 2014, 17(48):21-27.(in Chinese with English abstract)
[10]高自成,李立君,李昕,等.齒梳式油茶果采摘機采摘執行機構的研制與試驗[J].農業工程學報,2013,10(27):38-44.Gao Zicheng, Li Lijun, Li Xin, et al.Development and test of picking actor in oil-tea camellia fruit picking machine of tooth comb type[J].Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2013, 10(27): 38-44.(in Chinese with English abstract)
[11]章德賓,徐家鵬,許建軍,等.基于監測數據和BP神經網絡的食品安全預警模型[J].農業工程學報,2010, 1(6):55-62.Zhang Debin, Xu Jiapeng, Xu Jianjun, et al.Model for food safety warning based on inspection data and BP neural network [J].Transactions of the Chinese Society of Agricultural (Transactions of the CSAE), 2010,1(6):55-62.(in Chinese with English abstract)
[12] Yong J K, Shan B C, Sang B K, et al.A stiffness-adjustable hyper redundant manipulator using a variable neutral-line mechanism for minimally invasive surgery [J].IEEE Transactions on Robotics, 2014,2(30): 382-394.
[13] Matteo C P, Luca C, Giacomo P, et al.Analysis and design of a reconfigurable 3-DoF parallel manipulator for multimodal tasks [J].IEEE/ASME Transactions on Mechatronics, 2015, 4(20): 1975-1984.
[14] Janne K, Jouni M.Stability-guaranteed force-sensorless contact force/motion control of heavy-duty hydraulic manipulators [J].IEEE Transactions on Robotics, 2015, 4(31): 918-934.
[15] Rong J W, Rajkumar M.Design of fuzzy-neural-networkinherited back stepping control for robot manipulator including actuator dynamics [J].IEEE Transactions on Fuzzy Systems, 2014, 4(22): 709-721.
[16] Ahmet D, Koksal E.Trajectory tracking control for a 3-DOF parallel manipulator using fractional-order control [J].IEEE Transactions on Industrial Electronics, 2014, 7(61): 3417-3423.
[17] Sehoon O, Kyoungchul K.Two-degree-of-freedom control of a two-link manipulator in the rotating coordinate system[J].IEEE Transactions on Industrial Electronics, 2015, 9(62): 5598-5607.
[18] Kim S M.Lumped element modeling of a flexible manipulator system[J].IEEE/ASMETransactionson Mechatronics,2015,2(20): 967-973.
[19] Chu Z Y, Cui J, Sun F C.Fuzzy adaptive disturbance-observerbased robust tracking control of electrically driven free-floating space manipulator[J].IEEE Systems Journal, 2014, 2(8):343-351.
[20] Luca C, Massimo C, Giacomo P.Analysis of kinematics and reconfigurability of a spherical parallel manipulator[J].IEEE Transactions on Robotics, 2014, 6(30):1541-1546.
[21]吳玉香,王聰.基于確定學習的機器人任務空間自適應神經網絡控制[J].自動化學報,2013,6(39):806-815.Wu Yuxiang, Wang Cong.Deterministic learning based adaptive network control of robot in task space[J].Acta Automatica Sinica, 2013, 6(39): 806-815.(in Chinese with English abstract)
[22] Mai T L, Wang Y N.Adaptive force/motion control system based on recurrent fuzzy wavelet CMAC neural networks for condenser cleaning crawler - type mobile manipulator robot[J].IEEE Transactions on Control systems Technology, 2014, 5(22):1973-1981.
[23] Long J, Zhang Y N.Discrete-time neural network for online timevarying nonlinear optimization with application to manipulator motion generation[J].IEEE Transactions on Neural Networks and Learning Systems, 2015,7(26):1525-1531.
[24]張凱良,楊麗,張鐵中.草莓收獲機器人采摘執行機構設計與試驗[J].農業機械學報,2011,9(42):155-160.Zhang Kailiang,Yang Li, Zhang Tiezhong.Design and experiment of picking mechanism for strawberry harvesting robot [J].Transactions of the Chinese Society for Agricultural Machinery, 2011, 9(42): 155-160.(in Chinese with English abstract)
[25] TanN.Calibrationforaccuracyimprovementofserialmanipulators based on compressed sensing[J].Electronics Letters, 2015, 11 (51): 820-822.
[26] Luciano C, Giovanni M, Marco N, et al.A joint-angle estimation method for industrial manipulators using inertial sensors [J].IEEE/ASME Transactions on Mechatronics, 2015, 5(20): 2484-2494.
[27] Hamid S, Luigi V, Mehdi K, et al.Task-space control of robot manipulators with null-space compliance [J].IEEE Transactions on Robotics, 2014, 2(30): 493-505.
[28] Liu YC,Nikhil C.Gravity-compensation-drivenpositionregulation forroboticsystemsunderinput/outputdelays[J].IEEETransactions on Control Systems Technology, 2014, 3(22): 995-1003.
[29] Kern J, Jamett M, Urrea C.Development of a neural controller applied in a 5-DOF robot redundant[J].IEEE Latin America Transactions, 2014, 2(12): 98-106.
[30] Spencer B B, Aaron M D.Robust resonant frequency-based contact detection with applications in robotic reaching and grasping[J].IEEE/ASME Transactions on Mechatronics, 2014, 5 (19): 1552-1561.
[31] Corrado G, Lo B, Fabio G.Real-time planner in the operational space for the automatic handling of kinematic constraints [J].IEEE Transactions on Automation Science and Engineering, 2014, 3(11): 730-738.
Series actuator end integrated positioning analysis based-on multilayer perceptron neural network
Hu Yanzhu, Li Leiyuan
(College of Automation, Beijing University of Posts and Telecommunications, Beijing, 100876, China)
Abstract:It is difficult to establish Jacobian matrix and determine the coordinate frames of links for non-standard actuator.A new analytical method to establish the Jacobian matrix and determine the coordinate frames for joints and links are proposed in this paper.The proposed method made the positioning analysis of end-effector easier in space.At the same time, it is necessary to prove the effectiveness of the proposed method theoretically and verify the localization and configuration capabilities through simulations.First of all, forward kinematics model was set up based on a non-standard five Degree Of Freedom(5-DOF)actuator.A frame transformation is performed from base coordinate to end-effector coordinate.The relation between two adjacent joints is defined by a homogenous pose matrix.Secondly, the necessary and sufficient conditions for comprehensive localization are derived.They can guide the actuator to perform various tasks, such as tracking, assembly and autonomous grasping.A 5-DOF actuator is considered here as an example and this holds good for any N-DOF.Thirdly, inverse kinematics solutions are obtained by using artificial Neural Network(NN)based on backpropagation Multi-Layer Perceptron(MLP, multilayer perceptron NN)and are not unique.A unique solution using nonlinear minimization optimization is found.A NN based on supervisory learning method including three inputs, twenty neurons and five outputs has been used.Excitation function tansigmoid and linear excitation function pureline are in hidden and outer layers respectively.In Cartesian coordinate space, NN is trained by means of Levenberg Marquardt(LM)algorithm.The training sets used are Denav Hartenberg(DH)parameters and Cartesian coordinates.The weights are updated continuously which reduces the Mean Square Error(MSE)gradually.When MSE reaches the threshold set up, NN training will be terminated.After training, the test sets are used to examine the capability of NN.Fourthly, there are two evaluation functions viz., localization and cost functions.The localization function is defined to evaluate the positioning property of end-effector.At the same time, in task space, it will check whether the actuator has reached the target point along the direction needed or not.The cost function is defined to evaluate the kinematics configuration.There is a great relevance between cost function and Jacobian matrix.Velocity mapping from each joint to the end-effector was described by Jacobian matrix.So the cost function could give expression for kinematic configuration.At the end, simulations and experiments are conducted.The settings include industrial computer UNO2184G, 5-DOF non-standard actuator, Windows 7, MATLAB2012a.Coordinate frames for each joint are established and D-H parameters are determined.Then relative pose matrix is obtained between each of the two adjacent joints.Initial end-effector pose is obtained following right multiplication rule.The end-effector space range is formed under each joint operation range.Then, simulation is performed using NN, obtained localization and cost functions.The following results are obtained.The rank of Jacobian matrix is equal to 5.Therefore, this actuator met necessary and sufficient conditions for comprehensive positioning.NN method for solving inverse kinematics has reduced the computational complexity compared to conventional method.There are 21 groups of solutions when positioning to(41.4, 89.0, 104.5).The optimal solution obtained is(21.61, 91.44, 135.52, 221.42, 0)according to localization function rule.The optimal solution obtained according to cost function rule is(21.61, 125.73, 108.42, 221.99.41, 0).NN accuracy is 89.9%(approximately)while conventional method is 87.5% .By approximate estimation, the errors for θ1,θ2,θ3,θ4and θ5are 3.7°, 3.1°, 3.5°, 3.3°and 4.5°respectively.NN used 1.2 seconds while conventional method completed in 0.9 seconds.Therefore, computation accuracy has improved by 20% and efficiency by 2.4%.If the system is linear, the conventional method is chosen when less demand in real-time.In contrast, if the system is nonlinear, new method proposed in this paper is chosen when more demand in real-time.The minimum value of localization function is 0.96.The maximum value of cost function is 4.0349×1014.These two parameters decide the comprehensive positioning and the kinematics configuration.From the results presented, it can be concluded that the non-standard actuator with MLP has better localization and optimal configuration.
Keywords:mechanization; control; models; manipulator; neural networks; inverse kinematics
作者簡介:胡燕祝(1970-),教授,博士后,博士生導師,主要從事視覺測量與機器人方向。北京北京郵電大學自動化學院,100876。Email:YZH@263.net
基金項目:北京市計劃課題《軌道交通事故現場應急處置裝備研制與示范應用》(Z131100004513006)
收稿日期:2015-07-25
修訂日期:2015-11-13
中圖分類號:TP212
文獻標志碼:A
文章編號:1002-6819(2016)-01-0022-08
doi:10.11975/j.issn.1002-6819.2016.01.003