劉艷君,陶太洋,丁鋒
?
多輸入動(dòng)態(tài)調(diào)節(jié)系統(tǒng)的正交匹配追蹤迭代辨識(shí)算法
劉艷君1, 2,陶太洋1, 3,丁鋒1, 2
(1. 江南大學(xué)物聯(lián)網(wǎng)工程學(xué)院,江蘇無(wú)錫,214122;2. 江南大學(xué)輕工過(guò)程先進(jìn)控制教育部重點(diǎn)實(shí)驗(yàn)室, 江蘇無(wú)錫,214122;3. 河南中光學(xué)集團(tuán)有限公司,河南南陽(yáng),473000)
針對(duì)含有未知時(shí)滯的多輸入單輸出動(dòng)態(tài)調(diào)節(jié)系統(tǒng),基于過(guò)參數(shù)化后系統(tǒng)參數(shù)向量的稀疏特性,在有限測(cè)量數(shù)據(jù)下,將壓縮感知理論和遞階迭代思想相結(jié)合,提出一種正交匹配追蹤迭代辨識(shí)算法。該算法可以辨識(shí)多輸入動(dòng)態(tài)調(diào)節(jié)系統(tǒng)的未知時(shí)滯、參數(shù)和部分階次。研究結(jié)果表明:與最小二乘迭代算法相比,該算法不需要大量的采樣數(shù)據(jù),可以節(jié)約采樣成本,提高辨識(shí)效率。該算法能夠有效地估計(jì)這類系統(tǒng)的參數(shù)與時(shí)滯。
系統(tǒng)辨識(shí);時(shí)滯與參數(shù)估計(jì);壓縮感知;正交匹配追蹤;有限采樣
系統(tǒng)的參數(shù)估計(jì)在很多領(lǐng)域如系統(tǒng)辨識(shí)[1?3]、自適應(yīng)控制[4?5]、信號(hào)處理[6?7]等得到廣泛應(yīng)用。常規(guī)的參數(shù)估計(jì)算法如最小二乘類算法[8?9]、牛頓迭代算 法[10?11]、隨機(jī)梯度算法[12]等需要有充分的采樣數(shù)據(jù)才能保證其辨識(shí)性能。時(shí)滯現(xiàn)象在過(guò)程工業(yè)中普遍存在且對(duì)系統(tǒng)的控制精度有較大的影響,因此,系統(tǒng)的時(shí)滯估計(jì)也非常重要[13?15]。對(duì)于具有未知時(shí)滯的多輸入動(dòng)態(tài)調(diào)節(jié)系統(tǒng),系統(tǒng)參數(shù)化后將得到1個(gè)高維的參數(shù)向量,若采用常規(guī)辨識(shí)方法則需要大量的采樣數(shù)據(jù),這將耗費(fèi)大量的辨識(shí)成本, 而且當(dāng)系統(tǒng)時(shí)滯較大時(shí)可能帶來(lái)較大的估計(jì)誤差。為此,本文作者將尋求一種在有限采樣數(shù)據(jù)下,能夠有效辨識(shí)這類系統(tǒng)時(shí)滯與參數(shù)的方法。通常用于控制的系統(tǒng)模型階次較低,參數(shù)化后的系統(tǒng)模型中僅有少量參數(shù)。對(duì)于含有未知時(shí)滯的系統(tǒng),若不直接考慮時(shí)滯,則參數(shù)化后的系統(tǒng)模型可由一高維參數(shù)向量表示,但該參數(shù)向量中僅有少量的非零參數(shù)且位置未知,這樣的向量稱為稀疏向量,由稀疏向量表示的系統(tǒng)稱為稀疏系統(tǒng)[16]。壓縮感知理論表明: 在一定條件下,稀疏系統(tǒng)可在采樣數(shù)據(jù)量低于系統(tǒng)參數(shù)維數(shù)的情況下實(shí)現(xiàn)系統(tǒng)參數(shù)的估計(jì)[17?21]。DUMITRESCU等[22]提出一種正交最小二乘算法辨識(shí)線性稀疏系統(tǒng);SANANDAJI等[16]針對(duì)含有未知時(shí)滯的多輸入單輸出ARX系統(tǒng)采用分塊正交匹配追蹤算法,在有限采樣數(shù)據(jù)下估計(jì)系統(tǒng)的參數(shù)和時(shí)滯;LIU等[23]針對(duì)白噪聲干擾且含有未知時(shí)滯的MISO-FIR模型采用閾值正交匹配追蹤算法,在有限的采樣數(shù)據(jù)下估計(jì)系統(tǒng)的參數(shù)和時(shí)滯。此外,1范數(shù)正則化算法[24]、梯度追蹤算法[25]等也相繼被提出,用于在有限采樣數(shù)據(jù)下辨識(shí)系統(tǒng)的參數(shù)和時(shí)滯。本文作者將遞階辨識(shí)原理和壓縮感知重構(gòu)理論相結(jié)合,研究具有未知時(shí)滯的多輸入動(dòng)態(tài)調(diào)節(jié)系統(tǒng)。
多輸入單輸出動(dòng)態(tài)調(diào)節(jié)系統(tǒng)有如下表述形式[26]:
式中:()為系統(tǒng)輸出;u()和d分別為第個(gè)通道的輸入和時(shí)滯;()為零均值白噪聲;(),B()和()為單位后移算子?1的常系數(shù)多項(xiàng)式(?1()=(?1)),可表示為
其中:階次n和n已知,時(shí)滯d和輸入通道的階次n未知。假設(shè)當(dāng)≤0時(shí),()=0,u()=0,()=0。
定義系統(tǒng)的噪聲模型
其中:()為噪聲模型的輸出,是不可測(cè)的內(nèi)部變量。
定義
(4)
;
;
;
;
;
;
其中:為輸入數(shù)據(jù)回歸長(zhǎng)度,滿足>>;為含有個(gè)零元素的零塊。
系統(tǒng)模型(1)可寫(xiě)為
(5)
當(dāng)=1,2,…,時(shí),可獲得組測(cè)量數(shù)據(jù)。定義
則辨識(shí)模型(5)可寫(xiě)為如下形式:
(6)
若輸入信號(hào)持續(xù)激勵(lì),則參數(shù)向量的估計(jì)值可由如下最小二乘迭代(LSI)算法得到[13, 27?28]:
;
;
其中:為與時(shí)間無(wú)關(guān)的迭代變量。上述算法要獲得良好的參數(shù)估計(jì),采樣數(shù)據(jù)量需滿足>>。由于參數(shù)向量的維數(shù)很大,因此,LSI算法辨識(shí)系統(tǒng)(1)需要大量的采樣數(shù)據(jù),將耗費(fèi)較高的采樣成本。
由于參數(shù)向量具有稀疏特性,受壓縮感知重構(gòu)理論啟發(fā),對(duì)上述LSI算法加以改進(jìn),使得在有限采樣數(shù)據(jù)量下,可以有效地估計(jì)系統(tǒng)參數(shù)和時(shí)滯。
辨識(shí)模型(6)中,參數(shù)向量為稀疏向量。由式(4)可知,參數(shù)向量的稀疏度為[29]
很明顯稀疏度<<。若輸入u()持續(xù)激勵(lì),則根據(jù)壓縮感知理論,式(6)在有限采樣數(shù)據(jù)下的辨識(shí)問(wèn)題即為如下最小零范數(shù)約束求解問(wèn)題[21, 29]:
(7)
式中:為設(shè)定的允許誤差。
根據(jù)壓縮感知重構(gòu)理論,在有限的采樣數(shù)據(jù)下,即當(dāng)<時(shí),式(7)可以通過(guò)正交匹配追蹤(OMP)算法求解。由于中間變量?(?)未知,導(dǎo)致信息矩陣中含有許多未知項(xiàng),使得OMP算法不能直接應(yīng)用。為解決這一問(wèn)題,借鑒上述LSI算法中的遞階迭代辨識(shí)思想:在第次迭代時(shí),未知項(xiàng)?(?)用前一次迭代的估計(jì)值代替。換句話說(shuō),就是把OMP算法嵌套于迭代算法中。
在第次迭代時(shí),定義
(9)
由式(10)可知:輸出向量可以表示為信息矩陣各列的線性組合加上白噪聲項(xiàng)的形式。由于參數(shù)向量的稀疏性,式(10)等號(hào)的右邊只有少量的非零項(xiàng)。OMP算法的主要思想就是將這些非零項(xiàng)逐個(gè)挑選出來(lái),并對(duì)這些非零項(xiàng)參數(shù)進(jìn)行估計(jì)。由于OMP算法每次只選出一個(gè)非零項(xiàng),與LSI算法相比,OMP算法只需要較少的測(cè)量數(shù)據(jù)量。
在第次迭代時(shí),將式(6)中的用代替,表示信息矩陣的第列。OMP算法是一種離線的迭代算法,令(為OMP算法的迭代變量)。在第次迭代中的第次OMP迭代,定義準(zhǔn)則函數(shù)
(12)
由式(12)對(duì)求導(dǎo)并令其為0,可得
將回代入式(12)可得
(13)
(15)
(17)
為抑制噪聲對(duì)參數(shù)估計(jì)誤差的影響,本文通過(guò)設(shè)定1個(gè)很小的閾值H=對(duì)每次估計(jì)的進(jìn)行濾波。令為中的第個(gè)參數(shù),若,則令,濾波后的估計(jì)值記為,并由和更新殘差[23, 29]:
(19)
(21)
正交匹配追蹤迭代(OMPI)算法的具體實(shí)施步驟如下。
(22)
考慮含有時(shí)滯的五輸入單輸出動(dòng)態(tài)調(diào)節(jié)系統(tǒng):
輸入通道的時(shí)滯分別為1=9,2=23,4=29,5=17,取數(shù)據(jù)回歸長(zhǎng)度=40。系統(tǒng)的參數(shù)向量為。系統(tǒng)參數(shù)向量的維數(shù),稀疏度為。
仿真時(shí),u()采用零均值單位方差不相關(guān)可測(cè)隨機(jī)信號(hào),()采用方差為2的零均值白噪聲。取數(shù)據(jù)長(zhǎng)度=150。當(dāng)噪聲方差為2=0.102時(shí),分別采用LSI算法和OMPI算法估計(jì)系統(tǒng)參數(shù),并使用相同的閥值對(duì)LSI的參數(shù)估計(jì)值予以濾波,參數(shù)估計(jì)相對(duì)誤差隨迭代次數(shù)的變化曲線如圖1所示。另外,當(dāng)噪聲方差分別為2=0.102和2=0.502時(shí),采用OMPI算法估計(jì)該系統(tǒng)參數(shù),估計(jì)誤差隨迭代次數(shù)的變化如圖2所示。
1—LSI;2—OMPI。
σ2:1—0.102;2—0.502。
當(dāng)2=0.102時(shí),OMPI算法得到的參數(shù)估計(jì)為
(23)
(24)
綜合圖1、圖2和式(23)、(24)有以下結(jié)論:
1) 該稀疏系統(tǒng)在有限采樣數(shù)據(jù)量下,OMPI算法的參數(shù)辨識(shí)精度明顯比LSI算法的高。
2) 當(dāng)噪聲方差2=0.502時(shí),OMPI算法的估計(jì)精度略低于2=0.102時(shí)的估計(jì)精度,且所需要的迭代次數(shù)也略多于2=0.102時(shí)的次數(shù),說(shuō)明OMPI算法適用于噪聲水平不太高的場(chǎng)合。
3) OMPI算法不僅能夠有效估計(jì)系統(tǒng)的參數(shù),而且能有效地估計(jì)系統(tǒng)的未知時(shí)滯d以及部分階次。
1) 研究具有未知時(shí)滯的多輸入單輸出動(dòng)態(tài)調(diào)節(jié)模型的時(shí)滯估計(jì)和參數(shù)辨識(shí)問(wèn)題。結(jié)合遞階迭代辨識(shí)思想和正交匹配追蹤算法,提出一種正交匹配追蹤迭代算法。
2) 該算法可以同時(shí)估計(jì)系統(tǒng)的參數(shù), 未知時(shí)滯以及系統(tǒng)中的部分階次。此外,該算法不需要大量的采樣數(shù)據(jù),因而可以節(jié)約辨識(shí)實(shí)驗(yàn)中采樣成本,提高了辨識(shí)效率。仿真結(jié)果證明該算法是有效的。
[1] S?DERSTR?M T. A generalized instrumental variable estimation method for errors-in-variables identification problems[J]. Automatica, 2011, 47(8): 1656?1666.
[2] DING Feng, DUAN Honghong. Two-stage parameter estimation algorithms for Box-Jenkins systems [J]. IET Signal Processing, 2013, 7(8): 646?654.
[3] DING Feng, LIU Ximei, CHEN Huibo, et al. Hierarchical gradient based and hierarchical least squares based iterative parameter identification for CARARMA systems [J]. Signal Processing, 2014, 97: 31?39.
[4] SHEN Yi, LIU Lijun, DOWELL E H. Adaptive fault-tolerant robust control for a linear system with adaptive fault identification[J]. IET Control Theory and Applications, 2013, 7(2): 246?252.
[5] KELOUWANI S, ADEGNON K, AGBOSSOU K, et al. Online system identification and adaptive control for PEM fuel cell maximum efficiency tracking[J]. IEEE Transactions on Energy Conversion, 2012, 27(3): 580?592.
[6] PHILLIP F. Pulse signal and source identification using fuzzy-neural techniques [J]. IEEE Aerospace and Electronic Systems Magazine, 2013, 28(1): 22?33.
[7] SHI Yang, FANG Huazhen. Kalman filter based identification for systems with randomly missing measurements in a network environment [J]. International Journal of Control, 2010, 83(3): 538?551.
[8] LIU Yanjun, DING Feng. Convergence properties of the least squares estimation algorithm for multivariable systems [J]. Applied Mathematical Modelling, 2013, 37(1/2): 476?483.
[9] LIU Yanjun, DING Feng, SHI Yang. Least squares estimation for a class of non-uniformly sampled systems based on the hierarchical identification principle [J]. Circuits, Systems and Signal Processing, 2012, 31(6): 1985?2000.
[10] DING Feng, MA Junxia, XIAO Yongsong. Newton iterative identification for a class of output nonlinear systems with moving average noises [J]. Nonlinear Dynamics, 2013, 74(1/2): 21?30.
[11] LI Junhong, DING Feng, HUA Liang. Maximum likelihood Newton recursive and the Newton iterative estimation algorithms for Hammerstein CARAR systems [J]. Nonlinear Dynamics, 2014, 75(1/2): 235?245.
[12] DING F, LIU Xingao, CHU Jian. Gradient-based and least-squares- based iterative algorithms for Hammerstein systems using the hierarchical identification principle [J]. IET Control Theory and Applications, 2013, 7(2): 176?184.
[13] 丁鋒. 系統(tǒng)辨識(shí)新論[M]. 北京: 科學(xué)出版社, 2013: 199?220. DING Feng. System identification-new theory and methods [M]. Beijing: Science Press, 2013: 199?220.
[14] 王貞, 吳斌. 基于最小二乘法的時(shí)滯實(shí)時(shí)在線估計(jì)方法[J]. 振動(dòng)工程學(xué)報(bào), 2009, 22(6): 625?631. WANG Zhen, WU Bin. A real-time approach to delay estimation based on the Least-Square algorithm [J]. Journal of Vibration Engineering, 2009, 22(6): 625?631.
[15] 王建宏, 王道波, 王志勝. 多個(gè)未知時(shí)延的MISO系統(tǒng)的遞推辨識(shí)[J]. 控制與決策, 2010, 25(1): 93?98. WANG Jianhong, WANG Daobo, WANG Zhisheng. Recursive identification of MISO systems with multiple unknown time delays [J]. Control and Decision, 2010, 25(1): 93?98.
[16] SANANDAJI B M, VINCENT T L, WAKIN M B, et al. Compressive system identification of LTI and LTV ARX models [C]//Proceedings of the 50th IEEE Conference on Decision and Control and European Control Conference (CDC-ECC). Orlando, Florida, USA, 2011: 791?798.
[17] DONOHO D L. Compressed sensing [J]. IEEE Transactions on Information Theory, 2006, 52(4): 1289?1306.
[18] CANDèS E J. Robust uncertainty principles: Extra signal reconstruction from highly frequency information [J]. IEEE Transactions on Information Theory, 2006, 52(2): 489?509.
[19] CANDèS E J, ROMBERG J R, TAO T. Stable signal recovery from incomplete and inaccurate measurements [J]. Communications on Pure and Applied Mathematics, 2006, 59(8): 1207?1223.
[20] CANDèS E J, WAKIN M B. An introduction to compressive sampling [J]. IEEE Signal Processing Magazine, 2008, 25(2): 21?30.
[21] TROPP J A. Greed is good: algorithmic results for sparse approximation [J]. IEEE Transactions on Information Theory, 2004, 50(10): 2231?2242.
[22] DUMITRESCU B, ONOSE A, HELIN P, et al. Greedy sparse RLS [J]. IEEE Transactions on Signal Processing, 2012, 60(5): 2194?2207.
[23] LIU Yanjun, TAO Taiyang. A CS recovery algorithm for model and time-delay identification of MISO-FIR systems [J]. Algorithms, 2015, 8(3): 743?753.
[24] TóTH R, SANANDAJI B M, POOLLA K, et al. Compressive system identification in the linear time-invariant framework [C]//Proceedings of the 50th IEEE Conference on Decision and Control and European Control Conference (CDC-ECC). Orlando, Florida, USA, 2011: 783?790.
[25] BLUMENSATH T, DAVIES M E. Gradient Pursuits [J]. IEEE Transactions on Signal Processing, 2008, 56(6): 2370?2382.
[26] 丁鋒. 系統(tǒng)辨識(shí)(2): 系統(tǒng)描述的基本模型[J]. 南京信息工程大學(xué)學(xué)報(bào)(自然科學(xué)版), 2011, 3(2): 97?117. DING Feng. System identification (Part B): basic models for system description [J]. Journal of Nanjing University of Information Science & Technology (Natural Science Edition), 2011, 3(2): 97?117.
[27] LI Xiangli, DING Ruifeng, ZHOU Lincheng. Least squares based iterative identification algorithm for Hammerstein nonlinear systems with non-uniform sampling [J]. International Journal of Computer Mathematics, 2013, 90(7): 1524?1534.
[28] WANG Dongqin, DING Feng. Least squares based and gradient based iterative identification for Wiener nonlinear systems [J]. Signal Processing, 2011, 91(5): 1182?1189.
[29] ELAD M. Sparse and redundant representions: From theory to applications in signal and image processing [M]. New York: Springer-Verlag, 2010: 20?62.
(編輯 陳愛(ài)華)
Parameter and time-delay estimation for MISO dynamic adjustment systems based on orthogonal matching pursuit iterative algorithm
LIU Yanjun1, 2, TAO Taiyang1, 3, DING Feng1, 2
(1. School of Internet of Things Engineering, Jiangnan University, Wuxi 214122, China;2. Key Laboratory of Advanced Process Control for Light Industry (Ministry of Education),Jiangnan University, Wuxi 214122, China;3. Henan Costar Group Company Limited, Nanyang 473000, China)
For multiple-input dynamic adjustment systems with unknown time delays, based on the characteristics of sparsity of the over parameterized parameter vector, and by combining the orthogonal matching pursuit algorithm with the hierarchical iterative idea, an orthogonal matching pursuit iterative (OMPI) identification algorithm was presented. This algorithm can estimate the parameters, time delays and some of the orders of the multiple-input dynamic adjustment systems with limited sampled data. The results show that the proposed method can effectively reduce the measuring cost, and improve the identification efficiency since it requires only a small number of sampled data compared to conventional identification methods such as the least square iterative (LSI) method. The proposed algorithm can effectively estimate the parameters and time delays of such sparse systems.
system identification; parameter and time delays estimation; compressed sensing; orthogonal matching pursuit algorithm; limited sampled data
10.11817/j.issn.1672?7207.2017.02.017
TP273
A
1672?7207(2017)02?0389?06
2016?04?05;
2016?07?25
國(guó)家自然科學(xué)基金資助項(xiàng)目(61304138);江蘇省自然科學(xué)基金資助項(xiàng)目(BK20130163)(Project(61304138) supported by the National Natural Science Foundation of China; Project(BK20130163) supported by the Natural Science Foundation of Jiangsu Province, China)
劉艷君,博士,副教授,從事系統(tǒng)辨識(shí)、自適應(yīng)控制等研究;E-mail:yanjunliu_1983@126.com