Proceedings of the 2006 IEEE International Symposium on Intelligent Control Munich, Germany, October 4-6, 2006 ThBO4.4 On Convergence of the Linear Extended State Observer Dongchul Yoo, Stephen S. -T. Yau, Fellow, IEEE, and Zhigiang Gao, Member, IEEE Abstract— Motivated by the gap between theory and practice in control problems, Linear Active Disturbance Rejection Con- trol (LADRC) addresses a set of control problems in the absence of precise mathematical models. LADRC depends on the quick convergence of a unique state observer, known as the extended state observer, proposed by Han. This paper shows numerically the tracking condition and the absolute error estimation of this “observer for a class of nonlinear and uncertain motion control problems. 1, INTRODUCTION |A feedback control system is utilized to manipulate the ‘output of the plant so that it follows a reference signal, In a digital implementation, the controller collects samples through the sensor, compares it to the reference and ‘computes the corresponding input to the plant based on the control law. The state feedback control laws require access to all state variables. For those states that are not directly measured, their estimates, obtained from a state observer, are used, Consequently, observer design has become a key factor in control design. Luenberger [1]-(3] introduced the state observer for linear systems, known as the Luenberger Observer, For nonlinear system, several methods have been proposed. Misawa et al [4] surveyed some of these methods. The performance of these observers and the resulting control system largely depend on the accuracy of the mathematical model of the plant, which poses a practical ‘concer. To address this issue, an ingenious observer, known as Extended State Observer (ESO), was proposed by Han [SHIT], where the states as well as the uncertainties in the plant are estimated, This allows the controller to actively ‘compensate for the uncertainties, and it led to the Active Disturbance Rejection Control (ADRC) [5]-[14]. By using a linear feedback instead of a nonlinear one, Gao [8]-{13] proposed the Linear Active Disturbance Rejection Control (LADRC) for the Single-Input Single-Output (SISO) nonlinear uncertain system. LADRC is easy to use and to tune because it has only two tuning parameters, namely, the closed loop bandwidth and the observer bandwidth, However it requires that the ESO converges quickly. The objective of this paper is to find the LADRC’s ob- server design condition in discrete time to make the estimate Dongehul Yoo is with Department of Mathemsaties, States and Com puter Science (M/C 249), University of Mines at Chicago, Chicago. TL {60607 USA (deyooGmath wie ed), Stephen 8. -T. Yau is with Department of Mathematics, Statistics and Computer Scioce (MUC 249}, Unwesity of Minis at Chicago, Chicago, 60607 USA (yautbuic edu). ‘Zhigiang. Gao is with Department of Hlecwical and Computer Ex gineenng. Cleveland. State University. Cleveland, Ohio 48115 USA (a gaoticceorp, 0-7803-8798-3/06/$20.00 ©2006 IEEE ‘converge to the real signal quickly, 2. LADRC BACKGROUND 21. Practical Motion Control Problem Motion contol is a broad term used to describe an open- oop or closed-loop electromechanical system wherein things are moving. Such a system typically includes a motor, ‘mechanical parts that move, and feedback sensors. Unlike classical control theory which is based on Laplace transform representations, modem control deals directly with systems in ordinary differential equation. In a typical application using a motor as the power source, the plant equation of motion can be described as HO = F(t, ¥), HO), wi) + d(hule) ao where y(t) is the position output, u(t) is the voltage to the power amplifier, b(t) is a time-varying coefficient, and tu(t) represents the unknown external disturbance such as vibrations and torque disturbances, We assume no noise is presented to the measurement of output y(t}. The friction, the effect of inertia and various other nonlinearities in a ‘motion system are all represented by the uncertain function J (>), which is generally a time-varying function. The math- ‘ematical model of the function f(:) is usually unavailable in practice. In this paper, we assume thatthe uncertain function F(t,y(t), (6), 2o(t)) on (1) is continuous and differentiable, and it satisfies [s(t v(t), (2), w(e))] < ka @ and 4 re yd.al 5 Fhlbsvlt) HO) w(0)] < o o where &, and ky are known constants, Practically speak- ing, the ranges of /(t,y(t),U(¢),w(®)) and 6(¢) are usually known but their analytical expressions are hard to come by. For the sake of simplicity, consider the plant dynamies in (1) normalized at b(t) = 1 and let f(-) absorb the discrepancy (b(E) — Lule). Let 2a(t) = y(t), za(t) = y(t), and alt) = Ft, 2(@), 2a(¢), w(@), u(o). wo ‘Then, a state space description of (1) is a(t) = a(t) a(t) ul) = alt) + u(t) ) = z(t) 1645 ‘shore consed stile lo: Indian Intute of Technology- Goa, Downloaded on December 29,2020 at 1609: UTC tar IEEE Xplore, Restos aolLet z9(l) = a(t) be an additional state vatiable in ($) and let A(t) = a(t), then (5) can be rewritten as ng =H Bt) = asl) + ule a(t) = hit) © ut) = n(t) ‘The key here is that the state augmentation in (6) allows a(t) 10 be estimated as a state 23. The control problem formulation takes a sharp tur here: instead of tying to find Fé, y(0, GO, 0(0), wld) in system identification, estimates it and compensates for it in real time! This is the basis of ADRC. 2.2. LADRC Tuning Method Gao's tuning method can be described as follows: Write (6) as (0) = Ax(t) + Bu(t) + Eh(2), uid) = Cx(0), 9 where o 10) oO A=|001|,B=/1],c=[1 0 0] 000, ° andE=[0 0 1)”. Note that | J? denotes transpose. “The state space observer, denoted asthe linear extended state observer (LESO), of (7) is constructed as HE) = Aa(t) + Bult) + Lit) — a(t), gy = Cat), and Lis the observer gain vector, which ean be obtained Using any Known method such asthe pole placement tech- igo, 8) L=[H 8 6) ® With the state observer properly designed and the controller is given by u(t) = —29(t) + uolt) role) = ky(v(t) — (0) Where ky is the gain of the derivative controller, ky is the gain of the proportional controller, and v(t) is the desired ‘uajectory of the position, With the PD gains chosen as heat) ke = 26, and ky =u, ay Where wand € are the desired closed-loop bandwidth and damping ratio. The closed-loop transfer function G(s) in Laplace transform is approximately a standard second order transfer function hy wi? Piksth + Mus bu, The ratio € can be conveniently set to unity to avoid any overshoot in the response and to allow the closed loop bandwidth, w, the only tuning parameter to be adjusted in implementation, Furthermore, if the observer gains are chosen as Gals) = ay a3) the characteristic polynomial of the observer is Dols) = 884 fis? + Bas Bs = (s+ wo) (ld) which means that the observer bandwidth, w,, is the only tuning parameter in the observer, Therefore, the number of tuning parameters in ADRC reduces to two. 3, NUMERICAL APPROACH TO THE CONVERGENT CONDITION IN DISCRETE TIME ‘To approximate the error between state estimate (2) and the actual state (2), let e(t) = 2(¢) ~ 2(¢) where e(t),2(¢) and 2(¢) are 3 x 1 column vectors. Then, from (7) to (8), the system for e(t) becomes &(0) = (A= LO)e(0) + BAC). as) ‘As shown on (15), the control term u(t), which contains the closed loop bandwidth, is eliminated on the analysis of extimate’s error, So, we need to consider only the ob- server bandwidth for the error approximation, By definition, dag llx(®)—2(0)la = Of and only if Jn |le(6)]2 = 0. The ‘onvergent condition means that for any" positive number € there exists fo > 0 such that le(4)|2 < efor every ¢ > to. In a view of ESO with digital implementation, we will find an ‘observer design condition in discrete time. By the numerical approach using Euler's forward method, the system (15) can bbe discretized. Assume that A, > 0 be the time step size for . The discretized form of (15) in time space becomes (E+ AKA Le)elk = (1+ AA ~LC))*e(0) A] + AEA ~ 1 i +A: UF ALA LE) “EAI 1}, U6) where 1-3Aw. Ae 0 ° (+ A(A-Le)) =| -3duw2 1 Ar ol, Awe 0 1 1 and k = [t4/A,). As long as the right hand side of (16) approaches to zero as is increased, it means that e[k] ‘converges to zero. Assume the initial error e(0) is bounded. Theorem 1: For mattix M only if p(M) <1 mim M* = 0 if and ‘Thus, the spectral radius of (1-4 A¢(A —LC)) must be less than 1 such that (1+ A;(A—LC))*e 0] exponentially decays independent of 0) as k +00. The spectral radius of (I+ AMA ~LC)) is (I+ AA LC) =|1- Awe an Ut should satisfy that |1 — Ajw| < 1. Therefore, the first ‘constraint for the convergent condition requites 0< uy < (2/An), cas) 1646 ‘ushorize consed so tried lo: Insian Intute of Technology- Goa, Downloaded on December 29,2020 at 1609: UTC tar IEEE Xplore, Restos aolAssumption I: Assume k > no where no isa sufficiently large natural number such that |\(I~ A,(A —LC))*C\], is less than (/3) for any positive number « and a fixed constant C From the assumption that ¢{0] is bounded, the upper bound ‘of the first term of (16) can be deduced as 1+ AA - Le)}*e(o}l], < (€/3) a9 ‘Therefore, the error term e[f] at time step k is dominated by the second term of (16) as fk 00. k llelKll-< Ae 33 + Ar(A - LO)! “EAfi — 1] +(€/3). For the sake of simplicity, introduce 3 1 vector G with its row components, Gj, Gi, and Gi such that (20) (4 A(A =[a @ &]". en ‘Applying (21) to assumption 1, we can get ec < (/3) 2) 4 fixed constant C, Since k is greater than no, applying the assumption 1 leads the first term of (20) to Jase “cali — 1) a Sevan aya Sema no i 2 Gt‘(nfi -1)) jor Fe Han $ eons) se ws ‘The boundedness of ae Se Tire tale | is easily proved by the componentowise triangle inequality (Gefer to lemma 1). Define the notation hz as kono-1Si 0 for all ' >0, Thus, applying the component-wise triangle inequality to the inside ‘of 2-norm simplifies the matrix power series term of (23) as 26) YS eae =a} thon bea HY mytaz ys iG —2}m-+4 A — mA S im? + (¢= 2)m-+1) (mye? a8 (Sop ( (mt (MER 4 (1m (ear me m9 Which means that the upper bound depends on the choice of ‘mand sampling time A,, but it is independent of raising the power i Case 2; Lem<2 ‘The radius (1 — m) of convergence is negative, and so raising the power of (1 — m) alters its sign in turn. So, applying the component-wise triangle inequality derives the ‘upper bound of the norm of the matrix power series term of, (23) as | $5 ctHae—ap) sme Iino sR — my2ag] KG 2)m-+ 1) my 2A] EMD? 4 (f — 2} +1) (Lm) , - 68 1) 202 < (2-4 1)0m— 2A, 155 | (UE Bm? + — 2)m 4-1) (m= 1)? (iF DL my? 0 = H@—2m+IA—m) Be as) laa, GEMED nt + (i —2)m +1) =m)? 2 |), 1647 ‘ushorize conse uke: nian Intute of Technlogy- Goa, Downloaded on Decomber 282020 at 1608: UTC tar IEEE Xplore, Reston apphr - [ol =1)4))) | |e (2+ dak — [O(m—))!) o(m—1))/ J], Mw 28) where O-notation is used for an asymptotic upper bound If we compare the steady state behaviors between (27) and (28) under the condition where they have the same radius of convergence, then asymptotically, O((1 — m)*) terms and O((m — 1}*) terms decay in the same rato, However, the case for 0 < m < 1 has smaller value than the one for 1 < m <2 in the worst case, By this reason, we consider the case only for 0 oc and 0 fast The design objective isto Fate the Toad one revolution in one second with no overshoot and the contol signal has the physical constraint [ul < 3.5 volt The simulation is performed by Simulink using odel (Goulet) witha fixed step of lms sampling time. A step torque disturbance of 10% of the maximum torque is added at ¢ — 0. “Three diferent velocity profile are used and compared on the simulation, Fig. 1 shows those profiles When the motion (2) Teapeoidal rosie (©) S-cuve prole (@) Smooth S-curve prose Fig. 1. Velocity Poties profile shifts from 1{@) to 1(@), We expect the changes of the acceleration are getting smooth, so (a) has relatively larger jetk and 1(@) has relatively smaller jerk. On the frst simulation, the value of the closed loop bandwidth w is changed from 20 (rs) to 166 (/) and the observer bandwidth &» is changed from 20 (es) to 1000 (Ws). We fixed the absolute error tolerance as 0,01, Define the convergence time (t) as te = min{ te, | llella [ta/ Ae As > O and a fixed ¢ } Fig. 2 shows the convergence time (1,) needed to track () Convergence time for Surv profil (©) Convergence tie for Smooth S-carve profile. Fig. 2. Esimat's convergence time to wack the eal signal i the specified absolteerorlleranee 001 the real signal y(t) in absolute error tolerance range on ‘cach velocity profiles. We can easily observe the convergent tuend. As observer bandwidth is increased up to the sampling frequency, it shortens the convergence time on overall fixed closed-loop bandwidth ranges. When the closed-loop band- width is on fixed low observer bandwidth ranges, it shows fan interesting result that it affects convergence time with considerable rate, But there are no significant changes on the faxed high observer bandwidth ranges as we expected. For the precise comparison of three profiles, we fixed the closed-loop 1649 ‘ushorize consed so tried lo: Insian Intute of Technology- Goa, Downloaded on December 29,2020 at 1609: UTC tar IEEE Xplore, Restos aol‘bandwidth as 150 (r/s), and altered the value of the observer bandwidth, The result is shown on fig. 3, The smooth S- ‘Doser bani Fig. 3. Convergence time comparison of tee profiles with the fixed ‘led bandwidth 150 (5). ‘curve shows the relatively faster convergent time, whereas trapezoid profile has relatively slower convergent time. As wwe can see from (31) and (32), I, directly affects the ‘convergence of the estimate. From the problem formulation (1) and (3), we can deduce that the trajectory position (y) traces the reference input, and so /, directly varies with the magnitude of the jerk of the reference input. From the assumption for hy ils values may vary as time goes on. When ‘hy hits the bounds in (31) or (32), the estimate (=) starts to trace the actual signal within the desired absolute ferror tolerance, Thus, an effect of a larger jerk requires rmore time to decay away. The variation ofthe absolute exror ‘Desens banish a) Fig. 4. Contergence time comparison on the vatiation of absolute eror tolerance wih he fined losed bandwidth 150 (8) on Surv profi, tolerance also affects the convergence time as shown in (31), (32) and in fig. 4 5. Conciusion In this paper, numerically sufficient convergence condi- tions are derived for the linear extended state observer in dis- ‘rete time domain, under the assumption that the uncertainty and ils derivative are bounded. The conditions show that the selection of observer bandwidth must take into consideration ff the upper bound of the uncertainty (related to jerk) and the sampling period. The resulls are verified in a simulation study of a practical motion control application, In particular, the simulation results show that the amount of jerk in the motion profile has marked effects on the convergence of the ‘observer, as predicted by the mathematical derivation, Since ‘we did not consider the closed-loop bandwidth because the ‘control term is eliminated on the analysis of error, the closed oop bandwidth must be invested for the best control of the ‘uncertain system within the consideration of convergence condition in theorem 2. 6. ACKNOWLEDGEMENT D. Yoo wishes to express his sincere gratitude to Qing ‘Zheng who by offering the practical examples and assistance. REFERENCES 1D. G. Laenberger, “Observing the sate of a linear system,” IEBE rans, Mi. Elecion, vol MUS, pp. 74-80, 1964 ‘Observers (or mulivariale ystems,” IEEE Trans, Automat ont, vol AC-11, no. 2, Api 1966, “Am intoduction to observers” IEEE Trans, Automat, Contr, vol, AC-16, pp. 592-802, December 1971. Is] EA. Misawa and JK. Hedrick, "Nonlinear observer sate of the tut suey,” Trans, ASME J. Dynamic Systeme Meaeurement Cont, ol II, pp. 344-852, 1988 [51 1 Han, “Control theory: I it a theory of model or contol?” Syet. Si dnd Mathematical St, vol.9, 26. pp. 328-335, 1989, in Chinese) "Nonlinear sae exror feedback control” Control and Decision vol 10, 0.3, pp. 221-295, 1994, (la Chinese). fended sate observer for a kind of uncersin systems Contol and Decision, vl 10, 6.1, p. 85, 1995, ls Chinese). [81 Z Gro, "Reformulation of motion contol problems anda new pespective on esearch” Presented a! the [th aniversary(2004) com ference forthe Mathematice Intute at East China Normal Univers dune 24 [9] —. “Sealing and bandwidh-parameterization bated controler tn ing” Proc. ofthe 2007 American Control Conference, pp. 4989-96, Jone 2003, denver, Co Uo} 2 Gao, ¥. Haung. and J. Han, “An alternative paradigm for contol sytem design” Proc. ofthe 20] IEEE Conference om Decision and Contos December 2001, [11} 2. Gao, SH, and F Jiang, “A novel motion con! design approach based on active disturbance rejection.” presented at 2001 IEEE Com ference on Decision and Control, 2001 [12] W. Wang and Z_ Guo. “Oa comparison of advanced sate observer Assign technique,” Proc. ofthe 2003 American Control Conference, pp. 4754-4759, June 2003, [13] Z Gao and R Rbinchan, "Theory vs. practice: The eballenges from indosty” Proc. ofthe 2008 American Control Conference, pp. LI 1349, Tune S0July 22004, boston MA. [U4] 3. Han, “Auto disturbances rejection controller and its applications Contol and Decision, Wl. 13, 2.1, p19, 1998, (la Canes). 1 Bl to nm 1650 ‘ushorize consed so tied: Indian Intute of Technology- Goa, Downloaded on December 29,2020 at 1609: UTC tar IEEE Xplore, Restos aol