Design of Adaptive Backstepping controller for Systems with Mismatched Perturbations to Achieve Asymptotical Stability Chien-Wen Chung Department of Electrical Engineering, Far East University, Tainan, Taiwan e-mail: cwen.chung@msa.hinet.net Abstract – Based on the Lyapunov stability theorem, a methodology of designing the block backstepping controllers for a class of multi-input multi-output (MIMO) systems with matched and mismatched perturbations is proposed in this paper. Some adaptive mechanisms are embedded in both the virtual input and the robust controllers so that not only are the mismatched perturbations suppressed, but also the knowledge of upper bound of perturbation is not required except those of the input uncertainties. Finally, the feasibility of the proposed methodology can be verified by controlling of a separated excited DC motor (SEDCM). Key Words – Lyapunov stability, backstepping, SEDCM. 1. Introduction Backstepping is a recursive procedure that interlaces the choice of a Lyapunov function with the design of feedback control [1]. One advantage of this technique is that it can circumvent the problem of mismatched perturbations and achieve the property of asymptotical stability. The other advantage is that it can exploit some extra flexibility that exists with lower order systems. In recent years, some researchers used MIMO backstepping technique to design robust controllers for systems with perturbations. Wu & Zhou [2] designed an output feedback controller for a class of MIMO nonlinear systems to achieve uniform ultimate boundedness stability. Lin and Qian [3] proposed the partial state and output feedback controllers, Liu et al. [4] designed a H ∞ controller and applied it in a class of chemical plants. Yao and Tomizuka [5] and Liu et al. [6] designed the backstepping controllers for a class of MIMO nonlinear systems. Although the works proposed in [3,5,6] can be employed for MIMO systems to achieve asymptotical stability; however, some controlled systems have to be in the strict feedback form [3,4], some control schemes [3,4] need the existence of the positive definite and proper Lyapunov function. In addition, only the au- Yaote Chang Department of Electrical Engineering, Kao Yuan University, Kaoshuing, Taiwan e-mail: yat.chang@msa.hinet.net thors of [5,6] considered the uncertainties of input gain; however, the knowledge of the upper bound of perturbations is required. Due to the limitations of each of the methodologies mentioned above, in this paper we proposed a design methodology of adaptive block backstepping controllers for a class of MIMO systems with mismatched perturbations as well as matched input uncertainty to solve regulation problems. The advantages of this proposed methodology are that not only are asymptotical stability can still be achieved, but also the knowledge of the upper bound of perturbations is not required except those of input uncertainties when designing the controllers. Finally, a SEDCM is demonstrated for showing the applicability of the proposed technique. 2. System Descriptions and Problem Formulations Consider the following MIMO nonlinear system: (1a) ẋ1 = f1 (t, x1 ) + G(t, x1 )x02 + Θf2 (t, x1 ) h i ẋ2 = f3 (t, x) + ∆f4 (t, x) + B(t, x) + ∆B(t, x) u, (1b) T T x1 xT2 where x = ∈ Rn , n = n1 + n2 represents measurable state vector, x1 ∈ Rn1 , x2 = h iT T T ∈ Rn2 , x02 ∈ Rn1 , x002 ∈ Rn2 −n1 , n1 ≤ x02 x002 n2 . The vector fi (t, x1 ) ∈ Rn1 , i = 1, 2 and f3 (t, x) ∈ Rn2 are known nonlinearities, where vector fi , i = 1, 2 is once differentiable. The known matrix G(t, x1 ) ∈ Rn1 ×n1 and B(t, x) ∈ Rn2 ×n2 are nonsingular, where G is once differentiable. u ∈ Rn2 is the control input. The unknown constant matrix Θ ∈ Rn1 ×n1 is mismatched external disturbance. The matrix or vector ∆(·) denote the unknown model uncertainties and/or parameter variations in (·). The objective of this paper is to design the virtual input x02 and the control effort u so that the system’s state x1 and x002 are able to track the desired piecewise continuous, bounded, signal y0 ∈ Rn1 and y00 ∈ Rn2 −n1 asymptotically respectively, where y0 is twice differentiable, y00 is once differentiable. 3. Design of the Virtual Input and Robust Con- troller In order to show that the state x1 can track the signal y0 asymptotically, one can design the virtual input as φ(t, x) = x02 − ψ(t, x1 ) ψ(t, x1 ) = −G−1 [Ke1 + f1 + Θ̂(t)f2 − ẏ0 ], (2) Theorem 1 Consider the system (1). The lumped T , ζ1 ∈ perturbations are defined as ζ , ζ 1 T ζ 2 T Rn1 , ζ 2 ∈ Rn2 −n1 , ζ 1 , ∆f40 + ∆B0 u + Ġζ−1 (Ke1 + f1 + Θ̂f2 − ẏ0 ) +G−1 (KΘf2 + ḟ1ζ + Θ̂ḟ2ζ ) 0 n1 where e1 = x1 − y , [e1 e2 · · · en1 ] ∈ R is the tracking error and K ∈ Rn1 ×n1 is a constant matrix designed in a way such that Re[λ(K)] > 0. The adaptive gain Θ̂(t) , [θ̂ij (t)] ∈ Rn1 ×n1 , 1 ≤ i, j ≤ n1 in (2) is defined as ˙ θ̂ij (t) = f2j n1 P pid e1d , ζ 2 , ∆f400 + ∆B00 u, iT iT h h T T ∆f4 , ∆f40 T ∆f400 T , ∆B , ∆B0 ∆B00 , and ζ can be satisfied by the constraints (3) d=1 kζ(t, x, u)k ≤ r0 + r1 kxk + r2 kuk, n1 ×n1 where f2 , [f21 f22 · · · f2n1 ]. P , [pjd ] ∈ R ,1 ≤ j, d ≤ n1 is a symmetric and positive definite matrix and can satisfies eT1 Qe1 ≥ 0, Q , PK + KT P. Before designed the backstepping controllers, we can also divide dfi /dt, i = 1, 2 into dfi /dt , dfin /dt + dfiζ /dt, where dfin /dt is the nominal part of dfi /dt, and dfiζ /dt is the part of dfi /dt which contains pertur∂fi 0 i bations, i.e., dfin /dt = ∂f ∂t + ∂x1 (f1 + Gx2 ), dfiζ /dt = ∂fi Similarly, one can also have dG−1 /dt , ∂x1 Θf2 . −1 −1 dG−1 n /dt+dGζ /dt and dGn /dt = Gx02 ), dG−1 ζ /dt = −1 ∂G ∂x1 Θf2 . 00 T T f3 ] , T [f30 ∂G−1 ∂t −1 + ∂G ∂x1 (f1 + 0 −1 u11 = −f30 − Ġ−1 n (Ke1 + f1 + Θ̂f2 − ẏ ) − G h i ˙ + Θ̂ḟ − ÿ0 , K(f1 + Gx02 − ẏ0 ) + ḟ1n + Θ̂f 2 2n x002 00 1 − r2 kB−1 k ` 4. Stability Analysis = ẋ1 − ẏ0 = −Ke1 + Θ̃f2 + Gφ. V1 = (6) (7) 0T 0 1 T e1 Pe1 + θ̃ θ̃ , 2 (8) , if kzk = 6 0 , r̂` (t0 ) = 0, (5) if kzk = 0 where α is a designed positive constant. = f1 + Gψ + Θf2 + G(x02 − ψ) = −Ke1 + Θ̃f2 + ẏ0 + Gφ, where Θ̃ = Θ − Θ̂ , [θ̃ij ] ∈ Rn1 ×n1 , θ̃ij = θij − θ̂ij , 1 ≤ i, j ≤ n1 , are the adaption errors of unknown constant Θ. Using (6), the dynamics of the tracking error e1 is given by kun k + kB−1 k + η 1 kB−1 k α−1 kzkkxk 0 ẋ1 Now one can choose a Lyapunov function as e2 = − y , r2 < and η > 0 are the designed positive constants. The adaptive rules of r̂` (t), 0 ≤ ` ≤ 1 in (4) are given by r̂˙` (t) = According to (2), (1a) can be rewritten as ė1 u12 = −GT e1 , u2 = −f300 + ẏ00 , ( −1 6 0 − (t) + ρ(t) Bkzkz , if kzk = , us = 0, if kzk = 0 h iT r2 z = φT e2 T , ρ(t) = Proof: Decomposes the vec- tor f3 as f3 = where f30 ∈ Rn1 and 00 n2 −n1 f3 ∈ R . The backstepping controllers are designed as u = un + us , (4) T where un = B−1 uT1 uT2 , u1 = u11 + u12 , (t) = r̂0 (t) + r̂1 (t)kxk, where r` , ` = 0, 1, are unknown positive constants and r2 is given in (4). If the virtual input and the backstepping controllers are designed as (2) and (4) respec T tively, the state trajectory of e = e1 T e2 T will approach zero asymptotically; and the adaptive gains θ̂ij , 1 ≤ i, j ≤ n1 , r̂` (t), 0 ≤ ` ≤ 1 are all bounded. 0 where θ̃ , [θ̃1 θ̃ 2 · · · θ̃n1 ]T ∈ Rn1 n1 ×1 , θ̃ i , θ̃i1 θ̃i2 · · · θ̃in1 ∈ R1×n1 , 1 ≤ i ≤ n1 . Differentiating (8) along the trajectory of (7) yields T 1 V̇1 = − eT1 PK + KT P e1 + f2T Θ̃ Pe1 + eT1 Gφ 2 0T (9) +θ̃ θ̃˙ 0 . ˙ ˙ According to (3) and θ̂ij = −θ̃ij , 1 ≤ i, j ≤ n1 , one can r̃˙` = r̂˙` , ` = 0, 1, one can obtain the time derivative of (14) as verify that T f T Θ̃ Pe1 2n1 P f2j θ̃1j = = n1 P j=1 n1 P d=1 n1 X f2j θ̃2j ··· j=1 n1 P p1d e1d f2j θ̃1j n1 X p2d e1d n1 P ··· n1 X p1d f1d + n1 X f2j θ̃n1 j ˙ θ̃1j θ̃1j − j=1 n1 X f2j θ̃2j j=1 j=1 n1 X pn1 d e1d T 1 V̇2 = − eT1 PK + KT P e1 + f2T Θ̃ Pe1 + eT1 Gφ 2 1 X 0T +θ̃ θ̃˙ 0 + zT ż + α r̃` r̃˙` T `=0 d=1 d=1 +···+ f2j θ̃n1 j j=1 d=1 j=1 =− n1 P n1 X 1 = − eT1 Qe1 + eT1 Gφ + φT (ζ 1 + u12 + B0 us ) 2 1 X +eT2 (ζ 2 + B00 us ) + α r̃` r̃˙` p2d e1d d=1 `=0 pn1 d e1d d=1 n1 X n1 X ˙ ˙ θ̃2j θ̃2j − · · · − θ̃n1 j θ̃n1 j j=1 j=1 0T ˙ = −θ̃ θ̃ 0 . 1 X 1 = − eT1 Qe1 + zT ζ + Bus + α r̃` r̃˙` 2 `=0 1 ≤ (10) X 1 − eT1 Qe1 + kzkkζk + zT Bus + α r̃` r̂˙` 2 `=0 1 ≤ − eT1 Qe1 + kzk( 2 Using (10), (9) can be rewritten as V̇1 = −eT1 Qe1 /2 + eT1 Gφ. (11) −kzk( 0 According to (4), one can obtain u11 + u12 = B un T and u2 = B00 un , where B , B0T B00T , B0 ∈ Rn1 ×n2 , B00 ∈ R(n2 −n1 )×n2 . From (1) and (4), the time derivative of φ is φ̇ = ẋ02 + Ġ−1 Ke1 + f1 + Θ̂f2 − ẏ0 ˙ + Θ̂ḟ − ÿ0 +G−1 Kė1 + ḟ1 + Θ̂f 2 2 = f30 + ∆f40 + B0 (un + us ) + ∆B0 u −1 0 +(Ġ−1 n + Ġζ ) Ke1 + f1 + Θ̂f2 − ẏ h +G−1 K(f1 +Gx02 +Θf2 − ẏ0 )+ ḟ1n + ḟ1ζ i ˙ + Θ̂(ḟ + ḟ ) − ÿ0 +Θ̂f 2 2n 2ζ = ζ 1 + u12 + B0 us (12) = ẋ002 − ẏ00 = f300 + ∆f400 + ∆B00 u + B00 (un + us ) − ẏ00 = ζ 2 + B00 us . (13) In order to prove the state trajectory of e will approach zero asymptotically, one can use the Lyapunov function V1 given by (8) and define another Lyapunov function candidate as 1 V2 = V1 + (zT z + α 2 1 X r̃`2 ), ` r` kxk + r2 kuk) `=0 r̂` kxk` + ρ) + `=0 1 X (r̂` − r` )kzkkxk` `=0 1 = − eT1 Qe1 + kzk(−ρ + r2 kuk) 2 n o 1 ≤ − eT1 Qe1 + kzk − ρ + r2 [kun k+kB−1k (ρ + )] 2 1 = − eT1 Qe1 2 h i +kzk − ρ(1 − r2 kB−1 k) + r2 (kun k + kB−1 k) 1 = − eT1 Qe1 − ηkzk ≤ 0. 2 and the time derivative of e2 is ė2 1 X 1 X (14) `=0 where z is defined in (4). r̃` , r̂` − r` , 0 ≤ ` ≤ 1 are the adaptive errors of the unknown constant r` . Using (4), (5), (12), (13), the result of (11) and noting that The preceding equation indicates that z and state e1 will approach zero asymptotically. On the other hand, z reaches zero means that state e2 and φ will reach zero asymptotically. Therefore one can conclude that T the state variable e = e1 T e2 T will approach zero as t → ∞. Since V2 > 0 and V̇2 ≤ 0, this indicates θ̂ij , 1 ≤ i, j ≤ n1 , ĝ` (t), 0 ≤ ` ≤ 1, are all bounded. 5. Application of Controlling SEDCM In order to demonstrate the applicability of the proposed control scheme in this section, we apply the backstepping controller to control a SEDCM dynamic system [7] are described as dw 1 = (Te − Bw − TLN − ∆TL ) dt J 1 dia ua = (−RaN ia − E − ∆Ra ia ) + dt La La 1 dif uf = (−Rf N if − ∆Rf if ) + , dt Lf Lf (15) where E = kif w and Te = kia if = E w ia . The nomenclature of each variable in (15) is listed in Table 1 [7]. ∆Ra = Ra − RaN , ∆Rf = Rf − Rf N and ∆TL = TL − TLN are the unknown deviations from the nominal values RaN , Rf N and TLN of Ra , Rf and TL , respectively. According to the notation of (1), one can rewrite (15) in the form as (1), i.e., x1 , w, x02 , ia , x002 , if , E L f1 (t, x1 ) , −(Bx1J+TLN ) , G(t, x1 ) , Jx , Θ , −∆T , J 1 −(RaN x02 +kx1 x00 2) , f300 (t, x) , La 0 −∆Ra x2 −∆Rf x00 2 , ∆f400 (t, x) , , La Lf T , [ua uf ] , f2 (t, x1 ) , 1, f30 (t, x) , −Rf N x00 2 , Lf 0 y = wm , ∆f40 (t, x) , y00 = if m , u B, " 1 La 0 0 # 1 Lf , ∆B , 0 0 0 0 . Rated power Rated speed Field voltage Armature resistance Field resistance Inertia 3.7kW 1800rpm 240V 1.2Ω 60Ω 0.208kgm2 Rated voltage Rated torque Armature inductance Field inductance Motor constant Damping coefficient 240V 18Nm 0.01H 60H 0.3Nm/A2 0.011kgm2s−1 The motor parameter data are listed in Table 1 [7]. The design parameters are chosen to be (K, Q, η) = (3.5, 33, 0.01). Reference speed is initially set up in wm (y 0 ) =1800r/min and is later modified to 1840r/min at 2 sec. A sudden change of mismatched disturbance θ (load torque uncertainty ∆TL = 14N ·m) is applied at t = 6 sec. Fig. 1 shows that the actual speed x1 achieved the robust performance in a short transient time even if mismatched load torque existed. Fig. 2 shows that the trajectory of x002 , which can track the y 00 robustly. Fig. 3 shows that the adaptive gain θ̂ almost converges to the true θ. The adaptive gain ĝ0 , which is bounded, is shown in Fig. 4. [1] H. K. Khalil, Nonlinear Control, Prentice-Hall, New Jersey, 1996. 6. Conclusion An adaptive backstepping controllers for a class of MIMO nonlinear systems with mismatched perturbations is successfully proposed in this paper. The advantages of the proposed control scheme are not only can achieve asymptotical stability, but also it can be applied to a more general perturbed systems than those in [7]. The simulation results of a SEDCM show the performance of tracking is robust even the mismatched perturbation exists. [4] X. Liu, A. Jutan, and S. Rohani, “Almost disturbance decoupling of MIMO nonlinear systems and application to chemical processes,” Automatica, Vol. 40, pp. 465-471, 2004. Table 1: Nomenclature of the SEDCM [7] TL : load torque B: damp coefficient if : motor field current w: motor speed k: back emf constant Lf : field inductance uf : field control input Te = kia if = E w ia : develop torque E = kif w: back emf ia : motor armature current Ra : armature resistance Rf : field resistance La : armature inductance ua : armature control input J: damp inertia Table 2: Parameters of the SEDCM [7] References [2] Y. Wu and Y. Zhou, “Output feedback control for MIMO non-linear systems with unknown sign of the high frequency gain matrix,” International Journal of Control, Vol. 77, No. 1, pp. 9-18, 2004. [3] W. Lin and C. Qian, “Semi-global robust stabilization of MIMO nonlinear systems by partial state and dynamic output feedback,” Automatica, Vol. 37, pp. 1093-1101, 2001. [5] B. Yao and M. Tomizuka, “Adaptive robust control of MIMO nonlinear systems in semi-strict feedback forms,” Automatica, Vol. 37, pp. 1305-1321, 2001. [6] X. Liu, G. Gu, and K. Zhou, “Robust stabilization of MIMO nonlinear systems by backstepping,” Automatica, Vol. 35, pp. 987-992, 1999. [7] Z. Z. Liu, F. L. Luo, and M. H. Rashid “Speed nonlinear control of dc motor drive with field weakening,” IEEE Transaction on Industry Applications, Vol. 39. No. 2, pp. 417-423, 2003. 1860 1840 1820 x1 y' 1800 1780 0 2 4 6 8 10 Figure 1: Trajectories of angular speeds x1 and y 0 . 0.5 x 2" y" 0.4 0 2 4 6 Figure 2: Trajectories of field current 8 x”2 10 and y”. 10 0 -10 -20 -30 -40 -50 -60 -70 -80 -90 0 2 4 6 8 10 Figure 3: Adaptive gain θ̂ and θ. 400 g0 350 300 250 200 150 100 50 0 0 2 4 6 Figure 4: Adaptive gain ĝ0 . 8 10