Techset Composition Ltd, Salisbury Doc: {IEE} CTA20130568.tex www.ietdl.org 1 6 Published in IET Control Theory and Applications Received on 7th September 2012 Revised on 1st October 2013 Accepted on 1st December 2013 doi: 10.1049/iet-cta.2013.0568 71 76 11 16 21 Q1 ISSN 1751-8644 Discrete-time non-linear state observer based on a super twisting-like algorithm Iván Salgado 1, I. Chairez 2, Bijnan Bandyopadhyay 3, Leonid Fridman 4 and Oscar Camacho 1 1 CIC, Instituto Politecnico Nacional, Mexico City, Mexico 2 Department of Bioprocesses, UPIBI, Instituto Politecnico Nacional, Mexico City, Mexico 3 SYSCON, Indian Institute oftechnology, Bombay, Mumbai, India 4 Engineering Universidad Nacional Autonoma de Mexic, Mexico City, Mexico E-mail: jchairez@ctrl.cinvestow.mx 81 86 91 26 31 36 Abstract: The properties of robustness and finite-time convergence provided by sliding mode (SM) theory have motivated several researches to deal with the problems of control and state estimation. In the SM theory, the super-twisting algorithm (STA), a second-order SM scheme, has demonstrated remarkable characteristics when it is implemented as a controller, observer or robust signal differentiator although the presence of noise and parametric uncertainties. However, the design of this algorithm was originally developed for continuous-time systems. The growth of microcomputers technology has attracted the attention of researchers inside the SM discrete-time domain. Recently, discretisations schemes for the STA were studied using majorant curves. In this study, the stability analysis in terms of Lyapunov theory is proposed to study a discrete-time super twisting-like algorithm (DSTA) for non-linear discrete-time systems. The objective is to preserve the STA characteristics of robustness in a quasi-sliding mode regime that was proved in terms of practical Lyapunov stability. An adequate combination of gains obtained by the same Lyapunov analysis forces the convergence for the DSTA. The problem of state estimation is also analysed for second-order mechanical systems of n degrees of freedom. Simulation results regarding the design of a second-order observer using the DSTA for a simple pendulum and a biped model of seven degrees of freedom are presented. 96 101 106 41 46 51 56 61 1 Introduction Nowadays, the advance in the development of microcontrollers with capabilities to solve advanced mathematical algorithms has allowed the implementation of complex control schemes without the use of personal computers. In this way, the requirement of control and state estimation strategies downloadable in these microcontrollers has oriented the research of control strategies in two principal frameworks. The first one analyses the discretisation problem of several control techniques and its sampled time dependences in the discrete-time domain. On the other hand, as the second framework, new control strategies have been presented to study the case of pure discrete-time systems. These techniques include the well-known sliding mode (SM) theory [1. In addition, the implementation of integration schemes could be an option to keep working with continuous strategies. However, the computational requirements are increased when this option is implemented. The SM theory has been widely applied to solve the state estimation problem. The main features obtained by SM are robustness against parametric uncertainties and finitetime convergence. These advantages are evident in contrast with the conditions required for non-linear systems when a classical structure for observation is used, like a Luenberger one. Indeed, the necessity of the exact mathematical description constitutes a drawback in this kind of algorithms. Following the idea of Luenberger structures, the so-called high-gain observers have been developed to deal with complex non-linear systems [2. Sliding motion is obtained by including a discontinuity term in the algorithm structure. The discontinuous injection must be designed such that the trajectories of the system are forced to remain on a prescribed submanifold (sliding surface) on the state space. The resulting motion in the surface is named as a sliding motion [3, 4. If the system has relative degree greater than one with respect to the sliding surface, high-order sliding modes (HOSM) such as the second-order sliding modes (SOSM) could be implemented preserving the main characteristics of classical SM [5, 6. Moreover, SOSM reduce the undesired chattering effect. In continuous-time, the super-twisting algorithm (STA) has been successfully analysed and applied like a robust exact differentiator [7, state estimator [8 or controller [9. Although, there exist several researches in the continuous domain, the concept of high-order discrete-time sliding modes (DSM) has not deeply studied. DSM have not received much attention as the continuous counterpart. The first ideas of DSM were introduced in [10, 11, where a quasi-sliding mode (QSM) is established for systems with relative degree one. In [12, the study of discrete-time single-input single-output (SISO) non-linear 111 116 121 126 131 66 136 IET Control Theory Appl., pp. 1 10 1 doi: 10.1049/iet-cta.2013.0568 The Institution of Engineering and Technology 2014
141 146 151 156 161 166 171 176 181 186 191 196 201 systems with relative degree bigger than one was considered. A new definition of the QSM regime was addressed in [13. In this paper, the motion of the system was restricted in a certain band around the sliding surface. In [14, 15 some researches have been reported for several classes of discrete-time linear systems. The idea of SOSM control in discrete-time systems was introduced in terms of a certain classes of discretisation-like Euler discretisation or finite differences [8, 16, 17. However, a complete stability study has not been deeply analysed in terms of Lyapunov stability. There are several generic methods for designing sliding mode observers (SMO). For continuous-time linear and non-linear systems, most of these observers are based on the equivalent control concept for handling disturbances and modelling uncertainties. In the same context, discretetime sliding mode control (DSMC) methods have also been developed for linear and non-linear systems [12, 18, 19. However, discrete-time sliding mode-based observer (DSMO) design has not received much attention, especially for non-linear systems. DSMO design for linear systems was given in [18. In [20, the concept of sliding lattice for discrete-time systems was introduced and DSMO was designed for SISO linear systems using the Lyapunov min max method. Besides, in several works the accuracy of the discretisation in SOSM controllers depending on the sampling period has been analysed [21, 22. In this paper, a discrete-time SOSM algorithm is proposed based on the well known STA. Stability of discrete-time super twisting-like algorithm (DSTA) is analysed with a quadratic Lyapunov function. Stability conditions for the DSTA were obtained with the solution of a linear matrix inequality (LMI) and the trajectories were confined into a boundary layer in the vicinity of the sliding surface and stays inside it forever obeying a QSM behaviour. The boundary layer magnitude is proportional to the square of the sampled period. A natural application for the DSTA as a second order discrete-time observer is analysed under the same strategy. This analysis was done using quadratic Lyapunov functions an LMI techniques. In the following section, the structure of the DSTA is given. The result regarding the convergence of the proposed algorithm is established in terms of a simple quadratic Lyapunov function in Section III. In Section IV, numerical examples are presented. The first example is based on the state estimation problem of a Furuta pendulum and a second one to estimate the states of a biped model with seven degrees of freedom. Finally, in Section V, some conclusions are established. 2 Discrete-time super-twisting-like algorithm The DSTA is composed by the following two equations in differences x 1 (k + 1) = ρ 1 x 1 (k) + τx 2 (k) τk 1 x 1 (k) 1/2 sign(x 1 (k)) x 2 (k + 1) = ρ 2 x 2 (k) τk 2 sign(x 1 (k)) (1) The sign(κ) function is defined as 1 if κ < 0 sign(κ) := 0 if κ = 0 1 if κ > 0 where κ R. Equation in (1) is designed as a slightly modification of an Euler discretisation from the continuous SOSM. The system (1) can be represented as (2) x(k + 1) = Ax(k) + B(k)sign(x 1 (k)) (3) where A R 2 2 and B(k) R 2 are ρ1 τ τk1 x A := B(k) := 1 (k) 1/2 0 ρ 2 τk 2 2.1 Main result The result obtained in this paper is stated in the following theorem. Theorem 1: Consider the non-linear system given in (1), with gains selected as k 1 > 0 and k 2 > 0 and if the following LMI A (P + P P)A (1 ϱ)p + Q 0 (5) has a positive-definite solution P = P > 0 for a given Q = Q > 0 and = > 0, then, the trajectories of the dynamic system given in (1) converge asymptotically to a ball B r centred at the origin B r :={x : x 2 < r} characterised with a radius where r = c 1 ϱ 0 <ϱ<1 c := δ 2 + 1 4 δ 2 1 Q 1 2 F δ 1 := δ 1 + τ 4 k 2 1 k2 2 z2 12 ω 1, δ 2 := k 2 2 τ 2 (τ 2 k 2 1 z2 12 ω 1 1 +z 22) δ 1 := τ 2 k 2 1 z 11 (7) ω 1 R +, Z=[ 1 +P, Z,,P R 2 2 z11 z Z := 12 z 12 z 22 Proof: The proof of the previous theorem is given in the appendix Lemma 1: If ϱ is selected as 0 <ϱ<1 such that ( λ max {A} )/( (1 ϱ)) < 1, the LMI expressed in (5) always has positive definite solution P = P > 0 given by P = (4) (6) (à ) k Q 1 à k (8) k=0 211 216 221 226 231 236 241 246 251 256 261 266 271 206 where x i (k) Rare the states, ρ i R +, ρ i < 1, k i Rfor i = 1, 2, are the gains to be designed for ensuring the convergence of (1) to the origin and τ>0 is the sampling period. with à := ϱ 1 A and Q 1 := ϱ 1 Q. This solution can be obtained because (5) can be represented as a classical Lyapunov discrete-time equation. Also, for a particular value of P and Q we can exactly get the corresponding value of ϱ. 276 2 IET Control Theory Appl., pp. 1 10 The Institution of Engineering and Technology 2014 doi: 10.1049/iet-cta.2013.0568
281 286 291 296 301 306 311 316 321 326 331 336 341 346 Proof: The equation in (5) can be represented as A PA (1 ϱ)p = Q A A with = P 1 P 1. Therefore this equation becomes in A PA (1 ϱ)p Q (9) Because ϱ := 1 ϱ>0, equation (9) can be divided by ϱ and get à Pà P = Q 1 (10) Equation (10) is already a discrete-time Lyapunov equation and the explicit solution is (8). The structure of the proposed algorithm is straightforwardly to apply it into a class of state estimator for second-order discrete-time non-linear systems. 2.2 Particular application: a discrete-time second-order observer 2.2.1 Scalar case: The proposed algorithm in this paper can be directly extended to design a discrete-time super-twisting observer (DSTO). This subsection discuses how this algorithm can be implemented in the estimation process. Consider the non-linear system given in x 1 (k + 1) = x 1 (k)+τx 2 (k) x 2 (k + 1) = x 2 (k)+τf (x 1 (k), x 2 (k), u(k)) y(k) = x 1 (k) (11) where x i (k) Rfor i = 1, 2 are the states of the non-linear systems, u(k) Ris the control action applied at the time t = τk, τ R + is the sampled period and f (,, ) : R 2+1 R is a non-linear function describing the dynamics of the discrete-time non-linear system. The algorithm used in the state estimation process of the non-linear system (11) is given by the following two equations in differences ˆx 1 (k + 1) =ˆx 1 (k)+τ ˆx 2 (k) τk 1 e 1 (k) 1/2 sign(e 1 (k)) k 3 e 1 (k) ˆx 2 (k + 1) =ˆx 2 (k)+τ ˆf (x 1 (k), ˆx 2 (k), u(k)) τk 2 sign(e 1 (k)) k 4 e 1 (k) e 1 (k) := ˆx 1 (k) y(k) (12) The dynamics for the observation error e i (k) =ˆx i (k) x i (k), i = 1, 2 are described by e 1 (k + 1) = e 1 (k)+τe 2 (k) τk 1 e 1 (k) 1/2 sign(e 1 (k)) k 3 e 1 (k) e 2 (k + 1) = e 2 (k)+τ ˆf (x 1 (k), ˆx 2 (k), u(k)) τk 2 sign(e 1 (k)) k 4 e 1 (k) τf (x 1 (k), x 2 (k), u(k)) (13) The following statement is true by assumption (for a given bounded u). Consider the following function f (x 1, ˆx 2, u) := ˆf (x 1, ˆx 2, u) f (x 1, x 2, u) (14) Suppose that the systems states are bounded, then the existence of a constant f + is ensured, such that the next inequality f (x 1, ˆx 2, u) f + (15) with f + R + holds for any possible k, x 1, x 2. The previous assumption is commonly used in the design of SM observers [8, 23. Remark 1: For a physical second-order system, the constant f + can be found as the double maximal possible value of the derivative of x 2 (for example, the acceleration for a mechanical system or the flux for a electrical system). Moreover, the estimation constant f + does not depend on the control terms. Such assumption of the state boundedness is true too, if, for example system (11) is bounded-input bounded-output (BIBS) stable, and the control input u(k) is bounded. Equation (13) can be rewritten by e(k+1) = e(k) + (k)sign(e 1 (k)) 1 k3 τ τk1 e 1 (k) 1/2 :=, (k) := k 4 1 τk 2 + τ f (16) This equation has a similar form to the equation given in (3). According to the statement given in Theorem 1, the following result is obtained Corollary 1: Consider the non-linear system in (11), where the output available is x 1 (k), if the assumption required for equation (14) is fulfilled and the observer gains are selected such that the LMI given by (R + R R) (1 ϱ)r < G (17) with defined as in (16) and k 2 > f +, has a positive-definite solution R = R > 0, then, the observer trajectories of (12) converge to the real states values of (11) in a QSM regime. Proof: Equations in (13) have the same structure as (3). Then, it is straightforward to follow the proof of Theorem 1 given in the appendix to obtain the LMI given in (17). The exact solution of (17) can be obtained with similar arguments to the ones described in lemma 1. 2.2.2 Vectorial case: Consider the case for systems with more than one degree of freedom but with relative degree two. The dynamics of these kind of systems is described by xα (k + 1) x(k + 1) = x β (k + 1) [ x = α (k) + τx β (k) x β (k) + τ (f (x (k), u(k)) + ξ(k)) (18) 351 356 361 366 371 376 381 386 391 396 401 406 411 416 IET Control Theory Appl., pp. 1 10 3 doi: 10.1049/iet-cta.2013.0568 The Institution of Engineering and Technology 2014
421 426 431 436 where x α (k + 1) R n and x β (k + 1) R n are defined as x 1 (k + 1) x 1 (k) + τx n+1 (k) x 2 (k + 1) x α (k + 1) =. = x 2 (k) + τx n+2 (k). x n (k + 1) x n (k) + τx 2n (k) x n+1 (k + 1) x β (k + 1) =. x 2n 1 (k + 1) x 2n (k + 1) x n+1 (k) + τ(f 1 (x(k), u(k)) + ξ 1 (k)) =. x 2n 1 (k) + τ(f n 1 (x(k), u(k)) + ξ n 1 (k)) x 2n (k) + τ(f n (x(k), u(k)) + ξ n (k)) (19) ˆx β (k) x β (k) the estimation error becomes [ (I n n β 3 ) x α (k)+τ x β (k) x(k + 1) = β 4 x α (k)+ x β (k)+τ( f (x(k), u(k))+ξ(k)) β1 0 ( xα (k)) τ S( x 0 β 2 I α (k)) (23) n n here f ( x α (k), x β (k), u(k)) := ˆf (x α (k), ˆx β (k), u(k)) f (x α (k), x β (k), u(k)), in the same way as in the scalar observer (15), f ( x α (k), x β (k), u(k)) f + and ξ(k) ξ, k Z + {0}. Let us denote f i ( x i (k), x n+i (k), u(k), u(k)) and ξ i (k) to the ith row of the functions f ( x α (k), x β (k), u(k)) and ξ(k) correspondingly. Owing to the boundedness assumption (15) an upperbound can be found for each couple of coordinates such that f i (x(k), u(k)) + ξ i (k) f + c i, f + c i R + (24) 491 496 501 506 441 446 Here, x(k) =[x 1 (k),..., x n, (k), x n+1 (k),..., x 2n (k) R 2n is the state vector and u(k) R n is the control action applied to the system at time t := τk belonging to U adm defined as U adm :={u: u(k) 2 v 0 + v 1 x(k) 2 u <, k Z + {0}} (20) This bound is justified by the same facts used in the scalar case. If the new variables and are defined as In n β := 3 τi n n β1 ( x, (k) := τ α (k)) β 4 I n n β 2 (25) with I n n being the n n matrix identity, the dynamics of the observation error can be represented in the form of (16) and following the results presented in Theorem 1, the next Lemma is easily verified 511 516 451 456 The signals ξ i (k) represent internal disturbances for each state x i (k) in the system. The vector ξ(k) R n known as the general perturbation is composed by ξ(k) := [ξ 1 (k),..., ξ n (k) and by assumption bounded, that is, ξ(k) 2 ξ ϒ, ξ = ξ > 0, k Z + {0}. Every couple of states for the system (18) can be seen as a decoupled equations and they can be treated as a class of independent systems taking into account the relationship between them. In this way, the proposed observer showed in (12) can be extended as Lemma 2: Consider the system given in (18), using the extended version of (12) given by (21), selecting β 1i > 0, β 2i > f c + i, β 3i > 2 and β 4i > 0 with i = 1:n, and if the following LMI ( (P + PϒP) (1 ϱ)p)< Q (26) has a positive-definite solution P = P > 0 with ϒ = ϒ > 0, P, ϒ R 2n 2n, then the observer states (ˆx α, ˆx β ) converge to the real states (x α, x β ) in a QSM regime. 521 526 461 ˆx α (k + 1) =ˆx α (k) τβ 1 λ( x α (k))s( x α (k)) β 3 x α (k) ˆx β (k + 1) =ˆx β (k) + τ(f (x α (k), ˆx β (k), u(k)) β 2 S( x α (k))) β 4 x α (k) (21) Proof: With the definitions given in (25) the estimation error becomes x(k+1) = x(k) + (k)s( x α (k)) (27) 531 466 471 where ˆx α, ˆx β are the estimates of the state vectors x α, x β respectively; the gain matrices β 1 R n n, β 2 R n n, β 3 R n n, β 4 R n n, λ( x α (k)) R n n and S( x α (k)) R n are defined as following the Theorem 1, the Lemma is proven. 3 Numerical results 3.1 Simulation of the DSTA 536 541 476 481 486 β 1 = diag{β 11, β 12,..., β 1n } β 2 = diag{β 21, β 22,..., β 2n } β 3 = diag{β 31, β 32,..., β 3n } β 4 = diag{β 41, β 42,..., β 4n } (22) λ( x α (k)) = diag{ x 1 (k) 1/2, x 2 (k) 1/2,..., x n (k) 1/2 } S( x α (k)) =[sign( x 1 (k)), sign( x 2 (k)),..., sign( x n (k)) β jk R + j = 1:4, k = 1:n where x i = x i x i according the definition in (18), (19) and (21) for i = 1:n. Defining x α (k) =ˆx α (k) x α (k), x β (k) = For the simulation, the DSTA parameters were chosen as ρ 1 = 0.9, ρ 2 = 0.3, the free gains were k 1 = 10.1 and k 2 = 10, and the sampled period was selected as τ = 0.01. Fig. 1 shows that the first state (x 1 ) converged to the origin and remains in a band near to zero. High-frequency oscillations when the state x 2 evolved to zero around the origin are depicted in Fig. 2. At the same time, the Lyapunov function remained in a band for bounded values of the states as shown in Fig. 3. Selecting Q = 3 10 3 and ϱ = 0.81, the matrix à defined in Lemma 1 becomes 0.9994 0.001 à = (28) 0 0.3331 546 551 556 4 IET Control Theory Appl., pp. 1 10 The Institution of Engineering and Technology 2014 doi: 10.1049/iet-cta.2013.0568
561 566 571 576 with [ this solution selecting = I 2 2, = 0.1688 0.6762 8.78 10 4. The radius of the ball B r defined in (6) becomes r = 0.0038, this fact can easily verified in the second graph of Fig. 3. The steady-state value of V (x) is 3 10 5, that is less than the ratio r. With this fact, the result claimed in the main theorem was numerically proven. In addition, in Figs. 1 and 2, it could be appreciated that the DSTA states x 1 and x 2 presented small oscillations in steady state (chattering effect); therefore the Euclidean norm and the trajectories of the Lyapunov function presented in Fig. 3 have a value of 0.075 and 3 10 5, respectively. Both graphs in Fig. 3 are quite similar because the Lyapunov function is the weighted Euclidean norm by the matrix P, that is V (x) = x 2 P. 631 636 641 646 Fig. 1 Trajectories for the first state of the DSTA 3.2 Non-linear DSTO 581 A second illustration of the results presented in this paper was the design of a DSTO for a non-linear pendulum system. Consider a pendulum whose state model is given by 651 x 1 (k + 1) = x 1 (k)+τx 2 (k) x 2 (k + 1) = x 2 (k)+τ 1 J u k τ mgl 2J sin(x 1(k)) τ V s J x 2(k) + τψ(k) y(k) = x 1 (k) 586 (30) 656 591 596 601 606 Fig. 2 Trajectories for the second state of the DSTA The previous equations represent an Euler discretisation of the continuous pendulum model, where x 1 (k) is the angle of oscillation, x 2 (k) is the angular velocity, m is the pendulum mass, g is the gravitational force, l is the pendulum length, J = ml 2 is the inertia arm, V s is the pendulum viscous friction coefficient. For simulation the bounded perturbation was expressed as ψ(k) = κ 1 sin(2τk) + κ 2 cos(5τk), with κ 1 = κ 2 = 0.5, the initial conditions were chosen as x 1 (0) = 3 and x 2 (0) = 1 for the model and ˆx 1 (0) = 10, and ˆx 2 (0) = 10 for the observer. The following numeric values were applied to simulate the pendulum parameters m = 1.1 kg, l = 1m, g = 9.81((m)/(s 2 )) and V s = 0.18(kg m)/(s 2 ). The input applied into the system is u(k) = sin(2τk) cos(5τk). The observer parameters were chosen as k 1 = 7.1, k 2 = 12, k 3 = 0.1 and k 4 = 0.01 for the SM observer. A Luenberger observer was designed for comparison purposes. The linear gains for the Luenberger observer were selected as l 1 = 20 and l 2 = 84 with the following structure 661 666 671 676 611 ˆx 1 (k + 1) =ˆx 1 (k) + τl 1 e 1 (k) 681 ˆx 2 (k + 1) =ˆx 2 (k)+τ 1 J u(k) 616 τ mgl 2J sin(x 1(k)) τ V s J x 2(k) + τl 2 e 1 (k) (31) 686 621 626 Q2 Fig. 3 a Euclidean norm of the DSTA states and b Trajectories of the proposed Lyapunov function The matrix P solution for the LMI (10) is 2.4337 1.87 10 5 P = 0.0034 (29) The numerical simulation of the angular velocity estimation is shown in Fig. 4. This brings out the fact that the DSTO behaves as the STA observer for continuous-time systems with small sampled times. It is clear that even when the Luenberger high-gain observer reaches the trajectories of the measurable state before the DSTO, the steady-state error for the DSTO was smaller than the Luenberger error. This fact was explained by the gains selected for the Luenberger observer that were greater than the DSTO gains. When the 691 696 IET Control Theory Appl., pp. 1 10 5 doi: 10.1049/iet-cta.2013.0568 The Institution of Engineering and Technology 2014
701 The upperbound of this term was given by 771 ψ(k) κ 1 sin(2τk) + κ 2 cos(5τk) κ 1 + κ 2 706 711 716 To complete the numerical test, one can chose κ 1 = κ 2, and then ψ(k) 2κ = κ +. A measure of robustness was studied in means of the magnitude of this perturbation. The following cases were considered in the paper, κ = 0.5, 1, 2, 5, then κ + = 1, 2, 4, 10. The following figure showed the Euclidean norm of the estimation. The solid line depicts the performance of the DSTO observer, while the dotted line does for the linear observer. One can note that the DSTO performance was unaffected by the increment of the perturbation norm (κ + ). Q3 776 781 786 721 726 731 736 Fig. 4 State estimation for the pendulum velocity by the DSTO error trajectories were faraway from the real ones, the linear gain has a more important effect than those including the signum function. The zone of convergence can be calculated with the result provided in Theorem 1. Selecting Q = 1 10 7 I 2 2, ϱ = 0.9999, the solution for the LMI (17) were 5.0191 0.0075 P = (32) 0.0075 0.001 and the radius r defined in (6) was r = 0.0078. In Fig. 5, the performance indexes for the observers were depicted for several values of the perturbation term ψ(k). The term ψ(k) was defined as ψ(k) = κ 1 sin(2τk) + κ 2 cos(5τk) 3.3 State estimation for 2n-dimensional systems The model of a biped robot was formed with five links connected with frictionless joints. The identical legs have knee joints between the shank and thigh parts, and one rigid body forms the torso. Fig. 6a shows the model structure and variables that were taken from [24. As the system can move freely in the x y-plane and contains five links, it has seven degrees of freedom. The corresponding seven coordinates were selected according to Fig. 6a q = [ x 0 y 0 α β L β R γ L γ R (33) The coordinates (x 0, y 0 ) fix the position of the torso centre of mass, and the rest of the coordinates describe the joint angles. The link lengths were denoted as (l 0, l 1, l 2 ) and masses as (m 0, m 1, m 2 ). The centres of mass of the links were located at the distances (r 0, r 1, r 2 ) from the corresponding 791 796 801 806 741 811 746 816 751 821 756 826 761 831 766 Fig. 5 Performance indexes ( e(k) 2 ) for different values of κ. 836 6 IET Control Theory Appl., pp. 1 10 The Institution of Engineering and Technology 2014 doi: 10.1049/iet-cta.2013.0568
841 911 846 916 851 921 856 Fig. 6 Bipedsim a Coordinates and constants b External forces c Leg tip touches the ground in point (x 0,0) (grey) and penetrates it; the current position of the leg tip is (x G, y G ) (black) 926 861 joints (Fig. 6). The model was actuated with four moments Table 1 Biped model parameters 931 866 871 M = [ M L1 M L2 M R1 M R2 (34) two of them acting between the torso and both thighs and two at the knee joints (Fig. 6b). The walking surface was modelled using external forces F = F Lx F Ly F Rx F Ry (35) Notation Value Units [ r = [l 1 l 2 l 3 0.8 0.5 0.5 m [ r = [r 1 r 2 r 3 0.4 0.25 0.25 m = [m 1 m 2 m 3 5 2 1 m kg 936 941 876 881 886 891 that affect the leg tips. When the leg should touch the ground, the corresponding forces are switched to support the leg. As the leg rises, the forces are zeroed (Fig. 6c). Using the Lagrangian mechanics, the dynamic equations for the biped system can be derived A (q) q = b(q, q, M, F) (36) Here A(q) R 7 7 is the inertia matrix and b(q, q, M, F) R 7 is a vector containing the right-hand sides of the seven differential equations. Matrices A(q) and b(q, q, M, F) are defined in [24. Using the Euler discretisation, the system (36) is implemented as a discrete one using the physical parameters for the biped model summarised in the following table Initial conditions for the biped robot were given by (see (37)) Initial conditions for the observer were selected as (see (38)) The gains for the DSTO were chosen as β 1 = diag { 15 15 15 25 25 25 25 } β 2 = diag { 25 25 25 35 35 35 35 } β 3 = diag { 0.01 0.01 0.01 0.01 0.01 0.01 0.01 } β 4 = diag { 0.01 0.01 0.01 0.01 0.01 0.01 0.01 } (39) To compare the results obtained by the DSTO a classical Luenberger structure was designed as [ x α (k) + τ x β (k) x(k + 1) = x β (k) + τ( f (x(k), u(k)) + ξ(k)) L1 0 [ xα (k) τ 0 L 2 x α (k) 946 951 956 961 896 x α = [ 5.730 1.37 0.011 0.299 0.033 0.206 0.240 x β = 0.8 [ 0.545 0.029 0.159 1.130 0.970 0.064 0.387 (37) 966 901 971 906 ˆx α = [ 4.011 0.959 0.008 0.209 0.023 0.144 0.168 ˆx β = [ 0.436 0.023 0.127 0.904 0.776 0.051 0.309 (38) 976 IET Control Theory Appl., pp. 1 10 7 doi: 10.1049/iet-cta.2013.0568 The Institution of Engineering and Technology 2014
981 1051 986 1056 991 1061 996 1066 1001 1071 1006 1076 Fig. 7 Position of selected states of the 24 1011 1081 1016 1086 1021 1091 1026 1096 1031 1101 1036 1106 Fig. 8 Velocity estimation of selected states of the biped simulator 1041 1046 with the following gains L 1 = 5 diag { 15 15 15 25 25 25 25 } L 2 = 5 diag { 25 25 25 35 35 35 35 } (40) The trajectories for the DSTO and the Luenberger observer are depicted in Fig. 7. It can be seen how the trajectories of the DSTO converged closer to the real system trajectories. The Luenberger state estimator did not converge to the real states. In the same way, a similar comparison with the velocities can be seen in Fig. 8. Again, a remark should be 1111 1116 8 IET Control Theory Appl., pp. 1 10 The Institution of Engineering and Technology 2014 doi: 10.1049/iet-cta.2013.0568
1121 1126 1131 biped model. This animation was modified from the one presented [24 in order to include a second biped that represented the trajectories of the state estimator. In Fig. 10, the first biped corresponded to the DSTO and the second one to the real system. In the instant t = τk = 0, the centre of mass of the biped represented by the observer is ubicated on the x-coordinates in x = 4.011 and the real system on x = 5.730. After some steps of time, the animation obtained with the DSTO reached the system. After t = τk = 1.75, the estimated biped reached the real biped animation. This confirmed the complete reconstruction of the biped model states by the observer ones (21). 1191 1196 1201 4 Conclusions 1136 1141 1146 1151 1156 Fig. 9 Norm of the estimation error to compare a classical Luenberger observer and the DSTO done in the gains selection. The gains for the linear observer were five times greater than the DSTO, this fact, implied a faster convergence into a zone for the Luenberger observer in similar way as it was claimed for the scalar case. However, the error was bigger in the linear case when the system was in the steady state. It is important to note that only four states were plotted with the objective of keeping the length of the paper small enough. A complete analysis including all the states of the biped model was done by means of the Euclidean norm of the error (Fig. 9). Moreover, following the results presented in [24, where an animation of a biped was performed, in Fig. (10) this animation was reprogrammed to present some clips of the animation with the observer trajectories. In this figure, the mesh line are used to differentiate the left leg from the righ left of the The Euler-like discretisation of the super-twisting algorithm was successfully analysed in terms of Lyapunov stability. The behaviour of the DSTA has similar characteristics like its continuous-time counterpart for small sampled periods. However, finite-time convergence cannot be proved with this quadratic Lyapunov function. The convenient gains to ensure the convergence of the DSTA were obtained by means of an LMI. A direct application of this new discrete structure is the problem of state estimation. Numerical results showed how the observer can reach the real trajectories of a n-degrees of freedom non-linear mechanical system in a small period of time with better performance than the results obtained by a Luenberger observer. 5 References 1 Spurgeon, S.: Sliding mode observers: a survey, Int. J. Syst. Sci., 2008, 39, (8), pp. 751 764 2 Atassi, A. N., Khalil, H.: Separation results for the stabilization of nonlinear systems using different high-gain observer design, Syst. 1206 1211 1216 1221 1226 1161 1231 1166 1236 1171 1241 1176 1246 1181 1251 1186 Fig. 10 Clips taken from a biped simulation of model given in ( 36) 1256 IET Control Theory Appl., pp. 1 10 9 doi: 10.1049/iet-cta.2013.0568 The Institution of Engineering and Technology 2014
1261 1266 1271 1276 1281 1286 1291 1296 1301 1306 Q4 Control Lett., 2000, 39, pp. 183 191 3 Utkin, V.: Sliding modes in control and optimization (Springer- Verlag, 1992) 4 Utkin, V.: Sliding mode control in electro-mechanical systems, Automation and control engineering (CRC Press, 2009, 2nd edn.). 5 Levant, A.: Sliding order and sliding accuracy in sliding mode control, Int. J. Control, 1993, 58, (6), pp. 1247 1263 6 Levant, A.: Principles of 2-sliding mode design, Automatica, 2007, 43, (4), pp. 576 586 7 Levant, A.: Robust exact differentiation via sliding mode technique, Automatica, 1998, 34, (3), pp. 379 384 8 Davila, J., Fridman, L., Levant, A.: Second-order sliding-mode observer for mechanical systems, IEEE Trans. Autom. Control, 2005, 50, (11), pp. 1785 1789 9 González, T., Moreno, J., Fridman, L.: Variable gain super-twisting sliding mode control, IEEE Trans. Autom. Control., 2012, 57, (8), pp. 2100 2105 10 Drakunov, S. V., Utkin, V.: On discrete-time sliding mode, Proc. IFAC Symp. Nonlinear Control Systems Design, 1989 11 Miloslavjevic, C.: General conditions for the existence of a quasisliding mode on the switching hyperplane in discrete variable structure systems, Autom. Remote Control, 1985, 46, pp. 679 684 12 Sira-Ramirez, H.: Nonlinear discrete variable structure systems in quasi-sliding mode, Int. J. Control, 1991, 54, (5), pp. 1171 1187 13 Bartoszewicz, A.: Discrete-time quasi-sliding mode control strategies, IEEE Trans. Ind. Electron., 1998, 45, (4), pp. 633 637 14 Sarpturk, S. Z., Istefanopulos, Y., Kaynak, O.: On the stability of discrete-time sliding mode control systems, IEEE Trans. Autom. Control, 1987, 32, (10), pp. 930 932 15 Chang, J. L.: Robust discrete-time model reference sliding-mode controller design with state and disturbance estimation, IEEE Trans. Ind. Electron., 2008, 55, pp. 4065 4074 16 Levant, A.: Finite differences in homogeneous discontinuous control, IEEE Trans. Autom. Control, 2007, 52, (7), pp. 1208 1217 17 Wang, B., Yu, X., Li, X.: Zoh discretization effect on higher-order sliding mode control systems, IEEE Trans. Ind. Electron., 2008, 55, (11), pp. 4055 4064 18 Velovolu, K., Soh Y.C., Kao, W.: Robust discrete-time nonlinear sliding mode state estimation of uncertain nonlinear systems, Int. J. Robust Nonlinear Control, 2007, 17, pp. 803 828 19 Spurgeon, S. K.; Hyperplane design techniques for discrete-time variable structure control systems, Int. J. Control, 1992, 55, (2), pp. 445 456 20 Koshkouei, J., Zinover, A. S.: Sliding mode state observers for SISO linear discrete systems, UKACC Int. Conf. Control, 1996 21 Bartolini, G., Pisano, A., Usai, E.: Digital second-order sliding mode control for uncertain nonlinear systems, Automatica, 2001, 37, (9), pp. 1371 1377 22 Salgado, I., Moreno, A., Chairez, I.: Sampled output based continuous second-order sliding mode observer, Workshop on Variable Structure Systems, 2010 23 Davila, J., Fridman, L., Poznyak, A.: Observation and identification of mechanical systems via second-order sliding modes, Int. J. Control, 2006, 79, (10), pp. 1251 1262 24 Havisto, H., Hyötyniemi, Simulation tool of a biped walking robot model, Technical Report, Control Engineering Laboratory, Helsinki University of Technology, 2004 25 Poznyak, A.: Advanced mathematical tools for automatic control engineers: volume 1: Deterministic systems (Elsevier Science, 2008) Using the so-called lambda inequality [25 X Y + Y X X 1 X + Y Y for 2x (k)a PB(k)sign (x 1 (k)) the following result was obtained 2x (k)a PB(k)sign(x 1 (k)) x (k)a P 1 PAx(k) + B (k) B(k) Equation (43) turns in V (k) x (k)(a (P + P P)A (1 ϱ)p)x(k) + B (k)( 1 +P)B(k) ϱv (k) (44) (45) Then, expanding the term B (k)zb(k) with Z := ( 1 + P), one has B (k)zb(k) = δ 1 x 1 (k) +δ 2 x 1 (k) 1/2 + δ 3 δ 1 := τ 2 k 2 1 z 11, δ 2 := τ 2 k 1 k 2 z 12, δ 3 := τ 2 k 2 2 z 22 (46) Using again the lambda inequality in the term containing δ 2 and if there exists a matrix Q = Q > 0 solution for the LMI given by A (P + P P)A (1 ϱ)p = Q, then V (k) becomes into (for any ω 1 R + ) V (k) x(k) 2 Q + δ 1 x 1 (k) + δ 2 ϱv δ 1 := δ 1 + τ 4 k 2 1 k2 2 z2 12 ω 1, δ 2 := k 2 2 τ 2 (τ 2 k 2 1 z2 12 ω 1 1 +z 22) Following the previous result V x(k) 2 Q + δ 1 QQ 1 x(k) αv + δ 2 ( = Q 1/2 x(k) 1 ) T 2 δ 1 Q 1 ( Q 1/2 x(k) 1 ) 2 δ 1 Q 1 αv + δ 2 + 1 4 δ 2 1 Q 1 2 (47) (48) then V (k) ϱv (k) + c (49) with c := δ 2 + 1 δ 2 4 1 Q 1 2. V (k + 1) (1 ϱ) V (k) + c (50) 1331 1336 1341 1346 1351 1356 1361 1366 1371 1376 1311 1316 1321 1326 6 Appendix Proof: (Proof of Theorem 1) Consider the following function like a candidate Lyapunov one Let V (k) := V (k + 1) V (k) then V (x) := x 2 P (41) V (k) = x (k + 1)Px(k + 1) x (k)px(k) (42) Substituting (4) into (42) the terms V (k) becomes V (k) = x (k)(a PA P)x(k)+2x (k)a PB(k) sign(x 1 (k)) + B (k)pb(k) (43) The last equation is well defined as a discrete linear one and the solution is given by V (k + 1) (1 ϱ) k V (0) + k (1 γ x ) i 1 c (51) If the upper limit when k goes to infinity is considered, one has lim k V (k) c (52) 1 ϱ and the radius of the region of convergence of the DSTA is defined as r c (53) 1 ϱ This result completes the proof. i=1 1381 1386 1391 1396 10 IET Control Theory Appl., pp. 1 10 The Institution of Engineering and Technology 2014 doi: 10.1049/iet-cta.2013.0568
1401 CTA20130568 Author Queries www.ietdl.org 1471 1406 1411 Iván Salgado, I. Chairez, Bijnan Bandyopadhyay, Leonid Fridman and Oscar Camacho Q1 Please provide the initial of author named Chairez. Q2 Please provide main caption for figure 3. Q3 Please cite table 1 in the text cited in text. Q4 Please provide editor name in ref. [4. 1476 1481 1416 1486 1421 1491 1426 1496 1431 1501 1436 1506 1441 1511 1446 1516 1451 1521 1456 1526 1461 1531 1466 1536