4 views

Uploaded by Bhethho

Método de integración con algoritmo de Euler

save

- Global Placement Application Form
- Sliding Mode Control of CNC Electromagnetic Levitation System Based On
- Undrained Stability of a Square Tunnel in a Soil Whose Strength Increases Linearly With Depth
- 2010-Improved Integral Sliding Mode Control Methods Pmsm
- EPE20110400010_80446125
- Robust Adaptive Controller for Uncertain Nonlinear Systems
- Eulers Method
- Euler Method
- An Improved Ray Tracing Algorithm
- Implementing Euler's Method
- 00768190.pdf
- 27 study
- EPYCF
- General Eng Standards
- Atkinson Introduction
- Te Margonari Enginsoft
- Mathematics
- Optimization of Impact attenuator design using Finite element method for Enhanced Vehicle Safety during collision
- runge.pdf
- BTech_(SVEC-10)_Syllabus_(EEE)
- 19 Noraini Jamaludin(1)
- ListofResearchtopics2012_1st
- 2 Ftp
- De Eulers Method
- Tutorial AUTOMGEN 8
- Tutorial AUTOMGEN 8
- EW_IDCBookextract.pdf
- Automatizacion de Un Sistema Movil Por Trayectorias y Marcas
- Runge kutta
- Stockbridge Damper
- Ejercicio 11 Capítulo 6 APA 1
- Early Enteral Feeding in a Very Low Birth Weight Infant
- Ecm Titanium ENG
- Cómo se Llega a Ser Psicólogo Reik, Theodor.pdf
- Least Common Multiple
- Anclaje Para Muros Pantalla
- Integrales Dobles Por El Met de Gauss Legendre
- Catalago de artesania
- Chords Construcciòn
- tm5_818_5
- Meaning of Genesis 4, Verse 26
- Gestión de Inventarios 2017
- VITULLO & SCAVO - O Liberalismo e a Definição Bobbiana de Democracia - RBCP v.13, n1
- Discutindo Exposições Mast Colloquia 8 Conceito Construção Avaliação
- OTIMIZAÇÃO ESTRUTURAL TOPOLOGICA.pptx
- Foundations and Pre-Calculus 10
- Fear of Failure With Experimentation –Thai Cultural Effects on Education
- Guia de Porcentajes
- BscFT Kapurthala Syl
- Rv 2
- QUESTIONÁRIOS SOBRE NORMAS REGULAMENTADORAS
- Forgotten Ingredients Final
- PERT sample question
- Bacilus Subtilis Sobre Clavibacter
- 01hoyt.pdf
- Memorias de Una GEISHA Karen
- Paso 2_Resignificar Conceptos Del Diseño de Proyecto_551101_7
- Per Esp Pin Tope Pos Sim 15kV
- BCH Memoria 2009

You are on page 1of 39

**Implicit Euler numerical simulations of sliding mode
**

systems

N° 6886

March 2009

apport

de recherche

ISRN INRIA/RR--6886--FR+ENG

Thème NUM

ISSN 0249-6399

inria-00374840, version 1 - 10 Apr 2009

Vincent Acary — Bernard Brogliato

inria-00374840, version 1 - 10 Apr 2009

**Implicit Euler numerical simulations of sliding mode
**

systems

Vincent Acary∗ , Bernard Brogliato†

Th`eme NUM — Syst`emes num´eriques

´

Equipe-Projet

Bipop

inria-00374840, version 1 - 10 Apr 2009

Rapport de recherche n 6886 — March 2009 — 36 pages

**Abstract: In this report it is shown that the implicit Euler time-discretization of some classes
**

of switching systems with sliding modes, yields a very good stabilization of the trajectory and

of its derivative on the sliding surface. Therefore the spurious oscillations which are pointed out

elsewhere when an explicit method is used, are avoided. Moreover the method (an event-capturing,

or time-stepping algorithm) allows for accumulation of events (Zeno phenomena) and for multiple

switching surfaces (i.e., a sliding surface of codimension 2). The details of the implementation

are given, and numerical examples illustrate the developments. This method may be an alternative

method for chattering suppression, keeping the intrinsic discontinuous nature of the dynamics on

the sliding surfaces. Links with discrete-time sliding mode controllers are studied.

Key-words: Switching systems, Filippov’s differential inclusions, complementarity problems,

backward Euler algorithm, sliding modes, maximal monotone mappings, mixed linear complementarity problem, ZOH discretization.

∗

†

vincent.acary@inrialpes.fr

Bernard.brogliato@inrialpes.fr

**Centre de recherche INRIA Grenoble – Rhône-Alpes
**

655, avenue de l’Europe, 38334 Montbonnot Saint Ismier

Téléphone : +33 4 76 61 52 00 — Télécopie +33 4 76 61 52 52

Cette mthode peut tre une alternative aux mthodes complexes de suppression des oscillations. Dans ce rapport.e. Mots-cl´ es : Systmes commuts. oprateurs.Simulations numriques par la mthode d’Euler implicite des systmes modes glissants inria-00374840. Les oscillations artificielles qui sont gnralement mentionnes pour l’implmentation discrte de ce type de systmes sont vites. Mthode d’Euler implicite. les dtails de l’implmentation sont donns et des exemples numriques illustrent ses proprits. et de leurs drives sur la surface de glissement. en gardant la nature intrinsquement discontinue de la dynamique sur les surfaces de glissement. version 1 . la mthode (de type “event-capturing” ou “time–stepping”) permet de traiter des accumulations d’vnements (Phnomne de Zenon) et des surfaces de commutations multiples (i. Inclusion Diffrentielles de Filippov. De plus. Problme linaire de complmentarit mixte. Le lien avec les commandes modes glissants en temps discret est tudi. maximaux monotones. modes glissants.10 Apr 2009 R´ esum´ e : Dans ce rapport. Discrtisation Bloqueur d’Ordre Zro (BOZ) . des surfaces de glissement de codimension 2). on montre que la discrtisation en temps de type Euler implicite conduit une trs bonne stabilisation d’une classe de systmes commuts avec des modes glissants. Problmes de complmentarit.

contrary to other discrete sliding mode controllers. the digital implementation of sliding mode controllers has been studied thoroughly in [27. version 1 . 23. Roughly speaking. 7]. which have an interest not only for the sake of numerical simulation. These central results of the paper show that the implicit Euler implementation of the differential inclusion yields a chattering free convergence in finite time on the sliding surface. [4. our study shows that convergence and order results may not be sufficient to guarantee that the derivative of the state is correctly approximated on the switching surface. where the Zero–Order Holder (ZOH) discretization is used. Both event-driven methods and time-stepping methods have been developed. The advantage of such “dual” methods in terms of their accuracy on the sliding surface has already been noticed in [49. The multivalued part of the sgn(·) function. Solutions to cope with chattering or reduce its effects have been proposed. aircraft control [44. 16. 12. Moreover. 57]. In Section 3. Fundamentally. Through several examples. see e. Section 5 is devoted the study of Discrete–time Sliding Mode Control and the extension to ZOH discretization. besides convergence and order results. The purpose of this paper is to analyze the implicit (backward) Euler method for some particular classes of differential inclusions. In this paper we focus on time-stepping methods. 22. 1 see [55] for a discussion on this point. Some important fields of application are induction motors [43. 8. 31]. 49. 50] in an event-driven context. 36]. these control schemes are of the switching discontinuous type and they yield closed-loop systems that can be recast into Filippov’s differential inclusions. From a numerical point of view. a multifunction.10 Apr 2009 1 3 Introduction Sliding mode controllers are widely used because of their intrinsic robustness properties [41. One drawback of these solutions is that they usually destroy the intrinsic discontinuous nature of sliding mode control. However they are known to generate chattering which renders their application delicate.g. 5. 34. some convergence and chattering free finite–time stabilization results are given. 55]. where the motivation was the simulation of mechanical systems with Coulomb friction. but by a multiplier (a slack variable in a nonlinear programming language).g. 51]. RR n 6886 . 51. It is shown that. 26. 35]. 32. that is not the case of most of the other methods which become quite cumbersome as soon as more than two switching surfaces are considered. 20].e. extend to ZOH discretizations of sliding mode systems. A further advantage of the proposed method is that contrary to other methods that have been studied and which destroy the intrinsic discontinuous nature of sliding mode systems 1 (like the so-called boundary layer control.g. 54. see e. this is due to the fact that the switches are no longer monitored by the state at step k. The numerical simulation of such nonsmooth dynamical systems is non trivial and it has received a lot of attention. to cite a few. our method keeps the multivalued discontinuity and consequently the fundamental aspects and properties of sliding mode control from a Filippov’s systems point of view. avoiding stiff problems. 53.Implicit Euler numerical simulations of sliding mode systems inria-00374840. i. Some hints on the numerical implementation of the implicit Euler scheme are given in Section 6 and the paper ends with some numerical experiments in Section 9. sampling rates need not be high to reduce chattering. the Equivalent–Control–Based Sliding–Mode–Control (ECB-SMC) and the Lyapunov–based discontinuous robust control are shown to fit well within this class of differential inclusion. In parallel. Recently it has been shown that the explicit Euler method generates unwanted effects like spurious oscillations (also called chattering effects) around the switching surface [25. [1] for a survey. but also for the real implementations of sliding mode controllers on discrete-time systems [52]. solar energy systems [28]. that include sliding mode controllers. or various filtering techniques). 37. 15. a class of differential inclusions is introduced and existence and uniqueness results are given under the maximal monotonicity assumption. see e. A second contribution of this paper is to show that the results that hold for the backward Euler scheme. 52. which also have their own limitations [55]. hard disk drives [33. the advantage of the implicit method is that it allows one to get a very accurate and smooth stabilization on the switching surface (of codimension one or larger than one). is then correctly taken into account. In Section 4. The paper is organized as follows: Section 2 presents a motivating example for using an implicit Euler implementation of the simplest sliding mode system. [50. The implicit method adapts naturally to an arbitrary large number of switching surfaces.

2 A simple example inria-00374840. redoing the xk +n+1 − xk0 +n approximating the derivative. sk = sgn(xo ) while xk ∈ / [−h. i ∈ α]T . We obtain sk0 +1 = −1 and xk0 +1 = ε that is zero at the machine precision. n 3. the solution of the time-discretization (2) is given by xk = x0 − sgn(x0 )kh. For n = 2. Clearly then the terms 0 h for any h > 0. The same holds for all xk0 +n . . The open ball of radius r > 0 centered at a point x ∈ IRn is denoted by Br (x). Let us now consider that xk0 ∈ [−h. Let us now state a result which shows that once the iterate xk has reached a value inside some threshold around zero for some k. INRIA .1] if x(t) = 0 (1) with x(t) ∈ IR. . Lemma 1 For all h > 0 and x0 ∈ IR. This result is robust with respect to the numerical threshold that can be encountered in floating point operations. h]. Otherwise. ε > 0 is zero at the machine precision. The backward Euler discretization of (1) reads as: xk+1 − xk = −hsk+1 (2) s k+1 ∈ sgn(xk+1 ) This method converges with at least order 12 (see Proposition 2 below). are zero same reasoning. then A•i is the ith column and Ai• is the ith row. there exists k0 such that xk0 +n = 0 and 0 for all n 1. that is xα = [xi . then k0 = 0. ε we obtain xk0 +2 = 0 and sk0 = . 1]. This system possesses a unique Lipschitz continuous solution for any x0 . version 1 . If x0 ∈ [−h. For a set of indices α ⊂ {1. .sk0 +n . we have to solve h xk0 +2 = −hsk0 +2 s k0 +2 (4) ∈ sgn(xk0 +2 ) and we obtain xk0 +2 = 0 and sk0 +2 = 0. h]. the column vector xα will denoted the sub-vector of corresponding indices in α.10 Apr 2009 To start with we consider the simplest case: if x(t) < 0 1 x(t) ˙ ∈ −sgn(x(t)) = −1 if x(t) > 0 . h]. let us assume that xk0 − h = ε ≪ 1. and k0 = ⌈ |x(0)| h ⌉ − 1. n} and a column vector x ∈ IRn . This robustness stems from the fact that the dynamics is not h only monitored by the sign of xk but also by the fact that the “dual” variable sk+1 belongs to [−1. xk0 +n+1 − xk0 +n = h Proof: The value k0 is defined as the first time step such that xk0 ∈ [−h. The only possible solution for xk0 +1 − xk0 = −hsk0 +1 (3) s k0 +1 ∈ sgn(xk0 +1 ) is xk0 +1 = 0 and sk0 +1 = xk0 . x(0) = x0 [-1. The symbol ⌈x⌉ is the ceiling function which gives the smallest integer greater than or equal to x. .4 Acary & Brogliato Notations and definitions: Let A ∈ IRn×m . For the next iteration. Indeed. h] for k < k0 . that is. then the dual variable sk+1 keeps its value and so does xk+n for all n 1.

Other time-stepping methods exist. Let us introduce equivalent formulations of the inclusion RR n 6886 . tk+1 ) as uk = −sgn(xk+1 ) as a function of xk and h. where h is the sampling time.10 Apr 2009 Figure 1: Iterations of the backward Euler method. which means that even a large time step assures a smooth stabilization on the sliding surface. Consequently this result shows that there are no spurious oscillations around the switching surface. ∞ for s = sgn(x). 37]. (6) 1 and . Remarkably Lemma 1 holds for any h > 0. From a control perspective the input is implemented on [tk . and from a control point of view it introduces a high-gain feedback that may not be desirable in practical applications. the discrete state xk and the control sk are displayed for x0 = 1. like the so-called switched model [1. This is illustrated on Figure 1. On Figure 2(a)-(c). or more exactly sgn(ǫ) where ǫ is a very small quantity whose sign is uncertain. can respectively observe the convergence with order 1 on Figure 2(d). 26].01 at t0 = 0 and for various values of the time–step h that are sufficiently large to illustrate the behavior of the time–stepping scheme and its convergence. however it fails to correctly solve the integration problem when the number of switched surfaces is too large (see also [4] for similar issues when the so-called sigmoid blending mechanism is implemented). contrary to other time-stepping schemes like the explicit Euler method [25. We can compute that es ∞ = 1 for all h > 0 and therefore there is no convergence in infinite norm .Implicit Euler numerical simulations of sliding mode systems 5 si h xk−1 xk xk+1 xk+2 xi −h inria-00374840. Moreover this method may yield a stiff system. we Complementarity framework Let us end this section by restating the systems (1) and (2) into the complementarity framework. In . It is noteworthy that in the implicit method there is absolutely no issue related to calculating sgn(0). It is easy to show that the explicit method yields an oscillation around x = 0. The implicit method automatically computes a value inside the multivalued part of the sign multifunction and may be considered as the time-discretization of the multifunction sgn(·). as shown in more general situations in [25. 26]. Let us define two discrete function norms to measure the convergence: ef ef ∞ p N i=0 = = (h |fk − f (tk )| N i=0 (5) p 1/p |fk − f (tk )| ) . version 1 . where few iterations are depicted until the state reaches zero. It is noteworthy that solving the system (2) with unknown xk+1 and sk+1 is equivalent to calculate the intersection between the graph of the multivalued mapping xk+1 → −hsgn(xk+1 ) and the straight line xk+1 → xk+1 − xk . There is no problem of causality in such an implementation. 2 .

10 Apr 2009 (a) State and Control vs. INRIA . es 2 (dotted line (iii)). es with respect to h in logscale 1 (dashed line (ii)).02 10 (ii) (i) (iii) 1 0.5 0 0 -0.5 2 t x(t) −s(t) 1 x(t) −s(t) 1 0.01 at t0 = 0.6 Acary & Brogliato x(t) −s(t) 1 0.5 -1 0 0. Time h = 0.001 0.5 0.001 0.5 -1 -1 0 0.0001 1e-05 1e-05 0.5 0 -0.0001 0.01 0. Time h = 0.01 (b) State and Control vs.5 0 2 0.1 errors inria-00374840. version 1 .5 1 1.5 1 1.5 1 1.2 0.5 -0.5 2 t t (c) State and Control vs.1 1 h (d) Numerical error es ∞ (solid line (i)).01 0. Time h = 0. Figure 2: A simple example for x0 = 1.

1] (s(t)) ⇔ s(t) ∈ [−1. 1[ s(t) ∈ sgn(x(t)) ⇔ x(t) ∈ N[−1. where x(t) ∈ IRn : ˙ ∈ −A(x(t)) + f (t. x2 ∈ IRn . T ]. a other complementarity formulation can be defined. 1]. for all x ∈ D(A). 0 v2 ⊥ 1 − s 0} yields the following complementarity representation of the sign multi-valued function x(t) = −v1 (t) + v2 (t) x(t) ∈ N[−1. By setting λ1 (t) = 21 (1 − s(t)) and λ2 (t) = v1 (t). i. 0 v1 ⊥ s + 1 0. a.T ).Implicit Euler numerical simulations of sliding mode systems 7 s(t) ∈ sgn(x(t)) such that x(t) = 0 if s(t) ∈] − 1.1] (s(t)) ⇔ 0 λ1 (t) ⊥ x(t) + λ2 (t) 0 (10) 0 λ2 (t) ⊥ 1 − λ1 (t) 0 3 A class of differential inclusions Let us now introduce the following class of differential inclusions. Proposition 1 Let Assumption 1 hold. (iii) There exists a function Φ(·) such that for all R Φ(R) = sup ∂f (·. one gets s(t) = 1 − 2λ1 (t) x(t) ∈ N[−1. one has ||f (t. version 1 . on (0. inria-00374840.. 1] and x(t) 0 if x(t) = −1 x(t) 0 if x(t) = 1 (7) where N[−1. T ) → IRn that is Lipschitz continuous. T ) x(t) (11) x(0) = x0 The following assumption is made: Assumption 1 The following hold: (i) A(·) is a multivalued maximal monotone operator from IRn into IRn .e.IR ) | v 0: n L2 ((0.1] (s) = {−v1 + v2 . for all x1 . Then the differential inclusion (11) has a unique solution x : (0.1] (s(t)) ⇔ 0 v1 (t) ⊥ s(t) + 1 0 0 v2 (t) ⊥ 1 − s(t) 0 (8) (9) In order to directly substitute the value of s(t) into the dynamics x(t) ˙ = −s(t). y ∈ D(A) and all x′ ∈ A(x). RR n 6886 . The following is proved in [10. and let x0 ∈ D(A). x1 ) − f (t.e. v) ∂t n L2 ((0.1] is the normal cone in the sense of Convex Analysis to the interval [−1.10 Apr 2009 N[−1.IR ) R < +∞ . y ′ ∈ A(y).T ). The definition of the normal cone in the present case. x2 )|| L||x1 − x2 ||. one has (x′ − y ′ )T (x − y) 0 (12) (ii) There exists L 0 such that for all t ∈ [0. x(t)). 9]. with domain D(A).

Equ.e. and Sgn(Cx(t) + D) = (sgn(C1 x + D1 ). We can rewrite the i=1 RB•i sgn(Ci• R i=1 R system as m z(t) ˙ ∈ Rf (t. Theorem 5.7] it follows that fi (·) is convex. sgn(Cm x + Dm ))T ∈ IRm . Let R be the symmetric square root of P .. R2 = P . Proof: The proof uses a state variable change introduced in [13]. Sgn(Cx + D) = (sgn(x1 + 2x2 ). [57]). Let us consider the following example taken from [55]. Then the condition in (14) holds with P = c1 1 assures that P > 0.2]. It will be shown how to recast (13) into (11). Example 1 (Equivalent-control-based sliding-mode-control (ECB-SMC)) Consider a system x(t) ˙ = F x(t) + Gu. Exercise 12. Example 2 Consider the sliding mode system in [25.g. From [29. Corollary 31.(1)–(4)].. whose closed-loop dynamics is given by: x(t) 1 x(t) ˙ 0 0 0 0 e(t) 0 k −k −k e(t) ˙ 0 − sgn(Cx(t)) (17) x˙ s (t) = 0 0 0 1 xs (t) 0 1 2 1 x˙ s (t) 0 0 −τ2 −τ x ¨s (t) τ2 INRIA . . Suppose that (ii) and (iii) of Assumption 1) hold. By [46. on (0. Then the closed-loop system x(t) ˙ = (F − G(HG)−1 HF )x(t) − αG(HG)−1 Sgn(Hx(t)) fits within (13). D = 0. with an equivalent-control-based sliding-mode-control (ECB-SMC) of the form u(x) = −(HG)−1 HF x − α(HG)−1 Sgn(Hx). x(t)) − BSgn(Cx(t) + D). One has B = (0 1)T . α > 0 (see e. then for any initial data the differential inclusion (13) has a unique solution x : (0. Being the subdifferential of a convex function. Proposition −1 T 1..8 Acary & Brogliato In this paper we shall focus on inclusions of the form: ˙ ∈ f (t.4] it follows that each T multivalued mapping z → R−1 Ci• sgn(Ci• R−1 z(t) + Di ) is monotone. inria-00374840. Let us perform the state transformation z = Rx. Corollary 1 Consider the differential inclusion in (13). sgn(2x1 − x2 ))T . i. If there exists an n × n matrix P = P T > 0 such that T P B•i = Ci• (14) for all 1 i m. T ) → IRn that is Lipschitz continuous. the multivalued mapping z → ∂fi (z) is maximal (monotone) [45.5. Therefore RBSgn(CR−1 z(t) + m m −1 −1 T D) = z + Di ) = Ci• sgn(C•i R−1 z + Di ). Then we get z(t) ˙ ∈ Rf (t. 1 2 . By [45. version 1 . R−1 z(t)) − RBSgn(CR−1 z(t) + D) (15) m Notice that BSgn(CR−1 z(t) + D) = i=1 B•i sgn(Ci• R−1 z + Di ). Trajec2 −1 tories may slide on codimension one surfaces x1 + 2x2 = 0 or 2x1 − x2 = 0 and on the codimension 2 surface (x1 + 2x2 = 0 and 2x1 − x2 = 0).10 Apr 2009 Let us now state a well-posedness result which is a consequence of Proposition 1. R −1 z(t)) − T R−1 Ci• sgn(Ci• R−1 z(t) + Di ) (16) i=1 The multivalued mapping ξ → sgn(ξ) is monotone.3. p11 c1 and p11 > (c1 )2 C = (c1 1). Therefore by Proposition 1 the inclusion in (16) possesses a unique Lipschitz solution on (0. Example 3 Consider B = Example 4 One solution to reduce chattering is the observer based SMC. T ) x(t) (13) x(0) = x0 ∆ with B ∈ IRn×m .e. T ) for any T > 0 and since R is full–rank so does (13). a.11] it follows that R Ci• sgn(Ci• R−1 z(t) + Di ) = ∂fi (z) with fi (z) = |Ci• R−1 z(t) + Di |.

and is not straightforwardly avoided. T ] with a backward Euler scheme as follows: xk+1 − xk + A(xk+1 ) ∋ f (tk . It is noteworthy that (14) does not imply that B has full column rank. This is made possible because of the multivalued characteristic of the discontinuous input.. for all k ∈ {0. . The restriction to quadratic Lyapunov functions stems from monotonicity preserving conditions. f (·) satisfies assumption 1. For the notations see [55. P = P T > 0.i P x.. II. In particular it does not preclude m > n.. all t 0 and some finite ρi .Implicit Euler numerical simulations of sliding mode systems with C = (1 − 1 1 −1 with P = 0 0 9 0 0). because of the multivaluedness property. Such a controller assures the global asymptotic stability of the equilibrium x = 0. Corollary 2 Suppose that V (x) = 12 xT P x. p > 1. p33 > 0. and γ(·) ∈ IRm is a bounded disturbance satisfying |γi (t)| < ρi for all 1 i m. The problem is the stabilization of the system at the origin x = 0. This system satisfies the condition (14) −1 0 0 p22 0 0 . T Proof: We have ∇V (x)T B•i = B•. hence to (20). x(0) = x0 (18) where x(t) ∈ IRn . Dissipative systems with no feedthrough matrix satisfy an input-output constraint similar to (14) [14]. where R > 0 is the symmetric square root of P . knowing that there exists a function V (·) such that the uncontrolled undisturbed system x(t) ˙ = f (x(t)) admits V (·) as a Lyapunov function. p22 > 0. one has V˙ (x(t)) = ∇V (x(t))T f (x(t)) 0 along the trajectories of the free system. 4 Convergence results and Chattering Free Finite–time Stabilization The differential inclusion (11) is time-discretized on [0. B ∈ IRn×m . N − 1} h (21) x0 = x(0) RR n 6886 . We obtain: m m x(t) ˙ ∈ f (x(t)) − i=1 ρi B•i sgn(∇V (x)T B•i ) + i=1 B•i γi (t) (20) We can state the following result. version 1 . 0 p33 0 22 0 0 p44 T T T Notice that the condition (14) implies that B•i P B•i = B•i Ci• = Bi• C•i > 0.10 Apr 2009 x(t) ˙ = f (x(t)) + Bu(t) + Bγ(t). The system in (20) has a unique Lipschitz continuous solution on [0. Let z = Rx. Example 5 (Lyapunov-based discontinuous robust control) Let us show how the above material adapts to this type of feedback controller. When m = 1 this is a relative degree one condition. xk ).C]. In particular. Let us rewrite the system in (18) as m m x(t) ˙ = f (x(t)) + i=1 B•i ui + i=1 B•i γi (t) (19) Let us propose the control input ui (x) = −ρi sgn(∇V T (x)B•i ). The closed-loop system possesses the origin as its unique equilibrium. +∞) for any x0 . We may rewrite (20) as m m z(t) ˙ ∈ Rf (R −1 z(t)) − i=1 T Rz) + ρi RB•i sgn(B•i i=1 RB•i γi (t) Then following the same steps as for the proof of Corollary 1 we conclude that Proposition 1 applies to this system. The class of dynamical systems is inria-00374840.

Proposition 5]. Solving such generalized equations thus boils down to computing the intersection between the graph of Fs (·) and the graph of Fm (·) as illustrated in section 2. The result of Proposition 2 applies to (23). Similarly it is possible that time-stepping methods converge for systems that satisfy such a criterion. then the following bound is satisfied in the neighborhood of the sliding subspace. T ) h (23) x(0) = x0 One sees that advancing the implicit method from step k to step k + 1 involves solving generalized equations with unknown xk+1 . Thus the numerical scheme in (21) has at least order 21 . Before stating the smooth stabilization result. The next result is proved in [10]. Let us denote the output of the dynamical as: ∆ y(t) = Cx(t) + D (24) Lemma 2 Let us assume that a sliding mode exists for some indices i ∈ α ⊂ {1 . . xk ).(ii) holds. ∃κ > 0. one has (Cα• f (x(t). such an implicit method also assures a good estimate of the derivative x˙ and a smooth stabilization of the discrete-time solution on the sliding surface. the precision of the method may be much better than what is to be expected from (22) on large portions of the trajectories. m} such that ∃t∗ > 0. on (0. . t) − hCα• BSgn(y(t)). of the form 0 ∈ Fs (xk+1 ) + Fm (xk+1 ) where Fs (·) is singlevalued while Fm (·) is multivalued. tk and xk appear as parameters of the generalized equations. a. The differential inclusion in (13) is therefore discretized as follows: xk+1 − xk ∈ f (tk . like Filippov’s criterion for uniqueness of solutions [18. T ]. version 1 .10 Acary & Brogliato T . despite no result seems to be available in the literature. As we shall see next. The fully implicit method uses f (tk+1 . The convergence where h = N and order results stated in Proposition 2 below have been derived for the semi-implicit scheme (21) in [10]. ∃r > 0. ∀¯ x ∈ Br (x). xk ) − BSgn(Cxk+1 + D).10 Apr 2009 Moreover limh→0+ maxt∈[0. As seen in Lemma 1. Proposition 2 Under assumption 1. xk+1 ) instead of f (tk . The conditions of Assumption 1 for the existence and uniqueness results of Proposition 1 are sufficient only. (Cα• f (¯ x. t) ∈ Cα• BSgn(y(t)) (28) INRIA . t)) ρ (26) Furthermore. For t > t∗ . and convergence holds. t)) κr + ρ (27) for all x(t) such that Cα• x(t) + Dα = 0. yα (t) = Cα• x(t) + Dα = 0. for all t > t∗ . h. So the analysis in this section is based on such a discretization. let Assumption 1. Proof: From (25). ∃ρ > 0 such that ∀t > t∗ . Other criteria exist. However this is only a particular case of a more general θ−method which is used in practical implementations. ||x(t) − xN (t)|| inria-00374840. (25) Then there exists ρ > 0 such that for all t > t∗ and for all x(t) such that Cα• x(t) + Dα = 0. there exists η such that for all h > 0 one has For all t ∈ [0.e. we have y˙ α (t) ∈ Cα• f (x(t). the sliding mode yα (t) = 0 implies that y˙ α (t) = Cα• x(t) ˙ = 0 for all t > t∗ and therefore Cα• f (x(t).T ] ||x(t) − xN (t)||2 + t 0 η √ h (22) ||x(s) − xN (s)||2 ds = 0. let us consider a preliminary result.

. we have to solve for yk+1 = Cxk+1 + D and sk+1 the generalized equation yk+1 = yk + hCf (tk .k0 +1 ) h(K1 + hκK2 + h(ρ + ρ1 )) (34) and therefore it is possible to choose h1 such that for all h < h1 −h(Cα• B•α )−1 [yα. for t > t∗ almost everywhere and using (33) we obtain that there exists ρ1 > 0 such that √ √ yα. that is yα (t) = 0. i (35) If (35) is satisfied. t)) ρ (29) for all x(t) such that Cα• x(t) + Dα = 0.• f (tk0 . (30) Combining (29) and (30) ends the proof. . Choosing hc as the minimum of the considered time steps h1 .k0 √ for Cα• x(t1 ) + Dα = 0. Our results may therefore be considered as the continuation of studies on the finite-time convergence for algorithms of the proximal type. yk0 +2 = hCf (tk0 +1 .k0 +2 = 0.k0 +1 )] 1.k0 +2 )] (39) i and therefore the solution of (37) is The bound (29) is uniform and can be applied for the next steps. n 1.k0 +n . for all i ∈ α. xk0 +1 ) − hCα• B•β Sgn(yβ. Proof: At each time–step. xk ) − hCBsk+1 (31) sk+1 ∈ Sgn(yk+1 ) Under condition (14). Lemma 1 extends to (23) as follows when the sliding surface of codimension |α| is attained.• f (tk0 . the Lipschitz continuity of f (·. Let C and B be such that (14) holds and Cα• B•α > 0. yi (t) = Ci• x(t) + Di = 0}. . xk0 ) − hCα• B•β Sgn(yβ. ∃t1 > t∗ such that yα.k0 +1 = Cα• xk0 +1 + Dα = 0. we have the following bound √ √ h(K1 + hκK2 + hρ) (33) yα. there exists k0 ∈ IN such that yk0 +n = Cxk0 +n+1 + D = 0 for all integers n 1.k0 +2 )] 1. RR n 6886 .• f (tk0 +1 .• f (tk0 +1 . t > t∗ . t)) ∆ Cα• Lr = κr.k0 + hCα. xk0 +1 ) − hCBsk0 +2 . xk0 ) − hCα• B•β Sgn(yβ.(ii). the convergence of the time–stepping scheme is ensured by Proposition 2.k0 + hCα. we can use (29) to conclude that there exists h2 such that for all h < h2 −h(Cα• B•α )−1 [hCα. By the assumption 1. sα. version 1 . xk0 ) − hCα• B•β Sgn(yβ. . ∃K1 > 0. . Using (27) for x(t1 ) and a sufficiently small h such that r = K2 h.k0 + hCα. Cα• (f (¯ x(t). m}. t) − f (x(t). for all i ∈ α.. The finite-time convergence of the time-discretization of similar nonsmooth dynamical systems (essentially mechanical systems with dry friction) is proved in [6]. Then there exists hc > 0 such that ∀h < hc .k0 +2 = −h(Cα• B•α )−1 [hCα.k0 +1 = −h(Cα• B•α )−1 [yα. the proof is obtained for yα. The convergence and the existence of the sliding mode ensure that √ √ K2 h (32) K1 h and xk0 − x(t1 ) ∃k0 .• f (tk0 . ∃K2 > 0.k0 + hCα.Implicit Euler numerical simulations of sliding mode systems 11 The inclusion (28) yields ∃ρ > 0. ·) allows us to write for some κ > 0 ∀¯ x(t) ∈ Br (x(t)). the unique solution of (31) at the iteration k0 + 1 is given by yα. sk0 +2 ∈ Sgn(yk0 +2 ) (37) Using the fact that yα. (38) yα.• f (tk0 . sα.10 Apr 2009 Lemma 3 Let us assume that a sliding mode occurs for the index α ⊂ {1 .k0 +1 )] (36) The next iterate will by given by the solution of the generalized equation.k0 +1 = 0. xk0 +1 ) − hCα• B•β Sgn(yβ. xk0 ) Introducing the complementary index set β = {i. inria-00374840. h2 . (Cα• f (x(t).

u(x) = −(CG)−1 (CF x + αSgn(Cx)). tk+2 ) one obtains sk+2 = 0 ht ht x(t) = xk+1 − xk+1 = 0 − 0 = 0 (41) h h and so on on the next intervals. h]. Therefore on [tk . and h Φ = exp(F h) − exp(F τ )dτ G(CG)−1 CF (44) 0 h Γ exp(F τ )G(CG)−1 dτ = (45) 0 with G ∈ IRn×m . see Section 6 below).10 Apr 2009 5.2 Extension to ZOH discretized systems The ZOH discretization of linear time invariant systems x(t) ˙ = F x(t) + Gu(t) with an ECB-SMC controller. Suppose that xk ∈ [−h. Then following the same calculations as in the proof of Lemma 1. C ∈ IRm×n . On the next sampling interval [tk+1 . if we suppose that xk ∈ / [−h. Remark 1 The fact that the function sgn(·) generates only binary values (+1 or −1) does not hamper the above method to work. we obtain that sk+1 = xhk . tk+1 ) one has x(t) = xk + ht uk . version 1 . 56]. α > 0 results in a discrete-time system of the form: xk+1 = Φxk − Γsk for all t ∈ [kh.1] ( xk ) h (42) where projC denotes the Euclidean projection operator onto the set C. It requires at each step to solve a rather simple multivalued problem which a Mixed Linear Complementarity Problem (MLCP. As alluded to above. 5. where ht = t − tk . let us set uk = −(CG)−1 (CF xk + sk+1 ) sk+1 = Sgn(Cxk+1 ). For this simple system. the value of sk+1 is 1 or −1 according to the sign of xk . inria-00374840. h]. (46) INRIA . Let us implement a “backward” controller uk = −sgn(xk+1 ) at time tk .1 Example of an implicit Euler controller (IEC) Let us come back to the inclusion in (1). It is not of the high gain type. To summarize the control is given explicitly in terms of xk and h by uk = −proj[−1. following the above lines. when a explicit Euler implementation of the control is performed [52. such an “implicit” input is causal and can be computed at tk with the values of the state at tk by (42). (k + 1)h) (43) where h > 0 is the sampling period.12 5 Acary & Brogliato Discrete–time Sliding Mode Control (SMC) This section is devoted to show how the above discretizations may be used in a digital control framework. For an implicit Euler implementation. the ZOH and the Euler discretization yield the same–discrete system. Assume the integrator x(t) ˙ = u(t) is sampled with sampling period h > 0. The controller u(x) = −sgn(x) is known as the equivalent control-based SMC [57]. where the zero value is obviously some small value at the machine accuracy. tk+1 ): ht (40) x(t) = xk − xk h and it follows that x(tk+1 ) = xk+1 = 0. On the time interval [tk . Indeed the implicit Euler method allows us to compute values of the sign multifunction inside its multivalued part at xk = 0.

w. a vector q ∈ IRm and lower and upper m bounds l. RR n 6886 . We will see in the next Section how the computation can be carried out in practice. −∞}. one has to solve x = Φxk − Γsk+1 k+1 yk+1 = Cxk+1 + D . yk the discrete output and sk the discrete input. The discrete system (50) is a common representative for the discretization given by (23). The matrices C and D are given by their definition in (13).2. sk+1 ∈ Sgn(yk+1 ) (47) (48) Inserting the first line of (48) into the second line we obtain the following one–step system inria-00374840. xk the discrete state. v ∈ IRm + such that Mz + q = w − v l z u (51) (z − l)T w = 0 (u − z)T v = 0 where IR = IR ∪ {+∞. F . (21) or (48) and the matrices R ∈ IRn×n . u ∈ IR . We therefore get on each sampling period: xk+1 = Φxk − Γsk+1 for all t ∈ [kh. (49) Comparing with the time–discretized systems in (23) and (31) one sees that the term hCB is replaced in case of a ZOH method by the term CΓ. h. Let us define what is a MLCP in its general form with bounds constraints as it has been proposed in [21]: Definition 1 (MLCP) Given a matrix M ∈ IRm×m . G and C. Provided the problem has a unique solution one can compute the controller in (46) with the knowledge of xk . (k + 1)h) At each time–step.1 Mixed Linear Complementarity Problem (MLCP) The time–discretized system (50) appears to be a Mixed Linear Complementarity Problem (MLCP) that we have to solve at each time–step. 6. 6 Implementation of Discrete–Time Systems Let us consider in this section the following discrete–time system: x = Rxk + p − Ssk+1 k+1 yk+1 = Cxk+1 + D sk+1 ∈ Sgn(yk+1 ) (50) where k 0 is an integer.10 Apr 2009 yk+1 = CΦxk + D − CΓsk+1 sk+1 ∈ Sgn(yk+1 ) . find z ∈ IRm . version 1 .Implicit Euler numerical simulations of sliding mode systems 13 which corresponds to the implicit discrete time version of the ECB-SMC controller. S ∈ IRn×m and the vector p ∈ IRn are determined by the chosen time–discretization method and detailed in Section 6.

. the problem (50) can be recast into a MLCP by observing that sk+1 ∈ Sgn(yk+1 ) yk+1 ∈ N[−1. x(t)) as it has been given in (21). xk ). . m} The MLCP (51) is a well-known problem in the mathematical programming theory arising for instance from the Karush–Kuhn–Tucker optimality conditions of a quadratic program or from the primal/dual optimality conditions of a linear program. . ui = 1. .14 Acary & Brogliato Note that the problem (51) implies that −(M z + q) ∈ N[l. yk+1 = w − v M = CS. . The box [l. i = 1 . S = hB. 6. M z + q)) and v is the negative part of M z + q. 1[ and yj. . m}. .k+1 ∈] − 1.2 Some Time–Discretization Methods In this Section.k+1 0 if sj. At each time step. q = −(CRxk + Cp + D) li = −1.k+1 = 0 if sj. The results of existence and uniqueness of solutions of (51) are related to the properties of M (P-properties or coherent orientations of the associated affine map (normal map) for particular cases of bounds constraints). xk ) + Cxk + D) (56) INRIA . ui ].k+1 = −1 yj. that is v = (M z + q)− = max(0. Without entering into further details. the computations are done with the help the Siconos/Numerics open source Library [3] and/or the PATH solver [21]. The normal cone to a convex set is a standard instance of a multi–valued mapping [45].10 Apr 2009 where the notation NC (x) is used for the normal cone in the Convex Analysis sense to a convex set C at the point x. version 1 . 17] b) extensions of projection/splitting techniques for MLCP [24. u] ⊂ IRm is defined by the Cartesian product of the intervals [li . ·) Let us start with the explicit Euler discretization method of the term f (t. 1]m yj. . The assumptions on the matrix M drives the choice of particular solvers that can be in polynomial time rather than standard exponential time for brute force enumerative solvers. m. the variable xk+1 is condensed into the second line such that yk+1 = CRxk + Cp − CSsk+1 + D (53) sk+1 ∈ Sgn(yk+1 ) and the following variable and parameters are defined as follows z = sk+1 . j ∈ {1. 21.k+1 0 if sj. we refer to [30.1]m (sk+1 ) sk+1 ∈ [−1. that is w = (M z + q)+ = max(0. . q = −(hCf (tk . (52) inria-00374840. the formulation of the discrete–time system (50) is related to the continuous time system (13) through a given discretization method. Several families of solvers may be cited: a) extensions of Lemke and principal pivotal techniques for LCP to MLCP [48. The relation (52) is equivalent to the MCP (51) if we assume that w is the positive part of M z + q. 24] for the main results. The MCLP enjoys a large number of numerical algorithms and several reliable solvers have been implemented. In this paper. i ∈ {1.u] (z). p = hf (tk . 19] and c) semi–smooth Newton methods [40]. In order to state the problem (50) as a MLCP. M = hCB. the matrices in (50) and in the MLCP (51) can be identified as R = I. Explicit Euler discretization of f (·.k+1 = 1 (55) . 47. −(M z + q)). . (54) Finally.

The main motivation for doing in this way is the higher accuracy and stability that we can obtain for such a numerical integration scheme (see [2] for an example of instability with the Explicit Euler method). the control scheme is depicted showing that the controller is causal and computed form xk . Φ and Γ which amounts to compute the matrix exponential and its time integral [39].2 can be also formalized into the form (50) and then (51) with R = Φ. Bi ∈ IRn×1 . The second class of differential inclusions is: m ˙ ∈ f (t. x(t)) by a implicit Euler scheme or a θ-method. x(t)) = F x(t) + g. For θ = 1. the implicit Euler scheme is used to discretize f . x(t)) − i=1 (Ai x(t) + Bi )sgn(Ci x(t) + Di ).method In a more general way. q = −(CΦxk ) (58) In practice. In this case. M = hC(I − hθF )−1 B. Let the assumptions of Corollary 1 be satisfied with B full–column rank (CB = B T P B > 0). Let us consider first that the mapping f (·. the explicit Euler case is retrieved. ·) is affine.e.Implicit Euler numerical simulations of sliding mode systems 15 uk − xk Discrete-time Plant (CG)−1 CF xk − sk+1 MLCP solver inria-00374840. In this work. q = −((I − hθF )−1 (I + h(1 − θ)F )xk + (I − hθF )−1 g + D) (57) for θ ∈ [0. We refer to [2] for a detailed presentation of these developments. Furthermore. This result ensures the existence and uniqueness of a solution of the MCLP. standard pivotal techniques such as Lemke’s method or projection/splitting such as Projected Successive Over-Relation (PSOR) compute the solution.10 Apr 2009 Figure 3: Control system schema with implicit Euler implementation. p = (I − hθF )−1 g. numerous methods are available to compute the ZOH discretization. ·) is nonlinear. p = 0. The matrices in (50) and in the MLCP (51) can be identified as R = (I − hθF )−1 (I + h(1 − θ)F ).e. i. If the mapping f (·. the solution at each time step is sought as a limit of solutions of successive MLCPs. on (0.. S = h(I − hθF )−1 B. RR n 6886 . T ) x(t) (59) x(0) = x0 with Ai ∈ IRn×n . Di is a scalar. Implicit Euler and θ. we introduce other classes of differential inclusions which extend (13). 1]. M = CΓ. version 1 . we can choose to time–discretize the term f (t. that is f (t. For θ = 0. S = Γ. On the Figure 3. Zero–Order Holder (ZOH) method The ZOH discretization presented in Section 5. a. the numerical computation is performed using an explicit Runge–Kutta method with high order of accuracy and a numerical tolerance near the machine precision threshold. a newton linearization can be invoked. Ci ∈ IR1×n . 7 Two other classes of differential inclusions In this section.

Example 8 The mapping F : IR → IR. Corollary 4 Consider the differential inclusion in (60). Then F (x) = (kx + 1)sgn(x). Suppose that the multivalued mappings x → g•i (x)sgn(hi (x)). Indeed the jump of the vector field is equal to 2Ai x = 0 on Σi . Proof: First notice that the set-valued mapping x → αCiT sgn(Ci x + Di ) is maximal monotone [46. Then for any initial data the differential inclusion (59) has a unique solution x : (0. Suppose that the multivalued mappings x → (Ai x + Bi )sgn(Ci x + Di ). Hence Proposition 1 applies. 1 i m. Sgn(h(x)) = [sgn(h1 (x)..16 Acary & Brogliato The third class that we shall analyze is: ˙ ∈ f (t.10 Apr 2009 The proof is straightforward and is omitted. −k1 x + 1 if x 0 −k2 x − 1 if x 0 . on (0. version 1 . Suppose that (ii) and (iii) of Assumption 1) hold. Example 6 The mapping IR → IR. Suppose that (ii) and (iii) of Assumption 1) hold. T ) x(t) (60) x(0) = x0 where g : IRn → IRn×m and h : IRn → IRm are smooth functions. is Example 7 Let k1 . Lemma 4 Let Bi = αCiT for some α > 0.4]. x → 9 . The proof is straightforward and is omitted. |k2 |). are hypomonotone. Then for any initial data the differential inclusion (60) has a unique solution x : (0. Let k1 = −k2 = k. one sees that x → Ai xsgn(Ci x) is continuous on the surface Σi = {x ∈ IRn | Ci x = 0}. a. T ) → IRn that is Lipschitz continuous. are hypomonotone. Let us state other cases where (59) fits within Proposition 1. is hypomonotone with any Remark 2 As noted in [55] chattering may be due in sliding mode control appplications to the presence of parasitic dynamics.. The linearized Stribeck friction model (with multivalued part at zero tangential velocity) [38] is hypomonotone. 1 i m. The mapping F : IR → IR. Di = 0 and Ker(Ci ) ⊆ Ker(Ai ).. .. Exercise 12. Then for any initial data the differential inclusion (59) has a unique solution x : (0. Indeed (x + 1)sgn(x) = |x| + sgn(x) and x → |x| + kx is monotone for any k 1. k 8√ 3 1 1+x2 sgn(arctan(x)). Simple modelling of these yield when inserted in (1) the differential inclusion (see [55. Moreover it is Lipschitz continuous as it is piececewise linear. x → [−1. k2 be reals. x(t)) − g(x(t))Sgn(h(x(t)). T ) → IRn that is Lipschitz continuous.e. x → (x + 1)sgn(x) is hypomonotone. 1] if x = 0 hypomonotone with constant k for any k max(|k1 |. sgn(hm (x)]T ∈ IRm . Under the lemma’s conditions. (7) (8)]) 0 0 0 x(t) ˙ x(t) 1 x˙ s (t) = 0 0 1 xs (t) − 0 sgn(Cx(t)) (61) 1 1 2 x ¨s (t) − − x ˙ (t) 0 2 2 s τ τ τ INRIA . T ) → IRn that is Lipschitz continuous. inria-00374840. Corollary 3 Consider the differential inclusion in (59).

that is known to greatly influence the stability of systems and usually may yield instability. The mere existence and uniqueness of solutions for such relative degree 3 systems is not trivial. version 1 . It follows from [24. k 0. sk+1 = sgn(xk ) (67) xk+1 = 1 + hsgn(xk ) RR n 6886 .e. Lemma 5 For all 1 > h > 0 and x0 ∈ IR. a. Let us now illustrate this on the following system with hypomonotone multivalued part: x(t) ˙ ∈ −(x(t) + 1)sgn(x(t)) + u(t) (64) with x(t) ∈ IR.3.1 (63) x(0) = x0 m i=1 ρi = ρ. on (0. xk0 +n+1 − xk0 +n = h Proof:If x0 ∈ [−h. We may discretize it as: −(xk+1 − xk − huk − hρxk ) ∈ h(xk+1 + 1)sgn(xk+1 ) + hρxk+1 . Such parasitics may be seen as a non collocation issue. x0 = x(0) (65) Notice that we may rewrite (65) as 0 ∈ xk+1 − (xk + huk + hρxk ) + h(xk+1 + 1)sgn(xk+1 ) + hρxk+1 (66) Let us denote the mapping in the right-hand-side of (66) as F (xk+1 ). Definition 2. where 0 0 0 0 1 . in practice the intersection may be computed as follows. This system does not fit within the above classes B = (1 0 0)T and A = 0 1 2 1 − − τ2 τ2 τ of inclusions. T ) x(0) = x0 (62) xk+1 −xk h ∈ f (tk .3. h]. One interesting question is to determine what kind of solution is approximated by the backward Euler method applied to (61) which. 7. xk ) − ρxk − m i=1 (Ai xk+1 + Bi )sgn(Ci xk+1 + Di ) + ρi xk+1 . then k0 = 0. For u(t) = 0. Similar conclusions hold for the other form of parasitics in [55. on (0. inria-00374840. T ) where the ρi are the hypomonotonicity constants and applies to (62) and (63). there exists k0 such that xk0 +n = 0 and 0 for all n 1. Theorem 1] has a unique forward solution since its leading Markov parameter is CA2 B = τ12 > 0.1] for all ρ 1. the solution is given by : xk − hsgn(xk ) . but a unique so-called forward solution. the following Lemma extends the Lemma 1. xk ) − ρxk − g(xk+1 )Sgn(h(xk+1 ) + ρxk+1 . C) of this system is r = 3. a. according to [42. A possible solution for this non collocation issue is the observer design of Example 4. Therelative degree of the triplet (A. Otherwise.10 Apr 2009 The differential inclusions in (59)–(60) are therefore discretized as follows: and xk+1 −xk h ∈ f (tk . B. The result of Proposition 2 A simple hypomonotone case As shown in Section 2 on a simple monotone example. for k < k0 and xk ∈ / [−h. In [42] an example is given that possesses an infinity of absolutely continuous Filippov’s solutions. The set-valued mapping F (·) is strongly monotone [24.Implicit Euler numerical simulations of sliding mode systems 17 with C = (0 1 0). Theorem 2. h].3] that the generalized equation 0 ∈ F (xk+1 ) has a unique solution.e. (3) (4)].

Clearly then the terms 0 h for any h > 0. k 0. According to the class of systems (13). Let us now proceed with the time discretization of (70) by a fully implicit scheme : xk+1 = xk + hf (xk+θ . tN = T is constructed. version 1 . The main discrepancy lies in the choice of a θ-method to integrate the nonlinear term. and a grid t0 = 0. we have to solve h xk0 +2 = −hsk0 +2 inria-00374840. The approximation of a function f (·) on [0. h] can be easily found. let us introduce a new notation. (59) or (60) that we are studying the functions f (·) and g(·) are defined either in a fully nonlinear framework or by affine functions. For θ = 0. 1] and γ ∈ [0. The interval of integration is [0. The only possible solution for xk0 +1 − xk0 = −h(xk0 +1 + 1)sk0 +1 (68) s k0 +1 ∈ sgn(xk0 +1 ) is xk0 +1 = 0 and sk0 +1 = xk0 .sk0 +n . Let us now consider that xk0 ∈ [−h. As in [1].1 Time–discretization Starting from (60). xk+γ = γxk+1 + (1 − γ)xk . T > 0. constant on the intervals [tk . The time-step is h > 0. tk+1 = tk + h.4. This time-discretization is slightly more general than a standard implicit Euler scheme. n 3. we call the problem (71) the “one–step nonsmooth problem”. h]. redoing the xk +n+1 − xk0 +n approximating the derivative.10 Apr 2009 s k0 +2 (69) ∈ sgn(xk0 +2 ) and we obtain xk0 +2 = 0 and sk0 +2 = 0. tk+θ ) − hg(xk+γ )sk+1 yk+1 = h(xk+1 ) (71) sk+1 ∈ Sgn(yk+1 ) where xk+θ = θxk+1 + (1 − θ)xk . t) − g(x(t))s(t) y(t) = h(x(t)) (70) s(t) ∈ Sgn(y(t)) where s(t) ∈ IRm and y(t) ∈ IRm are complementary variables related through the Sgn(·) multi– valued mapping. 8 Detailed Implementation of Implicit Euler Discretization of the general case (60) This section is devoted to the implementation and the study of the numerical algorithm. 8. T ] is denoted as f N (·). We conclude that in this case also the system and its derivative are correctly approximated at the zero value on the sliding surface. The same holds for all xk0 +n . T ]. We present the time-discretization in its full generality and specialize the algorithms for each case in Section 8. the step k0 for which xk0 ∈ [−h. 1]. x(t) ˙ = f (x(t). we INRIA . and is a piecewise constant function. and tk+θ = θtk+1 + (1 − θ)tk+1 . We denote f N (tk ) as fk .18 Acary & Brogliato From the solution (67). For the next iteration. There is no spurious oscillation around the switching surface. with θ = [0. are zero same reasoning. tk+1 ).

where IR Note that the problem (72) implies that −F (z) ∈ N[l.u] (z). i = n + 1 . (73) The relation (73) is equivalent to the MCP (72) if we assume that w is the positive part of F (z).2 Mixed Complementarity Problem inria-00374840. n +1. n 0. For instance. i = 1 . . . or if we have to use large step sizes for practical reasons. the term g(·) is explicitly evaluated. tk+θ ) + hg(xk+γ )sk+1 −h(xk+1 ) −∞. Definition 2 (MCP) Given a function f : IRq → IRq and lower and upper bounds l. . i = n + 1 . F (z)) and v is the negative part of F (z). n wi = . m . i = n + 1 . the choice of θ > 1/2 offers better stability properties with respect to h. . .Implicit Euler numerical simulations of sliding mode systems 19 retrieve the explicit integration of the smooth and single valued term f (·). i = 1 . m yi+ 0. . Moreover for γ = 0. . the stability and the numerical damping of the proposed method. the bounds u and l and the condition (z − l)T w = 0. m If z solves the MCP (74). that is w = F + (z) = max(0. (u − z)T v = 0 imply that 0. u ∈ IR find z ∈ IRq . . vi = (75) yi− 0. . . ui = (74) +∞. The flexibility in the choice of θ and γ allows the user to improve and control the accuracy. . i = 1 . that is v = F − (z) = max(0. w. i = 1 . n −1. 8. v ∈ IRq+ such that F (z) = w − v l z u (72) (z − l)T w = 0 (u − z)T v = 0 ¯ = IR ∪ {+∞. −F (z)). −∞}. . . m The MCP is then given by xk+1 − xk − hf (xk+θ ) + hg(xk+γ )sk+1 0 = −h(xk+1 ) y − (xk+1 ) − y + (xk+1 ) = −y(xk+1 ) −1 < sk+1 < 1 (76) (sk+1 + 1)T y − (xk+1 ) = 0 (1 − sk+1 )T y + (xk+1 ) = 0 It is clear that the problem (71) is equivalent to the MCP defined by (74). i = n + 1 . version 1 . . The One–step nonsmooth problem as a MCP Let us define the MCP by xk+1 sk+1 z= F (z) = li = xk+1 − xk − hf (xk+θ .10 Apr 2009 The so-called ”one–step nonsmooth problem” (71) appears to be a Mixed Complementarity Problem (MCP) that we have to solve at each time–step. Let us define what is a MCP : ¯ q. if the smooth dynamics given by f (·) is stiff. . . The results of existence and uniqueness of solution of (72) or equivalently (73) are related to the monotonicity properties RR n 6886 .

tk+θ ) + hγ∇x g(γx + (1 − γ)xk ) ⊗ (81) At each time–step. version 1 . we get the following linear relation between yk+1 and sα+1 k+1 . α+1 α+1 α α α −1 α α (xk+1 . denoted by M (x. we propose to solve the ”one–step nonsmooth problem” (71) by performing an external Newton linearization. sk+1 ) = xk+1 − xk − hf (xk+θ ) + ∇x R(xk+1 .10 Apr 2009 ¯ denotes the simple contracted tensor product and the third–order tensor ∇x g(x) is the where ⊗ Jacobian of g with the respect x given by the following component: (∇x g(x))klp = 8. The same operation is performed with the second equation of (71) leading to the following linearized equation α+1 α+1 α α yk+1 = yk+1 + ∇x h(xα k+1 ) xk+1 − xk+1 (84) α+1 Inserting (83). s) = x − xk − hf (θx + (1 − θ)xk . sk+1 ) = 0 (79) with R(x. tk+θ ) + hγ∇x g(xk+γ ) ⊗ . tk+θ ) − hg(xk+γ )sk+1 .3 ∂gkl (x) . M (x. k+1 = xk+1 + M (83) that is The matrix M is clearly non singular for small h. (82) α+1 α −1 α α α α xα+1 (xk+1 . Without entering into further details. The Jacobian matrix is explicitly given in the case defined in (74) by ¯ sk+1 hg(xk+γ ) I − hθ∇x f (xk+θ . s) = I − hθ∇x f (θx + (1 − θ)xk . s) leads to ¯ s. sα k+1 ) xk − xk+1 + hf (xk+θ . α+1 α+1 α α α α α xα k+1 − xk − hf (xk+θ . The solution of this system α of nonlinear equations is sought as a limit of the sequence {xα k+1 . The numerical implementation of the MCP solvers are often based on the computation of the Jacobian matrix of the function F (·) with respect to z. Numerical Solvers The MCP (72) can be solved by a large family of solvers based on Newton– type Methods and interior-points techniques. ∇z F (z) = (77) ∇x h(xk+1 ) 0 inria-00374840. sk+1 }α∈IN such that 0 xk+1 = xk α+1 α+1 α+1 α α α α α α RL (xα+1 k+1 . tk+θ ) − hg(xk+γ )sk+1 ) k+1 ) M (85) INRIA . sk+1 )(xk+1 − xk+1 ) + hg(xk+γ )sk+1 = 0. tk+θ ) + hg(γx + (1 − γ)xk )s. sα yk+1 = yk+1 + ∇x h(xα k+1 )(xk − xk+1 + hf (xk+θ . 24] for the main results. sk+1 )(xk+1 − xk+1 ) + hg(xk+γ )sk+1 = 0 (80) The computation of the Jacobian of R with respect to x.20 Acary & Brogliato of F (P-properties or coherent orientations of the associated affine map for particular cases of bounds and affine function F (·). ∂xp (78) Newton’s linearization and Mixed Linear Complementarity Problems Due to the fact that two of the systems that are studied in this paper involve affine functions f (·) and g(·). s) = ∇x R(x. we refer to [30. tk+θ ) + M (xk+1 . which yields a Mixed Linear Complementarity Problems (MLCP). Newton’s linearization The first line of the problem (71) can be written under the form of a residue R depending only on xk+1 and sk+1 such that R(xk+1 . We refer to [11] for a comparison of several solvers based on Newton’s method. we have to solve the following linearized problem.

sk+1 . t) = I RR n 6886 (91) . θt + (1 − θ)tk ) (90) α+1 If CB > 0 then the matrix Wk+1 is also positive definite for sufficiently small h. α+1. interior point methods and semi-smooth Newton methods (see [1] for an overview).4 The special cases of the affine systems In this section.4. One obtains. 8. In the worst case α+1 when the matrix Wk+1 has no special properties. The matrix Wk+1 and bα+1 k+1 are then given by α+1 α Wk+1 = hCM −1 (xα k+1 . the MCLP can be always solved by enumerative solvers for which various implementations can be found.1 Time–discretization of the system (13) For the system (13). sα k+1 )M k+1 )g(xk+γ ) (87) inria-00374840. This result ensures the existence and uniqueness of a solution of the MCLP. With some positiveness properties [24]. Among these methods. we can cite the family of projection/splitting methods. Furthermore. α+1 that is h(x) = Cx + D. tk+θ )) k+1 = yk+1 + C M M (x. 8.+ α+1 α+1 α+1 yk+1 − yk+1 = −Wk+1 sk+1 + bα+1 k+1 α+1.Implicit Euler numerical simulations of sliding mode systems 21 Mixed linear complementarity problem (MLCP) To summarize. sα k+1 )(xk − xk+1 + hf (xk+θ . standard methods for LCP[19] can be straightforwardly extended. we specify the time–discretization to the two other classes of systems (13) and (59) and for particular value of θ and γ. tk+1 )(xk − xα bα+1 k+1 + hf (xk+θ .10 Apr 2009 α α −1 α α α bα+1 (xk+1 .− (88) 0 (sα+1 0 k+1 + 1) ⊥ yk+1 α+1. t) = I − hθ∇x f (θx + (1 − θ)xk . tk+θ )) k+1 = yk+1 + ∇x h(xk+1 ) M The problem (86) is equivalent to a MLCP which can be solved under suitable assumptions by many linear complementarity solvers such as pivoting techniques. the problem to be solved in each Newton iteration is: α+1 α+1 α+1 α+1 yk+1 = −Wk+1 sk+1 + bk+1 (86) α+1 α+1 sk+1 ∈ Sgn(yk+1 ) with W ∈ IRm×m and b ∈ IRm defined by α+1 −1 α α Wk+1 = h∇x h(xα (xk+1 . interior point techniques and splitting/projection strategies. The reformulation into a standard MLCP follows the same line as for the MCP in the previous section. version 1 . standard pivoting techniques such as Lemke’s method or projection/splitting such as Projected Successive OverRelation (PSOR) compute the solution.+ 0 (1 − sα+1 0 k+1 ) ⊥ yk+1 As for the MCP. the function g(x) is reduced to the matrix B and the function h(x) is affine. there exists numerous methods to numerically solve MLCP. Semi-implicit discretization with θ = 0 The matrix M is then reduced to M (x. tk+1 )B with (89) α α −1 α (xk+1 .

The MCLP defined by (99) can be simplified to Wk+1 = hCg(xk )bk+1 = yk + C [hf (xk . t) = Ex(t) + F .2 Time–discretization of the system (59) For the system (59). the matrix M reduces to M (x. s) = x − xk − hf (xk .kp xp + Bl . In this particular case. tk+θ )) k+1 = yk+1 + C M Semi–implicit discretization with θ = γ = 0 If γ and θ vanish. (95) The components of g(·) can be explicitly expressed by n gkl (x) = Al. = [hi (x) = Ci x + Di . the Jacobian of g(·) is given by (∇x g(x))klp = The Jacobian of h takes the following simple form: ∇h(x) = C = [Ci . we have to solve at each Newton’s loop the MLCP (86) with α+1 α α Wk+1 = hCM −1 (xα k+1 . Fully implicit discretization with an affine function The same conclusion can be drawn if ∇x f (θx + (1 − θ)xk . In this case.kp . tk ) + hg(xk )s (100) is linear in x and s. tk ))] (101) INRIA . Furthermore. t) = I − hθE inria-00374840. the residue R given by R(x. there is no need to perform Newton’s iterations. i = 1 . j = 1 . (96) ∂gkl (x) = Al. tk ) In this particular case. i = 1 . . . θt + (1 − θ)tk ) is equal to a constant matrix E that is when f (·) is linear time-invariant and given by f (x. tk+1 )g(xk+γ ) (99) α −1 α α bα+1 (xk+1 . m] ∈ IRm . .10 Apr 2009 and (93) W = hC(I − hθE)−1 B (94) bk+1 = yk + hC(I − hθE)−1 [Exk + a] 8. ∂xp (97) p=1 and therefore. . the MLCP has a solution for any h > 0 under the assumptions that CB > 0. . . (98) After the newton linearization.4. we recall that g(·) and h(·) are given by g(x) h(x) = [g•j (x) = Aj x + Bj . tk+1 )(xk − xα k+1 + hf (xk+θ . m] ∈ IRn×m .22 Acary & Brogliato and W = hCB (92) bk+1 = yk + hCf (xk . there is no need to perform some Newton iterations because the system to be solved at each time–step is linear. sk+1 . version 1 . m] ∈ IRm×n .

tk+1 ) − hγ∇x g(xk+θ ) ⊗ sk+1 ) α+1 α −1 α α α Wk+1 ← h∇x h(xk+1 )M (sk+1 .Implicit Euler numerical simulations of sliding mode systems 23 inria-00374840. 2. yk+1 with F. y0 ← y(0) = h(x(0)). . γ numerical integration parameters Require: ε Newton’s method tolerance Ensure: ({xk }. h Require: x(0) the initial condition Require: t0 . g. {yk }). tau0 ← 0 // Time integration loop while tk < T do Solve the MCP (76) for xk+1 . x0k+1 ← x0k . sk+1 ) > ε do −1 α α α ¯ α −1 . //time iteration tk ← tk+1 . y0 ← y(0) = h(x(0)). γ numerical integration parameters Ensure: ({xk }. {yk }). sk+1 . . . //Update xk ← xk+1 . tau0 ← 0 // Time integration loop while tk < T do 0 α ← 0. T time–integration interval Require: h time–step Require: θ. sk ← sk+1 . sk+1 ) ← (I − hθ∇x f(xα k+θ . x0 ← x(0). yk+1 ← yk0 //Newton’s loop α while R(xα k+1 . k ∈ {1. M (xk+1 . sk+1 )g(xk+1 ) α+1 α −1 α α α bk+1 ← yk+1 + ∇x h(xα k+1 )M (sk+1 ) xk − xk+1 + hf(xk+1 ) α+1 α+1 Solve the MLCP (88) for yk+1 . yk+1 ← yk+1 xk+1 ← xα k+1 . 2. sk+1 α+1 α+1 α sα yk+1 ← xk+1 k+1 ← sk+1 . T time–integration interval Require: h time–step Require: θ.} k ← 0. version 1 . . k ← k + 1 end while RR n 6886 . α+1 α α α −1 α α xα k+1 ← xk+1 + M (sk+1 ) xk − xk+1 + hf(xk+1 ) + hg(xk+1 )sk+1 α ←α+1 end while //Update α+1 α+1 sk+1 ← sk+1 . s0k+1 ← s0k .} k ← 0. h Require: x(0) the initial condition Require: t0 . yk ← yk+1 //time iteration tk ← tk+1 . l and u given by (74) and the Jacobian ∇z F (z) given by (77). k ∈ {1. x0 ← x(0).10 Apr 2009 Algorithm 1 Implicit Euler time-discretization with a generic MCP solver Require: System definition: f. . {sk }. {sk }. k ← k + 1 end while Algorithm 2 Implicit Euler time-discretization with an external Newton loop and a MLCP solver Require: System definition: f. . g.

INRIA . The Algorithm 1 describes the implementation with a generic MCP solver and the Algorithm 2 describes the numerical implementation of the algorithm with an external Newton linearization and a MCLP solver. In the case of the system (13) with an affine function f (·) or θ = 0. the MLCP matrix W can be computed before the beginning of the Newton loop.24 8. version 1 . In the case of the system (59) with θ = γ = 0.5 Acary & Brogliato Algorithms inria-00374840. the MLCP matrix W can be computed before the beginning of the time loop.10 Apr 2009 We propose in this section two algorithms to sum-up the numerical implementation of the implicit Euler time–stepping scheme. saving a lot of computing effort.

the LTI system with an ECB-SMC controller is defined by the following data. The results are reproduced on Figure 6(a). 9. The results displayed on Figure 5 show 0 5 that the system reaches firstly the sliding surface 2x2 + x1 = 0 without any chattering. (104) F = 1 −a1 −a2 Starting from the initial data. version 1 . Another example taken from [57] in the MIMO case is given by the following parameters.55. On Figure 6(b). 0 0 1 .2. the Implicit ZOH discretization as proposed in Section 5. This example illustrates Lemma 3 since CB = 9.01 and with c1 = 1 and α = 1. t) = 0 (103) 5 0 . Galias and Yu [27] have shown that the Explicit ZOH discretization of the system with a1 = −2. c1 = 1 and h = 0.3 exhibits a period–2 orbit. 0.3. Indeed. In the notation of Section 5. (105) 0 1 1 −1 −3 1 0 1 Similar results are depicted on Figure 8. D = 0. the sliding manifold is reached in finite time and without any chattering.2 Example 3: Multiple sliding surfaces Let us consider the example 3. On Figure 4.2 is free of chattering.fr/ RR n 6886 . The system can be defined in the form (13) with B= 1 2 2 −1 . h = 0. ·). the trajectories obtained by implicit discretization are shown using the implicit Euler discretization with h = 1. 0 0 1 0 0 1 0 1 1 1 . f (x(t). G = 1 0 .1 Chattering free stabilization Let us consider the following continuous–time closed loop system from [26] given by inria-00374840. (102) As it is shown in [26] the trajectories obtained by an explicit Euler discretization exhibit spurious oscillations which are described by period-2 cycle around the sliding manifold. a2 = 2. C = c1 1 . 3] which is designed for the simulation of multivalued nonsmooth systems. Note that the algorithm is also very robust in the sense that the simulation can be performed with relatively large time–steps. The system then slides on the surface up to reaching the second sliding surface 2x1 − x2 = 0 and comes to rest at the origin.inria. As it has been predicted by theoretical discussions of Section 4.10 Apr 2009 x˙ = 0 0 1 −c1 x− 0 α sgn( c1 1 x).3 Extensions to ZOH discretized systems The extension to ZOH discretized systems is illustrated on a first example taken from [27]. x0 = [0.1 and h = 0. 55]T .Implicit Euler numerical simulations of sliding mode systems 9 25 Numerical experiments Let us illustrate the above developments with numerical integrations performed with the siconos software platform of the INRIA2 [1. 9.gforge. G= . a comparison is given between the ZOH and the Euler discretization of the vector field f (·. C = F = 1 . C= 1 2 2 −1 . h = 0. On Figure 7. the matrix CB = α = 1 satisfies the assumptions of Lemma 3. 2 http://siconos.

2 0.4 0.2 0.1.6 0.5 2 0 0.6 0.5 2 2 1.4 0 s values 0 x2(t) state x1(t) and x2(t) -0.5 2 time t (c) sgn function s1 (t) and s2 (t) Figure 5: Multiple Sliding surface.5 -0.2 0.2 0.5 -0.2 0. c1 = 1. Implicit Euler (f) h = 0.6 0.3.2 x1(t) (e) h = 0.5 -1 -1 -1. Implicit Euler Figure 4: Equivalent control based SMC.4 0.5 1 x2(t) x2(t) 1 0.5 0 0 -0. −1.02.5 0 0.6 0.5 1 1. Explicit Euler (b) h = 0. Implicit Euler (d) h = 0.4 x1(t) 0.5 0.5 -0.5 -0.0.8 -1 -1 -1 0 0.7 0 x1(t) 0.8 1 1.4 0.6 -0.5 -1 -1 0 0.5 2.3 0.10 Apr 2009 0 0.8 0.5 2 1.1 0.5 0 0 -0.2 0.0]T INRIA .8 0.3 0.26 Acary & Brogliato 2.4 0.1 0.2 0.7 0.7 0.5 0.5 0 0 -0.8 -1 -0.5 2 2 1.4 0.5 0.6 0. State x1 (t) versus x2 (t).05.5 0.21]T . h = 0. 2.1 0 0.5 -0.8 1 (a) state x1 (t) and x2 (t) versus time (b) phase portrait x2 (t) versus x1 (t) 0 0.5 0.3.3 0.5 1.1 0.5 2.3 0.5 x2(t) inria-00374840. α = 1 and x0 = [0.1. Implicit Euler 2.8 1 x1(t) (a) h = 0. Explicit Euler 2. 0 x1(t) x2(t) 1 s1(t) s2(t) 1 0.6 0. x(0) = [1.4 0.9 x1(t) 0.5 1 1 x2(t) 2 1.5 1 x2(t) x2(t) 1 0.9 x1(t) (c) h = 1.5 -1 -0.5 1.5 0.9 1 0 x1(t) 0.2 0. version 1 .5 0.5 1 time t 1.6 0.5 -0.2 0.5 -0.5 2.1 0.7 0.

55.7 0.02 0.01 -0.8 -1 -0. 55]T State x1 (t) versus x2 (t). version 1 .3 x2(t) 0.3 0.05 0.6 0.4 -0.05 0.1 (a) h = 0. x0 = [0.6 0.04 -0.02 -0.5 0.2 -0.6 SlidingSurface ExplicitZOH ExplicitEuler SlidingSurface ImplicitZOH ImplicitEuler 0.1 0 0. a1 = −2.01 0 -0.6 0.1 0 x1(t) 0.4 0.2 0 x2(t) 0 x2(t) inria-00374840.04 -0.3 0.4 0.1 0.01 -0.2 0.Implicit Euler numerical simulations of sliding mode systems 0.6 0.4 0.6 x2(t) 0.1 -0. 0.55.06 0.3.1 0.2 -0.04 -0.4 -0.3.02 0. Explicit ZOH x1(t) 0.1 0 x1(t) 0.5 0.2 -0.2 0 0 x2(t) 0.3.3 0.01 0. Explicit ZOH (b) h = 0.8 -1 -0.1 0 0 -0.03 -0.02 -0.06 0.02 -0.2 0. c1 = 1 and h = 0.5 -0.3. −0.5.01 -0.01 0 -0.8 0 0. Implicit ZOH Figure 8: Equivalent control based SMC.1 x1(t) 0.2 0. a2 = 2.6 -0.03 -0.4 0.6 -0.4 0.2 0.2 0.5 0. 0.03 -0.3 0.8 x1(t) (a) Explicit implementation (b) Implicit implementation Figure 7: Comparison of Euler and ZOH discretizations of ECB-SMC system.6 -0.06 0.6 -0.7 0. Implicit ZOH Figure 6: Equivalent control based SMC.6 -0.05. x0 = [0. 0.03 -0.04 -0.4 0.02] RR n 6886 .01 0.5 0.3 x2(t) -0.3.3. a1 = −2.5 0.7 x1(t) (a) h = 0.05 -0.1 0. Trajectory with initial point x0 = [0.6 0.4 0.8 -0.4 -0.02 -0.4 -0.4 0. x3(t) x3(t) 0.7 0 0.4 0. a2 = 2.10 Apr 2009 27 -0.2 0.8 -0. 0.10 (b) h = 0.4 -0. 55]T State x1 (t) versus x2 (t).2 0. c1 = 1 and h = 0.2 0.05 0.2 -0.

the implicit method yields a smooth stabilization at x = 0 whereas the explicit Euler has significant chattering. The results displayed on Figure 9.1. t) = 0. h = 0. The sufficient conditions of the Lemma 3 are not satisfied when seems to indicated that these conditions has to be improved.2 -0. Figure 9(c) illustrates the fact that the controller varies inside the multivalued part of the sgn function in order to assure the existence of an equilibrium point.6 0.5 show that the system reaches the origin without any chattering.8 0. 9.2 -1 0 2 4 6 8 10 0 2 time t 4 6 8 10 time t (c) Control u(t) vs.2 0 2 4 6 8 10 0 2 4 time t 6 8 10 time t (a) State x1 (t) vs. h = 0. Explicit Euler Euler Figure 9: Lyapunov-based discontinuous robust control.4 0. that x = 0.1. Is is obvious that.6 0. Explicit Euler Euler 1 1 0.6 state x1(t) inria-00374840.5 control u(t) control u(t) 0. f (x(t).8 0.4 Acary & Brogliato Lyapunov-based robust control We propose in this section to give an numerical example which fits with the example (18) of a Lyapunov-based discontinuous robust control.2 0 0 -0. time. time. h = 0.8 0. h = 0.5 0 -0.28 9.2 -0.10 Apr 2009 state x1(t) with γ(t) = α sin(t) and u(t) = sgn(x(t)).2 0. D = 0.6 Example 4: Observer based SMC Let us illustrate the performance of our implementation on the observer based SMC described by the Example 4.1.4 0. Implicit (d) Control u(t) vs. Implicit (b) State x1 (t) vs.5 The Filippov example Example 9 Let us consider now the well known Filippov example which can be defined in the form (13) with B= 1 −2 2 1 . Let us consider the following system x(t) ˙ = −x(t) − u(t) + γ(t) (106) 1 1 0. 0. as expected.1. The dynamics is given by (17) with k = 1 and τ = 0. The initial conditions INRIA . C= 1 0 0 1 .1 α = 0. time.001. h = 0.1 9. version 1 . time.4 0 0. (107) The trajectories may slide on the codimension 2 surface given by Cx = 0.

6 2 -0.2 0.5 1 1.0]T RR n 6886 1 .4 0.4 -0.5 0 -0.5 0 0. h = 0. Filippov Example. x(0) = [1. −1.2 -0.Implicit Euler numerical simulations of sliding mode systems 29 0.10 Apr 2009 0 0.2 0 0.6 -0.5 -0.5 -1.8 -1 -1 -1.5 2 time t (c) sgn function s1 (t) and s2 (t) Figure 10: Multiple Sliding surface.4 0 x2 (t) state -0.002.2 -1.8 x1 (t) time t (a) state x1 (t) and x2 (t) versus time (b) phase portrait x2 (t) versus x1 (t) s1 (t) s2 (t) 1 0.5 s values inria-00374840.6 0.0.5 1 1.5 -1 0 0. version 1 .4 -0.2 x1 (t) x2 (t) 1 -0.

the parameter θ has a huge influence on the stability of the integration.0015 0. k = 1 and τ = 0.0005 0. For instance. h = 0.0005 0 -0.8 0. Unfortunately. On the contrary.0005 Control 0. The stability is retrieved for h < 0. 0.1 θ = 1. the instability of the scheme for h = 0.6 Control inria-00374840. the parameter γ has no influence on the numerical time–integration. 0]T . γ = 1. γ = 1 On the three previous examples. one sees a very accurate and smooth stabilization on the sliding surface. INRIA . This is not the case for the explicit Euler θ = 0 and for the chosen parameters k and τ .0.002 Error 0.2 -0.001 0.005.4 0 0.0035 0. On Figure 13. version 1 . 0. 0.1 does not allow to proceed to integration. k = 1 and τ = 0. On Figure 11. the implicit Euler integration (θ = 1) of the smooth term f (·) is unconditionally stable. 0]T . we can observe the behavior of the command without any chattering. the instability of the scheme is illustrated and appears as a chattering on the state x. 0.003 0. The numerical parameters are given by h = 0.0005 0 -0.001 0. the error between the reference command and the observer state is given.001 0 2 4 6 time (a) control u(t) = Sgn(Cx(t)) 8 10 0 2 4 6 8 10 time (b) Zoom on control u(t) = Sgn(Cx(t)) Figure 12: Observer based SMC: Control.001. γ = 1 1 0.0015 0 2 4 6 8 10 time Figure 11: Observer based SMC: Error e(t).10 Apr 2009 -0. Due to the fact that the function g(·) is linear and reduced to a matrix B = [1. we discuss the role of the numerical parameters θ and γ.1 θ = 1. Indeed.30 Acary & Brogliato are chosen as [2. even for values of h not so small. 0.2 -0. Influence of the integration parameters θ and γ In the following numerical experiments.001 -0.1 that is a sampling of 10Hz and θ = 1.0025 0. Other choices of θ can be made to improve the numerical time–integration of the smooth dynamics.001. On Figure 12. θ = 1/2 yields a second order scheme for the integration of f . h = 0. the scheme is not of second order due to the fully implicit integration of the multi-valued part.

6 -0. version 1 .5 0 0.5 1. γ = 1 RR n 6886 8 10 .001.5 1.01 θ = 0.4 1.5 2 0 2 4 time 6 time (a) state x1 (t) (b) control u(t) = Sgn(Cx(t)) 80 60 40 20 Error inria-00374840. k = 1 and τ = 0. h = 0.8 0 -20 -40 -60 -80 0 2 4 6 8 10 time (c) error e(t) Figure 13: Observer based SMC.3 -1.Implicit Euler numerical simulations of sliding mode systems 2 31 1 1.7 x1 1.10 Apr 2009 0.9 0 Control 1.5 -1 1.5 1 1.

Houston J. 24(3):545–569.01. Despite the backward Euler method has been studied and used for a long time in other fields like contact mechanics and electric circuits simulation [1]. 2009..5 -0. I: Blending. [2] V. The novelty compared to numerical simulation is that this time one has to consider not only the numerical simulation. INRIA. Seidman.8 1. even in case of codimension larger than one. http://hal. Alexander and T. This work therefore constitutes the introduction of a new discretization method for EBC-SMC systems. However as shown in this paper this is not the case. k = 1 and τ = 0.2 -0.5 0. γ = 1 10 Conclusions In this paper the backward Euler method is studied on specific classes of Filippov’s systems that encompass sliding mode control systems. P´erignon. On the Figure 14. Sliding modes in intersecting switching surfaces.32 Acary & Brogliato Nevertheless. 1999.I. The parameter τ has been modified to 0. but also the implementation on real processes. [3] V. 2 1 θ=1 θ = 1/2 θ=1 θ = 1/2 0.C.10 Apr 2009 -0. Sliding modes in intersecting switching surfaces. References [1] V. Alexander and T. 2007. Brogliato. version 1 .fr/inria-00162911/en/. Springer Verlag.4 0 5 10 time (a) state x1 (t) 15 20 0 5 10 15 20 time (b) control u(t) = Sgn(Cx(t)) Figure 14: Observer based SMC with θ = 1 and θ = 1/2. volume 35 of Lecture Notes in Applied and Computational Mechanics.6 1 state control 0.I. Technical Report RR 1234567890. INRIA . Technical Report TR-0340.inria. This paper paves the way towards the study of a new family of discrete-time sliding mode controllers.4 0. Perhaps one obstacle to the dissemination of the method is that at first sight. Acary and B.5 0 0 inria-00374840. [5] J.1.. Acary and B.5. the controller designed from a backward philosophy looks like a non causal controller. Numerical Methods for Nonsmooth Dynamical Systems: Applications in Mechanics and Electronics. 2008. INRIA. 25(1):185–211.2 0. Math. Brogliato. two simulations of the observer based SMC are presented for θ = 1 and θ = 1/2. Math. Acary and F. it can be interested to use of θ = 1/2 to decrease the numerical damping of the scheme on the smooth term. it seems it has not yet been used in the sliding mode control community. An introduction to siconos. 1998. Houston J. It is shown that such implict schemes allow a smooth accurate stabilization on the sliding surface. Seidman. h = 0.5 to correctly integrate the parasitic dynamics with the same time–step h = 0. [4] J.C. Implicit Euler numerical simulation of sliding mode systems. II: Hysteresis.

and E. and stability. 2008. volume 334 of LNCIS. I. An adaptive sliding mode control scheme for induction motor drives. Egeland. RR n 6886 . W. Billups.C. 1996. Modern sliding mode control theory. A comparison of large scale mixed complementarity problem solvers. Discontinuous dynamical systems.. 2006. Brogliato. Practical stabilization of a class of nonlinear systems with partially known uncertainties. Advances in Variable Structure and Sliding Mode Control. 1995. Boston. L. Cottle. 2008.H. 31(1):145–150. S. Int. Facchinei and J.H.. [15] B. Fridman. A. version 1 .P. [19] R. Bastien and M. 23-49 (2008). Springer Verlag NY. B. and O. Absolute stability and the Lagrange-Dirichlet theorem with monotone multivalued mappings. [18] J. Springer Verlag. Heemels. London. Stone. [14] B. ESAIM M2AN: Mathematical Modelling and Numerical Analysis. 7:3–25. editors. 1992.. Garrido. Computational Optimization and Applications. Fossas Colet. Persoz’ gephyroidal model described by a maximal monotone differential inclusion. Difference methods for differential inclusions: a survey.L. Numerical precision for differential incusions with uniqueness. [21] S. Springer Verlag. Springer Verlag. S. Mech. 1(1):73–78. Usai. Alkorta. [22] A. J. [8] G. Brogliato and A. 14(1):1–23. Modern Sliding Mode Control Theory.J. Lecture Notes in Control and Information Sciences 375. and J. An inertial proximal algorithm with dry friction: finite convergence results.Implicit Euler numerical simulations of sliding mode systems 33 [6] B. Schatzman.. 1995. L. 78(5). Pang. Cortes. 34(2):263–294. IEEE Trans. Dontchev and F. A. Cabot. 2002. Berlin: Springer. W. Systems and Signal Processing. Optimization Methods and Software. Edwards. E. Lozano. June 2008. Ferris. Brogliato. [7] O. 1992. 2007. 1997.M. [12] I. pages 36–73.) et al.10 Apr 2009 [10] J. Dirkse. A. Barambones. and M. a tutorial on solutions.P. Pang. [17] M. Automatica. K. volume I & II of Springer Series in Operations Research. volume 375 of LNCIS. Circuits. [24] F. Cao and M. 2006. 2nd edition. Bastien and C.C. A pivotal method for affine variational inequalities. 2003. Maseda. R. [13] B. [16] M. nonsmooth analysis. 2004. Giorgio (ed. Inc. Garrido. Boiko. and R. 36(3):427–460. Finite-dimensional Variational Inequalities and Complementarity Problems. 49:349– 357.M.C. Baji and A. C.-F. 5:123–156. [9] J. Systems and Control Letters. Maschke. Bartolini. editors. Inc. 38(1):184–191. Fridman. MA. Dirkse and M. Mathematics of Operations research. Trofino Neto.. Schumacher.J. Pisano. Danca. 2008. Appl. [20] M. Dissipative Systems Analysis and Control. Ferris. and E. Circuits and systems I. The PATH solver: A non-monotone stabilization scheme for mixed complementarity problems. Pisano. [11] S. 51:343–353. Academic Press. A comprehensive analysis of chattering in second order sliding mode control systems. Ferris.. and F. E. J. Camlibel. Bartolini. The Linear Complementarity Problem. 2007. Set-Valued Anal. New perspectives and applications. Consistency of a time-stepping method for a class of piecewise-linear networks. P. P. 21(1):44–64. Fridman. Chaos Solitons Fractals. 2002. Lamarque. SIAM reviews. Lempio. [23] C. and L. Numerical approximations of a class of switch dynamical systems. New Perspectives and Applications. Arch. Theory and Applications. Usai. IEEE Control Systems Magazine. 2008. inria-00374840.

1990. S. [27] Z.2008. IEEE Transactions on Circuits and Systems – II: Express Briefs. [31] G. Discrete-time sliding mode control with time-varying surface for hard disk drives. Kanzow. Garcia-Gabin. 52(9). V.J. Journal of Dynamic Systems. Springer verlag. Marton and B. Yu. J. Leine and H. Dumont. Dynamics and Bifurcations of Non-Smooth Mechanical Systems. algorithms and applications. May–June 2003. 2003. Galias and X.10. Control of mechanical systems with stribeck friction and backlash. Motreanu. Rochdi. Ferris. Wells. Yu. D. 53(8):652–656. and M. Harker and J. Hu. Facchinei.F. D. 1992. [26] Z.E. C. and Y. [29] D.I. 2003. F. [38] L. in press. Analysis of zero–order holder discretiation of two–dimensional sliding– mode control systems. Guo. Discontinuous Systems. L. [33] Q. 17(1):175– 183. Measurement. December 2000. [35] A. Koshkouei and A. London. IEEE Transactions on Control Systems Technology. Van Loan. Nijmeijer. December 2008. Variational and Hemivariational Inequalities: Theory.1016/j sysconle. Volume I: Unilateral Analysis and Unilateral Mechanics. August 2006. Moler and C. Hess and S. Methods and Applications.S. Mathematical Programming. Koren and M. Xie. M. IEEE Transactions on Automatic Control. SIAM Rev. Control Engineering Practice. The implicit midpoint rule applied to discontinuous differential equations. Y. C. January 2009. [30] P. Control. Practical implementation of a neural network controller in a hard disk drive. in press. and C. version 1 . 2009. 122:793–802. Herrmann. IEEE Transactions on Control Systems Technology. Sliding mode control of discrete-time systems. Galias and X. January–February 2008. Idan. Kastner-Maresch. [41] Y. [32] R. Munson. Sliding mode control applied to reconfigurable flight control design. Springer Verlag. Nonconvex Optimization and its Applications. and Control. Integrated sliding mode guidance and control for a missile with on-off actuators. Lecture Notes in Applied and Computational Mechanics 18. 2004. 49:45–62. [34] A.001. 45(1):3–49. A. 13(1):146–154. IEEE Transactions on Circuits and Systems – II: Express Briefs. Control. Kluwer Academic Publishers. The semismooth algorithm for large scale complementarity problems. Goeleven. of Guidance. September 2007. Zinober. Lantos.R. [39] C. Zambrano. S. Yu.-S. [40] T. Complex discretization behaviors of a simple sliding-mode control system. Fischer. 48:160–220. [36] A. Lyapunov Analysis and Robust Synthesis under Uncertainty Conditions. and Dynamics. 55(12). Ge. Galias and X.A.S.. Orlov.T. doi:10. Sliding mode predictive control of a solar air conditioning plant. 26(3):452–462. 13:294– 311. and G. twenty-five years later. and Dynamics. 2009. Computing. of Guidance. Systems and Control Letters. Pang. and E.34 Acary & Brogliato [25] Z. [37] R. Finite-dimensional variational inequality and complementarity problems: a survey of theory. Communications and Control Engineering. 2001. Du. Euler’s discretization of single input sliding-mode control systems.10 Apr 2009 [28] W. January 2005. INRIA . J. Camacho. Wang. 31(1):204–214. Nineteen dubious ways to compute the exponential of a matrix. INFORMS Journal on Computing. inria-00374840.

IEEE Transactions on Control Systems Technology. and H. Contents 1 Introduction 3 2 A simple example 4 3 A class of differential inclusions 7 4 Convergence results and Chattering Free Finite–time Stabilization 9 RR n 6886 . Princeton University Press. 2003. B. Ser. Hess. volume 317. X. Discretization behaviors of equivalent control based sliding-mode control systems.T. [54] S. [49] D. V.Implicit Euler numerical simulations of sliding mode systems 35 [42] A. Soc. Ozguner. 45:118–125.H. 1993. Variational Analysis.T. 7(3):328–342. September 2003. [45] R. A numerical method for friction problems with multiple contacts. 48(9):1641–1646.J.10 Apr 2009 [47] T. Numerische Mathematik. J.P. June 2005. Keyhani. Wang. 1978. 5(1):1–7. and G. Galias. May–June 2003. 7:36–54. [51] Y. Wang and J. W. Y. [43] A.J. Passivity-based sliding mode position control for induction motor drives. [53] W. 53(6):1563–1569. Eng. Chen. IEEE Transactions on Automatic Control. Utkin. A control engineer’s guide to sliding mode control. Austral. Heemels. of Guidance. and Technology. Rockafellar and R. 1997. Proca. Seshagiri. 2009. Promtun and S. IEEE Transactions on Automatic Control.R. 2009.M. IEEE Transactions on Energy Conversion.A. and G. and U. 1990.B. Stewart. [56] X.C. June 2003. version 1 . [50] D. Yu. A. Chen. and J. and Dynamics. Control. Multiple-surface sliding controller design for pneumatic servo systems. 1996. B. Mathematical Programming Study. of Applied Science. On solution concepts and wellposedness of linear relay systems. Rockafellar. A high accuracy method for solving ODEs with discontinuous rigth–hand–side. An efficient implementation of the Lemke algorithm and its extension to deal with upper and lower bounds. New York. Int. Springer Verlag.-B. inria-00374840.Y. Wets. Discretization effect on equivalent controlbased multi-input sliding-mode control systems. [44] E. Wells and R. Automatica. Miles: A mixed inequality and nonlinear equation solver. Young. Chen. Nijmeijer. Yu and G. J. Pogromski. Chen. [46] R. Tsai and A. W. 39:2139–2147. 18(2):205–212. Miller. 37:288–308. [55] K. 18(9):506–512. Sliding mode control of a pitch-rate of an f-16 aircraft.C. [57] X. Math. J. Stewart. Z. [48] R. Mechatronics. [52] B. Wang. Sensorless sliding-mode control of induction motors using operating condition dependent models. July 2008. Rutherford. May 1999. 20(2):316–321. 26(3):463–473. ZOH discretization effect on single-input sliding mode control systems with matched uncertainties. November 2008. 1970. Automatica.I. IEEE Transactions on Energy Conversion. H. Convex Analysis. Multi-input/multi-output sliding mode control for a tailles fighter aircraft. 58:299–328. Yu. Huang.D.M. Sargent.

. . . . . . . . . . . . . . . . . . 8. . . . . . 9. . . . . . . . . . . 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8. . . . . . . . . .1 Time–discretization of the system (13) . . . . . . . . . . .5 The Filippov example . . . . . . . . . . . . . . . . . . . . . . . . . . . 9. . . . . . . . . . . . . .4 The special cases of the affine systems . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Example of an implicit Euler controller (IEC) . . . . . . . . . . . . 13 13 14 7 Two other classes of differential inclusions 7. . . .2 Mixed Complementarity Problem . . . . . . . . . . . . . . .6 Example 4: Observer based SMC . . . .3 Extensions to ZOH discretized systems 9. . . . 8. . . 9. . . . . . 8. . . . . .10 Apr 2009 36 Acary & Brogliato 5 Discrete–time Sliding Mode Control (SMC) 5. . . . .5 Algorithms . . . . . . .1 A simple hypomonotone case . . . . . . . . . . . . . .4. . . . . . .3 Newton’s linearization and Mixed Linear Complementarity Problems . . . . . . . . . . .inria-00374840. . . . .2 Example 3: Multiple sliding surfaces . . . . . . . 8. . . . . . . . . . . . . . .2 Extension to ZOH discretized systems . . . . 15 17 8 Detailed Implementation of Implicit Euler Discretization of the general case (60) 18 8. . . . . . . . . . .4. . . . . . . . . . . . .1 Mixed Linear Complementarity Problem (MLCP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Some Time–Discretization Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 12 12 6 Implementation of Discrete–Time Systems 6. . 6. . . . . . . 32 INRIA . . . . . . . . . . . . . . . . . . .1 Time–discretization . . .1 Chattering free stabilization . . . . . . . . . . 18 19 20 21 21 22 24 9 Numerical experiments 9. . . . . . . . . . . . . . version 1 . . . . . 25 25 25 25 28 28 28 10 Conclusions . . . . . . . .4 Lyapunov-based robust control . . . . . . . . . . . . . . . . . . . . . .2 Time–discretization of the system (59) . 8. . . . . . . . . . . . 9. . . . . . .

35042 Rennes Cedex Centre de recherche INRIA Saclay – Île-de-France : Parc Orsay Université .351.inria-00374840.Rocquencourt . avenue de l’Europe .54602 Villers-lès-Nancy Cedex Centre de recherche INRIA Paris – Rocquencourt : Domaine de Voluceau .BP 105 . rue du Jardin Botanique . Technopôle de Nancy-Brabois .Rocquencourt.06902 Sophia Antipolis Cedex Éditeur INRIA .Campus scientifique 615. avenue Halley .78153 Le Chesnay Cedex Centre de recherche INRIA Rennes – Bretagne Atlantique : IRISA.BP 93 .38334 Montbonnot Saint-Ismier (France) Centre de recherche INRIA Bordeaux – Sud Ouest : Domaine Universitaire .78153 Le Chesnay Cedex (France) ØØÔ »»ÛÛÛº ÒÖ º Ö ISSN 0249-6399 .59650 Villeneuve d’Ascq Centre de recherche INRIA Nancy – Grand Est : LORIA.10 Apr 2009 Centre de recherche INRIA Grenoble – Rhône-Alpes 655. version 1 . cours de la Libération . rue Jacques Monod . BP 105 .40.Domaine de Voluceau .91893 Orsay Cedex Centre de recherche INRIA Sophia Antipolis – Méditerranée : 2004.ZAC des Vignes : 4.BP 101 . Campus universitaire de Beaulieu .33405 Talence Cedex Centre de recherche INRIA Lille – Nord Europe : Parc Scientifique de la Haute Borne . route des Lucioles .

- Global Placement Application FormUploaded byanandakumarmts
- Sliding Mode Control of CNC Electromagnetic Levitation System Based OnUploaded byAamir Akram
- Undrained Stability of a Square Tunnel in a Soil Whose Strength Increases Linearly With DepthUploaded byHumberto Jorge
- 2010-Improved Integral Sliding Mode Control Methods PmsmUploaded byHieu LE
- EPE20110400010_80446125Uploaded byAnimesh Ghosh
- Robust Adaptive Controller for Uncertain Nonlinear SystemsUploaded byijitcajournal
- Eulers MethodUploaded byMustijab Haider
- Euler MethodUploaded byThierry Rodrigues
- An Improved Ray Tracing AlgorithmUploaded byJose Fernando
- Implementing Euler's MethodUploaded bysunilverma2010
- 00768190.pdfUploaded byJulx Ale V
- 27 studyUploaded byHarry Long
- EPYCFUploaded byAndres Lopez Garcia
- General Eng StandardsUploaded byPageduesca Rouel
- Atkinson IntroductionUploaded byElenaa
- Te Margonari EnginsoftUploaded byoicfbd
- MathematicsUploaded byKosuri Kavitha
- Optimization of Impact attenuator design using Finite element method for Enhanced Vehicle Safety during collisionUploaded byIJSRP ORG
- runge.pdfUploaded byAdannKalamsyah
- BTech_(SVEC-10)_Syllabus_(EEE)Uploaded byPardha Saradhi Mavilla
- 19 Noraini Jamaludin(1)Uploaded byBianca Doyle
- ListofResearchtopics2012_1stUploaded byNadeem Khan
- 2 FtpUploaded byPurushottam Ardhapurkar
- De Eulers MethodUploaded bymfelzien

- Stockbridge DamperUploaded byjobpei2
- Ejercicio 11 Capítulo 6 APA 1Uploaded byJerry Mitchell
- Early Enteral Feeding in a Very Low Birth Weight InfantUploaded byJuanda Raynaldi
- Ecm Titanium ENGUploaded byBruno Duro
- Cómo se Llega a Ser Psicólogo Reik, Theodor.pdfUploaded byMakeilyn Chaviano
- Least Common MultipleUploaded byWilliam Bailey
- Anclaje Para Muros PantallaUploaded byigaranedavera
- Integrales Dobles Por El Met de Gauss LegendreUploaded byMarco Alanoca Choque
- Catalago de artesaniaUploaded byEral Antunez
- Chords ConstrucciònUploaded byIsrael Carvajal
- tm5_818_5Uploaded byPDHLibrary
- Meaning of Genesis 4, Verse 26Uploaded byvenz
- Gestión de Inventarios 2017Uploaded bySuzanne Macedo
- VITULLO & SCAVO - O Liberalismo e a Definição Bobbiana de Democracia - RBCP v.13, n1Uploaded byGabriel Eduardo Vitullo
- Discutindo Exposições Mast Colloquia 8 Conceito Construção AvaliaçãoUploaded byamatobrunogabriel
- OTIMIZAÇÃO ESTRUTURAL TOPOLOGICA.pptxUploaded bySergio Quispe Rodríguez
- Foundations and Pre-Calculus 10Uploaded byNathan Swift
- Fear of Failure With Experimentation –Thai Cultural Effects on EducationUploaded byHugh Fox III
- Guia de PorcentajesUploaded byalito8010
- BscFT Kapurthala SylUploaded byDurbaGhosh
- Rv 2Uploaded byCarlos A Barreto
- QUESTIONÁRIOS SOBRE NORMAS REGULAMENTADORASUploaded byDiones Duarte
- Forgotten Ingredients FinalUploaded bySperantza Spe
- PERT sample questionUploaded byrajkrishna03
- Bacilus Subtilis Sobre ClavibacterUploaded byRudy Altamirano Ruiz
- 01hoyt.pdfUploaded bychemsac2
- Memorias de Una GEISHA KarenUploaded bySPCarpeDiem
- Paso 2_Resignificar Conceptos Del Diseño de Proyecto_551101_7Uploaded byKatikaGatuela
- Per Esp Pin Tope Pos Sim 15kVUploaded byJorge Toledo
- BCH Memoria 2009Uploaded byRoberto Caceres Fernandez